CN106875936A - Voice recognition method and device - Google Patents

Abstract

An embodiment of the present invention provides a speech recognition method and device, the method comprising: acquiring the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation characteristics of each speech signal frame and the pronunciation of each speech signal The probability that the signal frame is mapped to the corresponding pronunciation; based on the probability included in the feature classification result, filtering the pronunciation included in the feature classification result; identifying the speech signal based on the filtered feature classification result. Implementation of the embodiment of the present invention, in the process of recognizing the speech signal, it is not necessary to perform the recognition operation related to the pronunciation that is filtered out, such as: no need to search for a path related to the pronunciation that is filtered out in the recognition network, so it can effectively The time spent in the speech recognition process is reduced, thereby improving speech recognition efficiency.

Description

Speech recognition method and device

technical field

The invention relates to the field of computer technology, in particular to a voice recognition method and device.

Background technique

With the development of computer technology, the application of automatic speech recognition (ASR) technology in human-computer interaction and other fields is increasing. At present, the speech recognition technology mainly converts the speech signal to be recognized into text information through the signal processing module, feature extraction module, acoustic model, language model (Language Model, LM), dictionary and decoder (Decoder), and completes speech recognition.

In the speech recognition process, the signal processing module and the feature extraction module first divide the speech signal to be recognized into multiple speech signal frames, and then enhance each speech signal frame by processing such as noise elimination and channel distortion, and then divide each speech signal frame The signal frame is converted from the time domain to the frequency domain, and the appropriate acoustic features are extracted from the converted speech signal frame. The acoustic model trained according to the characteristic parameters of the training speech library takes the acoustic features extracted by the feature extraction module as input, maps to the pronunciation that can describe the pronunciation characteristics of the speech signal frame, and calculates the mapping of the speech signal frame to each pronunciation. Probability to get feature classification results.

The language model contains associations between different words (such as words, words, and phrases) and their probabilities (possibility), and is used to estimate the possibility of various text information composed of different words. The decoder can establish a recognition network based on the trained acoustic model, language model and dictionary, and identify each path in the network corresponding to various text information and the pronunciation of each text information, and then output the pronunciation according to the acoustic model. Find the best path in the recognition network, and based on this path, the text information corresponding to the speech signal can be output with the maximum probability to complete the speech recognition.

However, the language model is generally a model trained based on a large amount of corpus, which contains a large number of associations and possibilities between words. Therefore, the recognition network established based on the speech model contains many nodes, and the number of branches of each node is also very large. many. When performing path search in the recognition network, the number of nodes involved in the pronunciation of each speech signal frame will increase exponentially, resulting in a large amount of path search, and the search process takes more time, which in turn will reduce the efficiency of speech recognition.

Contents of the invention

In view of this, the present invention provides a speech recognition method and device to solve the problems of time-consuming speech recognition process and low recognition efficiency.

According to a first aspect of the present invention, a kind of speech recognition method is provided, comprising steps:

Obtaining the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation characteristics of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation;

filtering the pronunciation contained in the feature classification result based on the probability contained in the feature classification result;

The speech signal is identified based on the filtered feature classification results.

In one embodiment, the filtering the pronunciation contained in the feature classification result based on the probability contained in the feature classification result includes:

Judging whether the probability that any speech signal frame is mapped to a corresponding pronunciation satisfies a predetermined filter rule;

If the corresponding pronunciation satisfies a predetermined filtering rule, the corresponding pronunciation is filtered out.

In one embodiment, if the probability difference between the probability that any speech signal frame is mapped to the corresponding pronunciation and the maximum mapping probability of the speech signal frame is within a predetermined difference range, then the corresponding pronunciation is determined Satisfy the predetermined filtering rules;

If the probability that any speech signal frame is mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to each utterance in the predetermined number of utterances, then it is determined that the corresponding utterance satisfies a predetermined filtering rule.

In one embodiment, the predetermined number is any of the following:

The number of pronunciations that are retained in the feature classification result in the pronunciation corresponding to the frame of the speech signal frame;

The product of the predetermined ratio threshold and the total number of utterances corresponding to the speech signal frame.

In one embodiment, the filtering the pronunciation contained in the feature classification result based on the probability contained in the feature classification result includes:

Obtain the histogram distribution of the probability that any speech signal frame is mapped to each pronunciation;

obtaining a beamwidth corresponding to the histogram distribution;

determining an utterance whose probability distribution is outside the beam width as an utterance satisfying the predetermined filter rule;

The utterances satisfying the predetermined filter rule are deleted from the utterances included in the feature classification result.

In one embodiment, the deleting the pronunciation satisfying the predetermined filter rule from the pronunciation contained in the feature classification result includes:

If the probability that any speech signal frame is mapped to a corresponding pronunciation satisfies a predetermined filter rule, determining the pronunciation as a candidate pronunciation;

If any frame in the predetermined number of adjacent speech signal frames of the speech signal frame, the probability mapped to the candidate pronunciation satisfies the predetermined filter rule, then the candidate pronunciation is deleted from the pronunciation contained in the feature classification result ;

If the probabilities of the predetermined number of adjacent speech signal frames of the speech signal frame mapped to the candidate pronunciation do not satisfy the predetermined filtering rule, then the candidate pronunciation is retained in the pronunciation contained in the feature classification result.

According to a second aspect of the present invention, a speech recognition device is provided, comprising:

The classification result acquisition module is used to obtain the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation for describing the pronunciation characteristics of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation;

A pronunciation filtering module, configured to filter the pronunciation contained in the feature classification result based on the probability contained in the feature classification result;

A speech recognition module, configured to recognize the speech signal based on the filtered feature classification result.

In one embodiment, the pronunciation filtering module also includes:

The first filtering module is used for mapping the probability of any speech signal frame to the corresponding pronunciation, and the probability difference between the maximum mapping probability of the speech signal frame, when within a predetermined difference range, the corresponding pronunciation to filter;

The second filtering module is configured to filter the corresponding utterance when the probability that any speech signal frame is mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to each utterance in the predetermined number of utterances.

In one embodiment, the pronunciation filtering module includes:

The probability distribution module is used to obtain the histogram distribution of the probability that any speech signal frame is mapped to each pronunciation;

A beam width determination module, configured to obtain a beam width corresponding to the histogram distribution;

a pronunciation determination module, configured to determine a pronunciation whose probability distribution is outside the beam width as a pronunciation satisfying the predetermined filter rule;

A pronunciation deletion module, configured to delete pