JP5585111B2 - Utterance content estimation device, language model creation device, method and program used therefor - Google Patents

Description

  The present invention relates to an utterance content estimation device that estimates utterance content, a language model creation device that creates a language model, an utterance content estimation method, a language model creation method, an utterance content estimation program, and a language model creation program.

  2. Description of the Related Art A voice recognition device that recognizes a word string corresponding to a voice from a voice (utterance) uttered by a user is known. As one of this type of speech recognition device, a speech recognition device that performs speech recognition processing based on a plurality of content-specific language models stored in advance is widely known.

  The content-specific language model is a model that represents the probability that a specific word appears in a word string that represents specific content (such as a topic). For example, in a word string that contains TV programs, the probability that a word such as a program name or performer name will appear increases, and in a word string that contains sports, a team name, exercise equipment name, player name, etc. The probability that this word will appear increases.

  By the way, in a series of user's utterances, the contents may change even within a group of utterances. In this case, when the speech recognition apparatus performs speech recognition processing based only on one content-specific language model, the accuracy of recognizing a word string may be reduced. A general voice recognition apparatus has a configuration in which a series of user's utterances are sequentially subjected to voice recognition processing for each utterance. As a result, the voice recognition result can be output immediately. Here, one utterance is a group of sounds that are divided in time according to the breathing of the speaker, between conversations, and the like, and is often equivalent to sentences and clauses in terms of language. At this time, it is possible to adapt to changes in content in a series of utterances by selecting a content-specific language model for each utterance for speech recognition. However, if the content changes within one utterance, it may not be possible to cope with such a method.

  Therefore, the speech recognition apparatus described in Non-Patent Document 1 is configured to use a plurality of content-specific language models mixed with different weights depending on positions in the speech within the speech recognition target speech.

Hitoshi Yamamoto, Ken Hanazawa, Kiyoichi Miki, "Speech recognition using dynamic control of language model based on category estimation", Proceedings of the 2009 Acoustical Conference of the Acoustical Society of Japan, p. 181-184

  However, in the speech recognition apparatus described in Non-Patent Document 1, the accuracy of recognizing words with specific contents decreases when different weights are not appropriately given depending on the position in the speech to be speech recognized. There's a problem.

  Accordingly, the present invention provides an utterance content estimation device, a language model creation device, which can recognize a specific phrase contained in an utterance with high accuracy even when the content changes within one utterance, It is an object to provide an utterance content estimation method, a language model creation method, an utterance content estimation program, and a language model creation program.

The utterance content estimation apparatus according to the present invention provides a first processing unit for each first processing unit, which is a processing unit obtained by dividing a processing target utterance into a plurality of time intervals corresponding to predetermined elements included in a speech recognition hypothesis . Features based on information obtained in the process of speech recognition processing, wherein a first score indicating the probability that the content of the processing unit is each of predetermined first specific content is extracted for the first processing unit A first content estimation means for calculating using a recognition feature, a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content in the utterance, and a plurality of utterances to be processed For each of the second processing units, which are processing units divided into the time intervals, the second specific content predetermined as the content in which the content of the second processing unit co-occurs with the first specific content the probability that each The to second score, the first score for a first respective specified contents of the first of each first processing unit, which is calculated by the contents estimation means, the respective relative occurrence of the first particular content 2 A second content estimation unit that is calculated based on the likelihood of appearance of the second specific content in the processing unit; and a first specific content in each first processing unit that is calculated by the first content estimation unit. together a first score for each, and a second score for a second respective specific contents of each of the second processing unit, which is calculated by the second content estimating means, shows the estimation results of the speech content An estimation result output means for outputting as information is provided.

In addition, the language model creation device according to the present invention relates to each of the first processing units, which are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability that the content of the first processing unit is each predetermined first specific content is used as information obtained in the course of the speech recognition processing extracted for the first processing unit. A first content estimation unit that calculates using a recognition feature that is a feature based on a feature and a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content in the utterance; against each of the second processing unit is a processing unit that is divided into a plurality of time intervals, said second specific contents of the second processing unit is predetermined as the content that co-occur with the first particular content each der of contents A second score indicating a probability, a first score for a first respective specified contents of the first of each first processing unit, which is calculated by the content estimating means, each for the emergence of the first particular content Second content estimation means for calculating based on the likelihood of appearance of the second specific content in the second processing unit, and a first content-specific language model corresponding to the first specific content are stored. 1 content-specific language model storage means, second content-specific language model storage means for storing a second content-specific language model corresponding to the second specific content, and first content estimation means A first score indicating the probability for each first specific content in each first processing unit, and a probability for each second specific content in each second processing unit estimated by the second content estimation means the second shows the 2 Using the score, the first content-specific language model stored in the first content-specific language model storage means, and the second content-specific language model stored in the second content-specific language model storage means And a language model creating means for creating a language model for each processing unit obtained by dividing an utterance to be processed into time sections.

In addition, the utterance content estimation method according to the present invention provides, for each of the first processing units, which are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability that the content of the first processing unit is each predetermined first specific content is used as information obtained in the course of the speech recognition processing extracted for the first processing unit. Calculating using the recognition feature that is based on the feature and the co-occurrence feature that is based on the set of linguistic information that appears simultaneously with the first specific content in the utterance, and divides the utterance to be processed into multiple time intervals For each second processing unit that is the processing unit, the probability that the content of the second processing unit is each second specific content that is predetermined as content that co-occurs with the first specific content. Second score to show , A first score for a first respective specified contents of the first processing unit, to the appearance of the ease of the second specific contents of the first particular each second processing units to the appearance of the contents Calculated based on the first score for each of the first specific contents in each first processing unit and the second score for each of the second specific contents in each second processing unit, The information is output as information indicating the estimation result of the utterance content.

In addition, the language model creation method according to the present invention provides, for each first processing unit, which is a processing unit divided into a plurality of time intervals corresponding to a predetermined element included in the speech recognition hypothesis. The first score indicating the probability that the content of the first processing unit is each predetermined first specific content is used as information obtained in the course of the speech recognition processing extracted for the first processing unit. Calculating using the recognition feature that is based on the feature and the co-occurrence feature that is based on the set of linguistic information that appears simultaneously with the first specific content in the utterance, and divides the utterance to be processed into multiple time intervals For each second processing unit that is the processing unit, the probability that the content of the second processing unit is each second specific content that is predetermined as content that co-occurs with the first specific content. Second scoop showing And a first score for a first respective specified contents of the first processing unit, and the ease of occurrence of the second specific contents of the first particular each second processing units to the appearance of the contents The first score for each of the first specific contents in each first processing unit, the second score for each of the second specific contents in each second processing unit, and pre-stored The first specific language model corresponding to the first specific content and the second specific language model corresponding to the second specific content stored in advance are used for processing. A language model is created for each processing unit obtained by dividing an utterance into time intervals.

Further, the utterance content estimation program according to the present invention allows a computer to process each first processing unit, which is a processing unit divided into a plurality of time intervals corresponding to predetermined elements included in a speech recognition hypothesis. The first score indicating the probability that the content of the first processing unit is each of the first specific content determined in advance is obtained in the course of the speech recognition processing extracted for the first processing unit. Processing using a recognition feature, which is a feature based on the information to be obtained, and a co-occurrence feature, which is a feature based on a set of language information that appears simultaneously with the first specific content in the utterance, For each second processing unit that is a processing unit divided into time intervals, each second specific content that is predetermined as content that co-occurs with the first specific content. A second score indicating a certain probability, the first score for a first respective specified contents of the first processing unit, the second specific in each of the second processing units to the appearance of the first particular content Processing based on the easiness of appearance of the contents, and the first score for each of the first specific contents in each first processing unit and the second specific contents in each second processing unit In addition to the second score, a process of outputting the information indicating the estimation result of the utterance content is executed.

In addition, the language model creation program according to the present invention allows a computer to process each first processing unit, which is a processing unit divided into a plurality of time intervals corresponding to predetermined elements included in a speech recognition hypothesis. The first score indicating the probability that the content of the first processing unit is each of the first specific content determined in advance is obtained in the course of the speech recognition processing extracted for the first processing unit. Processing using a recognition feature, which is a feature based on the information to be obtained, and a co-occurrence feature, which is a feature based on a set of language information that appears simultaneously with the first specific content in the utterance, For each second processing unit that is a processing unit divided into time intervals, the second specific content that is predetermined as the content that co-occurs with the first specific content. The In a second score that indicates a probability, the first score for a first respective specified contents of the first processing unit, the second in the first particular each second processing units to the appearance of the contents Processing calculated based on the easiness of appearance of specific content, the first score for each of the first specific content in each first processing unit, and the second specific content in each second processing unit The second score for each, the first content-specific language model corresponding to the first specific content stored in advance, and the second content corresponding to the second specific content stored in advance Using another language model, a process for creating a language model is executed for each processing unit obtained by dividing a speech to be processed into time intervals.

  According to the present invention, even when the content changes within one utterance, it is possible to recognize a phrase having a specific content included in the utterance with high accuracy.

It is a block diagram which shows the structural example of the speech recognition apparatus of 1st Embodiment. It is explanatory drawing which shows the example of the parameter (feature information) contained in a 1st content model. It is explanatory drawing which shows the example of the feature-value which the 1st content estimation part 21 extracts. It is explanatory drawing which shows the example of the parameter (weight information) contained in a 1st content model. It is explanatory drawing which shows the example of the process unit sequence which the 1st content estimation part 21 produces | generates. It is explanatory drawing which shows the example of the lattice which the 1st content estimation part 21 produces | generates. It is explanatory drawing which shows the example of the 1st score which the 1st content estimation part 21 computed. It is explanatory drawing which shows the example of the parameter (Assignment score to an adjacent process unit) memorize | stored as information which shows a 2nd content model. It is explanatory drawing which shows the example of the 2nd score which the 2nd content estimation part 31 computed. It is explanatory drawing which shows the other example of the 2nd score which the 2nd content estimation part 31 computed. It is explanatory drawing which shows the other example of the 2nd score which the 2nd content estimation part 31 computed. It is explanatory drawing which shows the example of a word list. It is a flowchart which shows the operation example of the speech recognition apparatus of 1st Embodiment. It is a block diagram which shows the outline | summary of this invention. It is a block diagram which shows the outline | summary of this invention.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of the speech recognition apparatus according to the first embodiment of this invention. A speech recognition apparatus 100 shown in FIG. 1 includes a speech recognition unit 11, a first content estimation unit 21, a first content model storage unit 22, a second content estimation unit 31, and a second content model storage. Unit 32, language model creation unit 41, first content-specific language model storage unit 42, and second content-specific language model storage unit 43.

  The speech recognition apparatus 100 is realized by, for example, a computer including a central processing unit (CPU), a storage device (memory and a hard disk drive (HDD)), an input device, and an output device.

  The output device is, for example, a display device. The output device displays an image made up of characters and figures based on the image information output by the CPU. The output device may be an interface device with a data storage medium or a network. In such a case, information on the speech recognition result is output via a data storage medium or a network.

  The input device is, for example, a microphone. The input device inputs, for example, a voice signal around the microphone (that is, outside the voice recognition device 100). The input device may be an interface device with a data storage medium or a network. In such a case, an audio signal is input via a data storage medium or a network.

  In this embodiment, the speech recognition apparatus 100 is configured to input a speech signal from the outside via an input device and output a speech recognition result corresponding to the input speech signal. .

  In the present embodiment, the speech recognition unit 11, the first content estimation unit 21, the second content estimation unit 31, and the language model creation unit 41 are stored in a storage device, for example, a CPU included in the speech recognition device 100. It is realized by operating according to the program. Alternatively, it may be realized by hardware such as a logic circuit. The first content model storage unit 22, the second content model storage unit 32, the first content-specific language model storage unit 42, and the second content-specific language model storage unit 43 include, for example, the speech recognition device 100. This is realized by a storage device provided.

  The speech recognition unit 11 performs speech recognition processing on the input speech signal and outputs a speech recognition hypothesis corresponding to the speech signal. In the present embodiment, the speech recognition unit 11 outputs a speech recognition hypothesis to the first content estimation unit 21 and an output device (not shown).

  In the speech recognition processing, the speech recognition unit 11 performs speech on the input speech signal according to a score given by a model for speech recognition (including an acoustic model, a language model, a word dictionary, etc.), for example. A general process such as searching for a word string that matches the signal may be performed. For example, the speech recognition unit 11 may use a hidden Markov model as an acoustic model and a word trigram as a language model. Note that the speech recognition apparatus 100 stores these models in a storage device in advance.

  In addition, the speech recognition unit 11 may output, as a speech recognition hypothesis, for example, result information (speech recognition result information) expressing a word / phrase candidate corresponding to the speech signal as one word string. For example, as a speech recognition hypothesis, result information expressed in the form of a word graph including a plurality of word strings or an N best word string may be output. At this time, the speech recognition unit 11 outputs result information including time information indicating which section of the input speech signal each word included in the word string output as the speech recognition hypothesis corresponds to.

  The first content estimation unit 21 receives the speech recognition hypothesis output from the speech recognition unit 11, generates a processing unit sequence to be used in the content estimation process, and outputs the content for each of the generated processing units. Is estimated. More specifically, the probability (score indicating likelihood) that the content of each processing unit is specific content is calculated. This value may be a value that increases as the probability that the content is specific content increases. For example, it may be the appearance probability of the specific content, or a value called likelihood or weight for the specific content. It may be.

  Based on the content model stored in the first content model storage unit 22, the first content estimation unit 21 obtains this score for specific content for each processing unit. Hereinafter, the score obtained by the first content estimation unit 21 is referred to as a first score. The first content estimation unit 21 outputs the obtained first score to the second content estimation unit 31 and the language model creation unit 41.

  The 1st content estimation part 21 uses what divided | segmented from the start end of the speech of speech recognition object into the several area as a process unit for content estimation. Each section which is a processing unit may be determined by the start time and end time of the section in the utterance.

  In this embodiment, the 1st content estimation part 21 uses the area corresponding to each word contained in the word sequence of a speech recognition hypothesis as a processing unit of content estimation. Note that a section corresponding to a character or phoneme included in the speech recognition hypothesis may be used as a processing unit. By doing in this way, the 1st content estimation part 21 can divide the input audio | voice into a several area without exception.

  In the present embodiment, a case where a search condition for information search is used as the specific content estimated by the first content estimation unit 21 will be described as an example. In the utterance for information retrieval, the probability that a specific word appears (exists) in the utterance varies depending on the retrieval condition. For example, words used as search conditions in the case of utterances for searching for TV programs include personal names (including talent names and group names), program names, program genre names (variety, sports, etc.), broadcasting stations Names, time expressions (evening, 8 o'clock, etc.) are used. In the case of this example, word wrapping names (search word item name, type used for search conditions such as “person name”, “program name”, “program genre name”, “broadcast station name”, “time expression”) , Also referred to as an attribute) may be the specific content for content estimation. Hereinafter, the specific content used by the first content estimation unit 21 for content estimation is referred to as first content.

  Thus, in order to estimate the utterance content, the first content estimation unit 21 represents the probability that the utterance content corresponding to the section is the first content for each section (processing unit) during the utterance. Calculate the score. Therefore, even when the content changes during the utterance, the utterance content can be estimated for each section.

  The first content model storage unit 22 stores content model information representing a relationship between a content estimation processing unit and a probability that the utterance content corresponding to the processing unit is each of the plurality of first contents.

  For example, a probability model based on the theory of conditional random fields (CRF) may be used as the content model. This content model is represented by the following equation (1).

P (C | W) = exp (Λ · Φ (W, C)) / Z (1)

  Here, W is a processing unit sequence for content estimation. C is a label string of contents corresponding to each processing unit (a label string of the first contents). That is, the left side P (C | W) of the equation (1) represents the probability that the label column of the content represented by the processing unit column W is C. Further, Φ (W, C) is a feature extracted from the processing unit sequence W. Λ is information representing a weighting coefficient corresponding to each element of the feature. Z is a normalization term. Note that exp () represents a function for obtaining the power of a numerical value with the natural logarithm e as the base.

  The first content model storage unit 22 may store, for example, the definition (extraction method) of the feature Φ and the weight Λ corresponding to each feature as information indicating the content model. The weight Λ may be a different value depending on the first content. By giving a different value for each first content, it is possible to make a difference in the appearance probability of each content in the processing unit. In this embodiment, a recognition feature and a co-occurrence feature are used as features to be extracted for each processing unit.

  The recognition feature is a feature based on information obtained in the process of speech recognition processing. For example, the reliability of speech recognition in the section may be used. The reliability of speech recognition is a value having a correlation with the correctness (accuracy) of the speech recognition result. For example, reliability based on word posterior probabilities calculated after voice recognition processing may be used. By using such a feature, the first content estimation unit 21 can easily detect misrecognition included in the speech recognition hypothesis. In particular, since a word used as a search condition for information search is often erroneously recognized, it is more preferable to use such a feature.

  A co-occurrence feature is a feature based on a set of linguistic information (for example, words and parts of speech) that appear simultaneously in an utterance. For example, when each processing unit has a certain first content, it is a feature expressed as “There is“ word X ”in the preceding section”.

  FIG. 2 is an explanatory diagram illustrating an example of feature information stored in the first content model storage unit 22. In the example illustrated in FIG. 2, an ID that is a feature identifier and a definition of the feature are stored in association with each other. For example, in FIG. 2, the characteristic “reliability of speech recognition in the section” is defined as ID = 1. This feature is an example of a recognition feature. Further, for example, a feature that “There is“ out ”in the preceding section” is defined as ID = 2. This feature is an example of a co-occurrence feature. In addition, ID = 3 has a feature that “there is“ out ”in the following section”, ID = 4 has a feature that “the preceding section has“ Asano ”, and ID = 5 has a“ in the following section ” An example of the feature “There is Asano” is shown. It should be noted that any word included in the word dictionary may be used as the word used for the co-occurrence feature (such as “Out” and “Asano” in this example). When these words need to be selected, it is more preferable to select words that co-occur with the first content.

  As shown in FIG. 2, the co-occurrence feature may be treated as a different feature depending on whether a certain word precedes or follows the processing interval of the feature extraction target. Regardless of whether it appears or not, it may be treated as one feature. In addition, the co-occurrence features may be treated as different features by distinguishing them according to the distance of the processing section to be extracted. For example, different features may be handled by distinction such as “adjacent”, “one sandwiched between”, and “two or more sandwiched between”. This distance may be the number of processing sections (that is, the number of words), or the format of time information based on the start time and end time in the utterance of the processing section (for example, “30 frames apart” )) May be used.

  For example, when “out” appears in a preceding section of a processing unit in the input speech recognition hypothesis, the first content estimation unit 21 has a specific first content with the processing unit content. The feature amount = 1 indicating the presence of co-occurrence may be given to the feature that “out” has appeared in the preceding section. If “out” does not appear in the preceding section, it is determined that these do not co-occur and feature value = 0 indicating no co-occurrence may be given. FIG. 3 is an explanatory diagram illustrating an example of the feature amount extracted by the first content estimation unit 21. In FIG. 3, for example, the first content estimation unit 21 has a value (feature amount) of 0.3 for a feature defined as ID = 1 (“reliability of speech recognition in the section”) for a certain processing unit. ) Is extracted. Further, for example, a value of 1 (feature amount) is extracted for the feature defined as ID = 2. This is because the word “out” appears in the preceding section of the processing unit. Further, for example, a value of 0 (feature amount) is extracted for the feature defined as ID = 5. This is because the word “Asano” does not appear in the subsequent section of the processing unit. Note that FIG. 3 shows an example of the feature amount extracted for the processing unit 3 in the processing unit sequence shown in FIG.

  FIG. 4 is an explanatory diagram showing an example of weight information stored in the first content model storage unit 22. In the example shown in FIG. 4, an ID for identifying a feature and a first content-specific weight Λ for the feature are stored in association with each other. FIG. 4 shows an example in which the first contents are three types of “person name”, “genre name”, and “others”. For example, in FIG. 4, for the feature with ID = 1, the weight Λ = 0.4 for “person name”, the weight Λ = 0.3 for “genre name”, and the “other” Defines a weight Λ = −0.2.

  Based on the product of the value (feature amount) extracted for each feature and the weighting factor (Λ) separately determined according to the first content, the first content estimation unit 21 Calculate the score. In addition, as shown in FIG. 4, it is preferable that the thing with which the weighting coefficient with respect to each feature differs according to the 1st content estimated by the 1st content estimation part 21 is contained. In other words, it is preferable that many features having different weighting factors are defined for each first content. Thereby, a different score is obtained according to the first content, and the appearance probability of each content in the processing unit can be differentiated.

  Here, a method of calculating the first score for each section in which the first content estimation unit 21 is speaking will be described using a specific example. In the following, a case where the content estimation processing unit is a section corresponding to each word of the speech recognition hypothesis and CRF is used as the first content model will be exemplified. FIG. 5 is an explanatory diagram illustrating an example of a processing unit sequence generated from a speech recognition hypothesis input for an utterance to be processed. FIG. 5A is an explanatory diagram illustrating an example of a processing target utterance. FIG. 5B is an explanatory diagram showing an example of a word string of the input speech recognition hypothesis. FIG. 5C is an explanatory diagram showing an example of the generated processing unit sequence. As shown in FIG. 5 (a), for example, when the utterance to be processed is “a music program in which tomorrow is exiled”, a speech recognition hypothesis as shown in FIG. 5 (b) is obtained. Suppose. In FIG. 5B, the word string shown as the speech recognition hypothesis is “Asano / Out / Absent / Ru / Toka / Goto / Tei / I / Music program”. . Here, “/” indicates a word break. In addition, the position in the utterance of each word is as follows: word 1: “Asano” = 0-30, word 2: “out” = 30-40, word 3: “absent” = 40-60, word 4: “ru” = 60-75, Word 5: “Toka” = 75-95, Word 6: “Ga” = 95-105, Word 7: “Out” = 105-115, Word 8: “Te” = 115-125, Word 9 : “I” = 125-140, Word 10: “Music program” = 140-200. In such a case, the 1st content estimation part 21 produces | generates a process unit for every area corresponding to each word contained in the word sequence of the input speech recognition hypothesis, for example. That is, the processing unit 1 is generated for the section corresponding to the word 1, the processing unit 2 is generated for the section corresponding to the word 2, and the process is repeated for the section corresponding to the word 10 at the end. By generating the unit 10, the processing unit sequence including the processing units 1 to 10 is generated.

  Next, the first content estimation unit 21 gives the content of each processing unit with reference to the content model stored in the first content model storage unit 22 for each generated processing unit. The probability of each of the first contents is calculated. In this example, the first content estimation unit 21 calculates a posteriori probability of each node in a lattice (graph structure) representing a combination of all the contents C that can be taken by the processing unit sequence W (processing unit word). .

  For example, when the content C candidates (that is, the given first content) are three types of “person name”, “genre name”, and “others”, the first content estimation unit 21 performs the lattice as shown in FIG. Is generated. FIG. 6 is an explanatory diagram illustrating an example of a lattice generated by the first content estimation unit 21. In the example illustrated in FIG. 6, the words in each processing unit represent a combination when the content is “a” “person name”, the content b is “genre name”, and the content c is “other”.

  In FIG. 6, for example, the node “1a” indicates a state in which the word 1 corresponding to the processing unit 1 (“Asano” in the example of FIG. 5B) is “person name”. For example, the node “1b” indicates a state in which the word 1 corresponding to the processing unit 1 is “genre name”. For example, the node “1c” indicates a state in which the word 1 corresponding to the processing unit 1 is “others”. Similarly, for example, the node “3a” indicates a state in which the word 3 corresponding to the processing unit 3 (“absent” in the example of FIG. 5B) is “person name”.

  The first content estimation unit 21 extracts, for each node, for example, values serving as feature amounts based on the feature definition illustrated in FIG. 2, and these extracted values (see FIG. 3) and FIG. A first score is calculated based on the product of weighting factors (here, the weighting factors given to the contents associated with the node) corresponding to each feature exemplified.

  Below, an example of the method of calculating | requiring the posterior probability in each node based on above-mentioned Formula (1) is shown. The a posteriori appearance probability (a posteriori probability) p (Ci = j | W) is calculated by a recursive calculation using a forward algorithm and a backward algorithm. Here, Ci = j indicates that the content of the i-th processing unit is content j. The first content estimation unit 21 obtains the posterior probability p as the appearance probability of each content in the section. FIG. 7 is an explanatory diagram illustrating an example of the appearance probability (first score) of each content in each processing section calculated by the first content estimation unit 21. As illustrated in FIG. 7, the first content estimation unit 21 outputs the appearance probability of each content (first content) for each section of the processing unit as a first score. In the example illustrated in FIG. 7, for example, the appearance probability of “others” and the appearance probability of “person name” in the processing unit 1 are values around 0.4 to 0.5, and the appearance probability of “genre name” is zero. It is shown that the value is close to.

  Note that the model parameter of the CRF that is the first content model is a combination of a model input (W: processing unit sequence) and a model output (C: content label sequence) associated in advance. The data may be learned by optimizing by an iterative calculation method or the like according to the criterion for maximizing the log likelihood of the above-described equation (1). Specific methods of identification using CRF, calculation of posterior probabilities of identification results, and learning of model parameters are described in, for example, the literature “J. Laffery, A. McCallum, F. Pereira,“ Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data ", Proceedings of 18th International Conference of Machine Learning, 2001, p.282-289" (Non-Patent Document 2) may be used.

  The second content estimation unit 31 stores the first score for each content (first content) for each processing unit output by the first content estimation unit 21 and the second content model storage unit 32. On the basis of the second content model, a probability (score indicating likelihood) that the content of the section is specific content is obtained for each processing unit. Hereinafter, the score obtained by the second content estimation unit 31 is referred to as a second score. The second content estimation unit 31 outputs the obtained second score to the language model creation unit 41.

  Similar to the first content estimation unit 21, the second content estimation unit 31 uses a speech recognition target speech divided into a plurality of sections as a processing unit. In the present embodiment, it is assumed that the processing unit generated by the first content estimation unit 21 is used. Note that the second content estimation unit 31 may use a time interval different from that generated by the first content estimation unit 21 as a processing unit for score calculation.

  In the present embodiment, the second content estimation unit 31 is content that co-occurs with the specific content (first content) used in the first content estimation unit 21 as the specific content to be estimated (that is, the front-rear expression). ) Is used. Hereinafter, the specific content used by the second content estimation unit 31 for content estimation is referred to as second content. In an utterance for information retrieval, specific expressions (phrases) that differ depending on the retrieval condition frequently appear before and after the word representing the retrieval condition. For example, in the case of an utterance for searching for a TV program, a phrase expression such as “appearing” often appears in a word indicating a search condition of “person name”. In this example, a corresponding phrase expression is defined as the second content for the first content.

  For example, when the first content estimation unit 21 uses the grouping of the expression “person name” as the first content, the second content estimation unit 31 uses “phrase expression of the personal name” as the second content. Further, for example, when the first content estimation unit 21 uses “person name”, “genre name”, and “others” as the first content, the second content estimation unit 31 selects “phrase expression of personal name” “genre name”. May be used as the second content. In this example, “others” is not used as the second content because “others” does not have a particularly frequent expression.

  The second content model storage unit 32 stores information on the second content model including information indicating the ease of appearance of the second content in each processing unit. FIG. 8 is an explanatory diagram illustrating an example of parameters stored as information indicating the second content model. In the example illustrated in FIG. 8, the second content model storage unit 32 holds, for each second content, a score (giving score to an adjacent processing unit) that is given to the likelihood of appearance. When the co-occurrence feature used in the first content estimation is defined by the positional relationship (front and rear or distance) co-occurring with the processing unit, a different value may be held for each positional relationship. .

  As the second score calculation method, the second content estimation unit 31 is based on the first content having the largest first score in each processing unit, and the first content in the processing unit adjacent to the processing unit. A score is given such that the second score for the second content (phrase expression) corresponding to is increased. For example, a score (a value given to the appearance probability) determined according to the second content corresponding to the first content having the largest first score in each processing unit is adjacent to the score. You may obtain | require the 2nd score which shows the probability which is a 2nd content about each process unit by performing the process provided to a process unit.

  Here, the calculation method of the 2nd score in the 2nd content estimation part 31 is demonstrated using the process unit 3 shown in FIG.5 (c) as an example. In the example shown in FIG. 7, the second content estimation unit 31 has the highest first score in the processing unit 3 as the “person name” score (first probability). According to the content model, a score of 0.4 is given to the processing unit adjacent to the processing unit 3. FIG. 9 is an explanatory diagram illustrating an example of the second score calculated by the second content estimation unit 31. In FIG. 9, according to the second content model shown in FIG. 8, up to four adjacent processing units are given scores. In addition, although the example shown in FIG. 9 shows an example in which a score is given except for the processing unit having the highest score of “person name”, a score may be simply assigned to an adjacent processing unit for each processing unit. Good. In such a case, for example, for a processing unit in which a score is given redundantly by the giving process to the adjacent processing unit for each processing unit, the score value to be given may be given after adjustment. As already described, it may be given so that it is simply doubled, or it may not be given a duplicate score.

  The second content estimation unit 31 may give the score by weighting the value of the second content model according to the magnitude of the first score when giving the score to the adjacent processing unit. For example, when the first score calculated for a certain processing unit is smaller than a predetermined value, processing such as decreasing the score to be given based on the first score may be performed. In addition, when a feature distinguished by the positional relationship (front and back or distance) co-occurring with the processing unit is used as the co-occurrence feature used in the first content estimation, the second content is determined for each positional relationship. You may give different scores. For example, different values may be given to the preceding processing unit and the subsequent processing unit, or different values may be given to processing units that are two or more away from the next processing unit.

  Further, as already described, the second content estimation unit 31 may use a time unit different from that of the first content estimation unit 21 as a processing unit for score calculation. In such a case, the second content estimation unit 31 determines the processing unit (first content) based on the first content having the largest first score in each processing unit of the first content estimation unit 21. With respect to the processing unit adjacent to the processing unit in the estimation unit 21 (here, the processing unit in the second content estimation unit 31), the second content corresponding to the first content (phrase expression) What is necessary is just to give a score so that the 2nd score of this may become large. For example, when the processing unit of the first content estimation unit 21 is a “word” unit and the processing unit of the second content estimation unit 31 is a “frame” unit, for several frames before and after the word, It is also possible to perform processing such as giving a score.

  In addition, the second content estimation unit 31 outputs both the first score and the second score, and at that time, each score (first score) for the first content obtained by the first content estimation unit 21 is output. 1) and each score (second score) for the second content obtained by the second content estimation unit 31 may be normalized and output for each processing unit. By normalizing and outputting, it is possible to output an utterance content estimation result obtained by comprehensively evaluating the probability of the first content and the probability of the second content for the input utterance to be processed. In addition, what is necessary is just to match | combine to the smaller processing unit, when using a different processing unit by the 1st content estimation part 21 and the 2nd content estimation part 31. FIG.

  When the present invention is implemented as an utterance content estimation device, the first score calculated by the first content estimation unit 21 and the second score calculated by the second content estimation unit 31 are used. In addition, it may be output as information indicating the estimation result of the utterance content to be processed. For example, an estimation result output unit (not shown) may be provided to input a first score and a second score, and output information indicating the estimation result of the utterance content to be processed based on the first score and the second score. Good. The estimation result output unit may output the information including the first score and the second score as information indicating the estimation result of the utterance content to be processed as it is, or the first result as described above. You may output what normalized the score and the 2nd score for every process unit as information which shows the estimation result of the speech content of a process target. In addition, each content estimation part may also use the function of this estimation result output part, and each may output the calculated | required score, or the 2nd content estimation part 31 outputs collectively, and an estimation result output part is output. You may also use it.

  10 and 11 are explanatory diagrams illustrating other examples of the second score calculated by the second content estimation unit 31. FIG. FIG. 10 is an explanatory diagram showing an example of giving a second score when the co-occurrence features are classified according to the distance as the relationship between the positions that co-occur with the processing units. FIG. 11 is an explanatory diagram showing an example of giving a second score when the co-occurrence features are classified according to the front-rear relationship of the position where the process unit co-occurs. Both FIG. 10 and FIG. 11 show an example of giving the second score for the “personal expression”.

  The language model creation unit 41 includes a calculated first score, a calculated second score, a first content-specific language model stored in the first content-specific language model storage unit 42, a first Based on the second content-specific language model storage unit 43 and the second content-specific language model 43, a language model representing a probability that a specific word appears for each processing unit among the speech recognition target speeches. create. In addition, the language model creation unit 41 outputs the created language model to the speech recognition unit 11.

  The language model created by the first content-specific language model, the second content-specific language model, and the language model creation unit 41 is defined, for example, as the probability that a certain word appears depends on the immediately preceding N-1 words. N-gram language model may be used.

In the N-gram language model, the appearance probability of the i-th word w _i is represented by P (w _i | W _{i−N + 1} ⁱ⁻¹ ). Here, W _{i−N + 1} ⁱ⁻¹ in the condition part represents the (i−N + 1) to (i−1) th word string. The model with N = 2 is called a bigram model, and the model with N = 3 is called a trigram model. A model constructed based on the assumption that it is not affected by the immediately preceding word is called a unigram model.

According to the N-gram language model, the probability P (W _i ⁿ ) that the word string W _i ⁿ = (w ₁ , w ₂ ,..., W _n ) appears is expressed by the following equation (2).

P (W _i ⁿ ) = Π _i P (w _i | W _{i−N + 1} ⁱ⁻¹ ) (2)

  In addition, a parameter composed of various conditional probabilities of various words used in such an N-gram language model is obtained by maximum likelihood estimation or the like for learning text data.

  The language model creation unit 41, for example, the probability of each content (first content and second content) in each section of the speech of speech recognition target indicated by the first score and the second score, and the first A language model may be created according to the following equation (3) using a plurality of content-specific language models stored in the content-specific language model storage unit 42 and the second content-specific language model storage unit 43: .

P _t (w _i ) = Σ _j α _j (t) P _j (w _i ) (3)

In Expression (3), P _t (w _i ) is a probability that the word w _i appears in the section t. Α _j (t) is a probability (score) that the content in the section t is the content j. P _j (w _i ) is a probability that the word w _i appears in the content-specific language model for the content j. In this example, the language model creation unit 41 uses the score calculated by the first content estimation unit 21 and the second content estimation unit 31 (the appearance probability of the content in each section in the utterance) of Equation (3). Used as α _j (t). Here, a product obtained by multiplying the first probability and the second probability by different coefficients may be used as α _j (t) in Expression (3).

  Here, t in Expression (3) may be a section corresponding to a time frame used in the speech recognition processing, or may be a time representing a time point in the utterance.

  The first content-specific language model stored in the first content-specific language model storage unit 42 uses, as learning data, a word string including words that express the first content. For example, a word list expressing search conditions may be prepared, and a word string including them may be used as learning data.

  The second content-specific language model stored in the second content-specific language model storage unit 43 uses, as learning data, a word string including words that express the second content. For each search condition, a list of wording expressions selected as specific expressions described later may be prepared, and a word string including them may be used as learning data.

  FIG. 12 is an explanatory diagram illustrating an example of a word list. FIG. 12A is an explanatory diagram illustrating an example of a word list having a first content, and FIG. 12B is an explanatory diagram illustrating an example of a word list having a second content. For example, in FIG. 12A, the first content “person name” includes a word such as “Exile”.

  Regarding the second content word list, in this example, a specific expression (phrase expression) included in the second content is determined according to the first content model. Here, the specific expression included in the second content is a specific expression such as “appearing” in relation to the expression of the expression “personal expression” in the second content. It is. Specifically, among the CRF parameters learned as the first content model, for each first content, a word that greatly contributes to the estimation of the content is selected. For example, in the case of the parameter (weight) example shown in FIG. 4, the word “out” used for the feature (ID = 2) having a large weighting coefficient with respect to the first content “person name” is a specific expression. , And others are not selected as specific expressions. Note that the number of specific expressions is not limited to one, and any number may be determined.

In such a case, the language model creation unit 41 assigns the second score to α _j (t) only for a specific expression selected according to the first content model in the word list of the second content. You may make it reflect. For example, in the example shown in FIG. 12B, for the second content “phrase expression of a person name”, examples such as “appear”, “out”, and “moderator” are shown. When only the word “out” is selected, the second score is reflected on α _j (t) only for the word “out” using “out”. Thereby, it is possible to prevent the second score from being excessively reflected in the phrase expression having a small co-occurrence relationship. Note that the determination of whether or not a specific expression is made is not limited to the method performed in units of words as described above, and it is also possible to determine so as to include, for example, a part of words that match. In such a case, the word “appear” will be selected. The selection of the specific expression for the second content may be performed at the same timing as the learning of the first content model. That is, it may be performed in advance before a series of processes of the present embodiment. It should be noted that the selection of “first content” and “second content”, that is, “first content” and “second content”, also determines the type of utterance to be processed and the content estimation. It is assumed that the process is performed in advance according to the purpose, and in the present embodiment, it is given as a set value when starting a series of processes.

  In this way, by learning the content-specific language model, different appearance probabilities are given to words expressing specific content in different content-specific language models.

  The content-specific language model storage units 42 and 43 may store the above-described word list as a list of words that are particularly desired to be recognized. In this case, the language model creation unit 41 may be configured to increase, for example, the probability that a word included in the word list for the content with the highest score appears in each section in the utterance by a predetermined value. .

  Note that the language model creation unit 41 may output all the information included in the language model when outputting the created language model, or may output only information specified from the outside.

  As described above, the language model creation unit 41 creates a weighted language model for each content-specific language model, using the estimation result of the appearance probability of each content in the section for each speech recognition target speech section as a weighting coefficient.

  In the subsequent stage, since the speech recognition unit 11 uses the language model created by the language model creation unit 41, speech recognition processing can be performed using an accurate language model for each section. The result can be output.

  Next, the operation of this embodiment will be described with reference to the flowchart shown in FIG. FIG. 13 is a flowchart showing an example of the operation of the speech recognition apparatus 100 of the present embodiment. As shown in FIG. 13, when activated, the speech recognition apparatus 100 reads necessary data from a storage device or the like, and recognizes the speech recognition unit 11, the first content estimation unit 21, the second content estimation unit 31, and the language. The first content model storage unit 22, the second content model storage unit 32, the first content-specific language model storage unit 42, and the second content-specific language model storage unit 43 can be referred to from the model creation unit 41. Initialization processing such as data expansion is performed (step S11).

  On the other hand, the voice recognition unit 11 receives a voice signal in response to a notification from the input device, and performs a voice recognition process (step S12). Further, the speech recognition unit 11 outputs the speech recognition hypothesis obtained by the speech recognition process to the first content estimation unit 21.

  Next, the first content estimation unit 21 generates a processing unit sequence generated from the speech recognition hypothesis by the method already described based on the first content model stored in the first content model storage unit 22. The first score, which is the probability that the content corresponding to each processing unit included in is a specific content (first content), is calculated (step S13).

  Next, the second content estimation unit 31 includes the appearance probability (first score) of each content (first content) of each processing unit estimated by the first content estimation unit 21 and the second content. Based on the second content model stored in the model storage unit 32, the content corresponding to each processing unit is the probability that the content corresponding to each processing unit is specific content (second content) by the method described above. A score of 2 is calculated (step S14).

  Next, the language model creation unit 41 generates an appearance probability (first content and second content) of each content (first content and second content) estimated by the first content estimation unit 21 and the second content estimation unit 31. Based on the first score and the second score), and the content-specific language model stored in the first content-specific language model 42 and the second content-specific language model 43. A language model representing the probability that a specific word appears is created for each position (step S15). In addition, the language model creation unit 41 outputs the created language model to the speech recognition unit 11.

  Next, the speech recognition unit 11 performs speech recognition processing on the speech to be recognized using the language model created by the language model creation unit 41, and outputs the speech recognition hypothesis as result information to the output device (step S16). ).

  Through such a series of processes, a more accurate speech recognition hypothesis can be obtained. Note that the speech recognition unit 11 may output the speech recognition hypothesis at this point to the first content estimation unit 21 again (return to step S13).

  That is, according to the present embodiment, when estimating the content of the speech of the speech recognition target, not only the first content (for example, which search condition is), but also based on the estimation result, The second content (for example, which search condition is a wording expression) that is an expression of a word that co-occurs with the first content is also estimated. Thereby, it can avoid estimating the content which changes within the speech of speech recognition object erroneously.

  This is because the second content estimation unit 31 performs content estimation using the front and rear expressions of the specific content used in the first content estimation unit 21 as the second content, so that the appearance probability of the content with respect to the utterance is accurately determined. Because it becomes. For example, when the output of the first content estimation unit 21 is the appearance probability shown in FIG. 7, in the processing unit 1, the probability of “others” and the probability of “person name” are comparable. This is because the probability of “person name” increases due to the influence of “out” in processing unit 2. If speech recognition is performed by the speech recognition unit 11 using a language model created by the language model creation unit 41, which will be described later, using this appearance probability, the weight of the language model classified by contents of “person name” in this section Therefore, the recognition result in this section is a word having the content of “person name” (for example, “Asano”), and there is a high possibility that a recognition error (search condition insertion error) will occur.

  However, when the second content estimation unit 31 gives the processing unit 1 more appearance probabilities of the content of the “personal expression”, the probability of being “person name” can be relatively reduced. Thereby, the appearance probability of the content with respect to the utterance becomes accurate. Then, an accurate speech recognition hypothesis (in this example, “tomorrow”) can be obtained in the subsequent speech recognition unit 11.

  In utterances for information retrieval, there are many recognition errors in words of search conditions, and words in phrase expressions tend to have few recognition errors. In addition, the former tends to be larger for variations of expression. Using the utterance tendency, the first content estimation unit 21 estimates the category of the search condition by using a phrase expression that is relatively easy to recognize and has few variations. For this reason, it is relatively difficult for the first content estimation unit 21 to simultaneously estimate the content of the “phrase expression”. On the other hand, since the content of the “search condition” is relatively easy to estimate in the first content estimation unit 21, the second content estimation unit 31 uses the search condition as a clue, and the phrases that exist before and after the search condition. The interval of expression can be estimated.

  As described above, the second content estimation unit 31 estimates the appearance probability for each section in the utterance not only for the first content (search condition) but also for the second content (phrase expression). It is possible to appropriately estimate the content of the speech to be recognized.

  Therefore, the speech recognition apparatus 100 can estimate the specific contents in the utterance more accurately. As a result, the speech recognition apparatus 100 can more accurately recognize words having specific contents included in the utterance.

  Next, the outline of the present invention will be described. 14 and 15 are block diagrams showing an outline of the present invention. FIG. 14 is a block diagram showing a configuration example when the present invention is applied to an utterance content estimation apparatus. The utterance content estimation apparatus shown in FIG. 14 includes first content estimation means 101, second content estimation means 102, and estimation result output means 103.

  The first content estimation unit 101 estimates the probability that the content of the processing unit obtained by dividing the utterance to be processed into time intervals is the first specific content. The first content estimation unit 101 is realized by the first content estimation unit 21 in the above embodiment.

  The second content estimation means 102 is the second specific content that is defined as the content of the word that co-occurs with the word of the first specific content, in which the content of the processing unit obtained by dividing the utterance to be processed into time intervals. Estimate a certain probability. The second content estimation unit 102 is realized by the second content estimation unit 31 in the above embodiment.

  The estimation result output unit 103 includes a probability of the first specific content in each processing unit estimated by the first content estimation unit 101 and a second value in each processing unit estimated by the second content estimation unit 102. Together with the probability of the specific content of the message, it is output as information indicating the estimation result of the utterance content. The estimation result output means 103 outputs or normalizes and outputs the first content estimation unit 21 and the second content estimation unit 31 or the first score and the second score in the above embodiment. This is realized by the second content estimation unit 31.

  In such an utterance content estimation apparatus, the second content estimation means estimates different probabilities for the second specific content depending on the position before and after the first specific content that co-occurs. May be.

  Further, the second content estimation means may estimate different probabilities for the second specific content according to the distance of the position from the co-occurring first specific content.

  In addition, the first content estimation means extracts the processing unit when calculating the first score indicating the probability of the first specific content in each processing unit using the conditional random field model. Using a recognition feature, which is a feature based on information obtained in the course of speech recognition processing, and a co-occurrence feature, which is a feature based on a set of linguistic information that simultaneously appears in the utterance, and at least one weighting factor for each feature In addition, the second content estimation unit uses a weighting factor having a different value depending on the first specific content, and the second content estimation unit calculates the first specific content in each processing unit calculated by the first content estimation unit. A score of 1 may be used to calculate a second score indicating the probability of the second specific content in each processing unit.

  The utterance content estimation device includes a first content-specific language model storage unit that stores a first content-specific language model corresponding to the first specific content, and a second specific content-specific language model storage unit. A second language model storage unit that stores a language model by content, and a language model creation unit that creates a language model for each processing unit obtained by dividing the utterance to be processed into time intervals. The first score, which is the probability of the first specific content in each processing unit estimated by the first content estimation means, output from the estimation result output means, and estimated by the second content estimation means A second score which is the probability of the second specific content in each processing unit, a first content-specific language model stored in the first content-specific language model storage means, and a second content-specific Language model memory By using the second content by language model stored in the stage, may create the language model for each unit of processing an utterance to be processed is divided time intervals.

  FIG. 15 is a block diagram showing a configuration example when the present invention is applied to a language model creation apparatus. The language creation model shown in FIG. 15 includes a first content estimation unit 101, a second content estimation unit 102, a first content-specific language model storage unit 104, and a second content-specific language model storage unit 105. Language model creation means 106.

  The first content estimation unit 101 estimates the probability that the content of the processing unit obtained by dividing the utterance to be processed into time intervals is the first specific content. The first content estimation unit 101 may be the same unit as the first content estimation unit 101 shown in FIG.

  The second content estimation means 102 is the second specific content that is defined as the content of the word that co-occurs with the word of the first specific content, in which the content of the processing unit obtained by dividing the utterance to be processed into time intervals. Estimate a certain probability. The second content estimation unit 102 may be the same unit as the second content estimation unit 102 shown in FIG.

  The first content-specific language model storage unit 104 stores a first content-specific language model corresponding to the first specific content. The first content-specific language model storage unit 104 is realized by the first content-specific storage unit 42 in the above embodiment.

  The second content-specific language model storage unit 105 stores a second content-specific language model corresponding to the second specific content. The second content-specific language model storage unit 105 is realized by the second content-specific storage unit 43 in the above embodiment.

  The language model creation unit 106 includes a first score that is a probability of the first specific content in each processing unit estimated by the first content estimation unit, and each process estimated by the second content estimation unit. A second score which is a probability of the second specific content in the unit, a first language model stored in the first language model storage means, and a second language model storage A language model is created for each processing unit obtained by dividing the utterance to be processed into time intervals using the second language model by content stored in the means.

  (Supplementary Note 1) Further, in the utterance content estimation method according to the present invention, different probabilities may be estimated for the second specific content depending on the position before and after the co-occurring first specific content.

  (Additional remark 2) Moreover, with the speech content estimation method by this invention, you may estimate a different probability about 2nd specific content according to the distance of the position with the 1st specific content to co-occur.

  (Additional remark 3) Moreover, when calculating the 1st score which shows the probability about the 1st specific content in each processing unit using the conditional random field model by the speech content estimation method by this invention, each process As the features to be extracted for each unit, a recognition feature that is a feature based on information obtained in the process of speech recognition processing and a co-occurrence feature that is a feature based on a set of linguistic information that simultaneously appear in the speech are used, A weighting factor having a different value depending on the first specific content is used as at least one of the weighting factors, and a weighting factor having a different value depending on the first specific content is used when calculating the second score. The second content estimation means uses the processing unit used by the first content estimation means as the processing unit, and the first specific content in each processing unit calculated by the first content estimation means. Using the first score may be calculated second score indicating the probability for the second specific contents of each processing unit.

  (Additional remark 4) Moreover, the speech content estimation method by this invention is the 1st score which is the probability about the 1st specific content in each processing unit, and each processing unit estimated by the 2nd content estimation means. A second score that is a probability of the second specific content, a first language model according to the first specific content, and a second content language according to the second specific content A step of creating a language model for each processing unit obtained by dividing the utterance to be processed into time intervals using the model may be included.

  (Additional remark 5) Moreover, you may make the matter shown to the said additional remarks 1-3 apply to the language model creation method by this invention.

  (Additional remark 6) Moreover, the speech content estimation program by this invention is a process which calculates a 2nd score to a computer, About the 2nd specific content, before and after the position with the 1st specific content to co-occur Depending on, different probabilities may be estimated.

  (Additional remark 7) Moreover, the utterance content estimation program by this invention is the process of calculating a 2nd score to a computer, About the 2nd specific content, The distance of the position with the 1st specific content to co-occur Depending on, different probabilities may be estimated.

  (Additional remark 8) Moreover, the speech content estimation program by this invention is a process of calculating | requiring a 1st score in a computer, using a conditional random field model, and the process of speech recognition processing as a characteristic extracted about each processing unit. Using a recognition feature, which is a feature based on information obtained in step 1, and a co-occurrence feature, which is a feature based on a set of linguistic information that simultaneously appear in an utterance, and at least one of the weighting factors for each feature In the process of calculating the second score by using a weighting factor having a different value depending on the contents of each, the processing unit used by the first content estimating means is used as the processing unit, and each calculated by the first content estimating means By using the first score for the first specific content in the processing unit, the second score indicating the probability for the second specific content in each processing unit is calculated. Good.

  (Additional remark 9) Moreover, the utterance content estimation program by this invention makes the computer estimate each 1st score which is the probability about the 1st specific content in each processing unit, and the 2nd content estimation means. A second score that is the probability of the second specific content in the processing unit, a first language model according to the first content according to the first specific content, and a second according to the second specific content A process for creating a language model may be executed for each processing unit obtained by dividing the utterance to be processed into time intervals using the language model for each content.

  (Additional remark 10) Moreover, you may make the matter shown to the said additional remarks 6-8 apply to the language model creation program by this invention.

  The present invention is not limited to the use of directly estimating the utterance content or the use of creating a language model, but may use the utterance content estimation result or the language model (regardless of device, system, method, or program). Can be suitably applied. For example, the present invention can be suitably applied to the use of speech recognition processing for recognizing a specific phrase included in an utterance.

DESCRIPTION OF SYMBOLS 100 Speech recognition apparatus 101 1st content estimation means 102 2nd content estimation means 103 Estimation result output means 104 1st content-specific language model storage means 105 2nd content-specific language model storage means 106 Language model creation means 11 Speech Recognition unit 21 First content estimation unit 22 First content model storage unit 31 Second content estimation unit 32 Second content model storage unit 41 Language model creation unit 42 First content-specific language model storage unit 43 Second Language model storage unit by content

Claims (10)

The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis . The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance First content estimation means for calculating using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content ;
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. was the second score indicating the probability of the second respective specific content, the first score for a first respective specified contents of the first of each of the first processing unit, which is calculated by the content estimating means And second content estimation means for calculating based on the likelihood of appearance of the second specific content in each second processing unit for the appearance of the first specific content ;
A first score for each of the first specific contents in each first processing unit calculated by the first content estimating means and a second score in each second processing unit calculated by the second content estimating means. An utterance content estimation apparatus comprising: an estimation result output unit that outputs a second score for each of two specific contents together with information indicating an estimation result of the utterance content. The first processing unit is a processing unit obtained by dividing an utterance to be processed into a plurality of time intervals corresponding to words included in the speech recognition hypothesis,
2. The utterance content estimation apparatus according to claim 1 , wherein the second processing unit is the same processing unit as the first processing unit or a processing unit obtained by dividing an utterance to be processed into time intervals corresponding to frames . The first specific content is specific content for content estimation,
The utterance content estimation apparatus according to claim 1, wherein the second specific content is a wording expression corresponding to the first specific content . When calculating the first score indicating the probability of the first specific content in each first processing unit using the conditional random field model, the first content estimation means includes at least a weighting factor for each feature. one, utterance estimating device according to claim 1, Ru using a weighting factor to any one of claims 3 with different values by a first particular content. First content-specific language model storage means for storing a first content-specific language model corresponding to the first specific content;
A second content-specific language model storage unit that stores a second content-specific language model according to the second specific content;
A language model creating means for creating a language model for each processing unit obtained by dividing an utterance to be processed into time intervals;
The language model creation means outputs a first score indicating a probability for each of the first specific contents in each first processing unit estimated by the first content estimation means, which is output from the estimation result output means; A second score indicating the probability for each second specific content in each second processing unit estimated by the second content estimation means and the first content-specific language model storage means are stored. For each processing unit obtained by dividing the utterance to be processed into time intervals using the first content-specific language model and the second content-specific language model stored in the second content-specific language model storage unit The utterance content estimation apparatus according to any one of claims 1 to 4, wherein a language model is created. The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis . The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance First content estimation means for calculating using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content ;
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. was the second score indicating the probability of the second respective specific content, the first score for a first respective specified contents of the first of each of the first processing unit, which is calculated by the content estimating means And second content estimation means for calculating based on the likelihood of appearance of the second specific content in each second processing unit for the appearance of the first specific content ,
First content-specific language model storage means for storing a first content-specific language model corresponding to the first specific content;
A second content-specific language model storage unit that stores a second content-specific language model according to the second specific content;
A first score indicating a probability for each of the first specific contents in each first processing unit estimated by the first content estimation means, and each second process estimated by the second content estimation means A second score indicating the probability of each second specific content in the unit , a first language model stored in the first content-specific language model storage means, and a second content-specific Language model creation means for creating a language model for each processing unit obtained by dividing the utterance to be processed into time intervals using the second content-specific language model stored in the language model storage means. Language model creation device characterized by The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance Using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content,
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. A second score indicating the probability of being each of the second specific contents, the first score for each of the first specific contents in each first processing unit, and the appearance of the first specific contents Calculate based on the likelihood of appearance of the second specific content in each second processing unit,
The estimation result of the utterance content by combining the first score for each of the first specific contents in each first processing unit and the second score for each of the second specific contents in each second processing unit An utterance content estimation method characterized in that the information is output as information indicating the utterance. The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance Using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content,
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. A second score indicating the probability of being each of the second specific contents, the first score for each of the first specific contents in each first processing unit, and the appearance of the first specific contents Calculate based on the likelihood of appearance of the second specific content in each second processing unit,
The first score for each of the first specific contents in each first processing unit, the second score for each of the second specific contents in each second processing unit, and the previously stored The utterance to be processed using a first content-specific language model corresponding to the first specific content and a second content-specific language model corresponding to the second specific content stored in advance A language model creation method characterized by creating a language model for each processing unit divided into time intervals. On the computer,
The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance Processing using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content;
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. A second score indicating the probability of being each of the second specific contents, the first score for each of the first specific contents in each first processing unit, and the appearance of the first specific contents Processing based on the likelihood of appearance of the second specific content in each second processing unit, a first score for each of the first specific content in each first processing unit, and each second Utterance content for executing processing that is output together with the second score for each second specific content in the processing unit, as information indicating the estimation result of the utterance content for each time interval of the utterance to be processed Constant program. On the computer,
The content of the first processing unit is determined in advance for each of the first processing units that are processing units obtained by dividing the utterance to be processed into a plurality of time intervals corresponding to predetermined elements included in the speech recognition hypothesis. The first score indicating the probability of each of the first specific contents is extracted for the first processing unit, and the recognition feature that is a feature based on the information obtained in the course of the speech recognition processing and the utterance Processing using a co-occurrence feature that is a feature based on a set of linguistic information that appears simultaneously with the first specific content;
For each second processing unit that is a processing unit obtained by dividing the utterance to be processed into a plurality of time intervals, the content of the second processing unit is predetermined as content that co-occurs with the first specific content. A second score indicating the probability of being each of the second specific contents, the first score for each of the first specific contents in each first processing unit, and the appearance of the first specific contents A process that is calculated based on the easiness of appearance of the second specific content in each second processing unit, a first score for each of the first specific contents in each first processing unit, and each A second score for each of the second specific contents in two processing units, a first language model according to the first specific contents stored in advance, and the previously stored language model Second special A language model creation program for executing a process of creating a language model for each processing unit obtained by dividing the utterance to be processed into time intervals using a second content-specific language model corresponding to a predetermined content.

Images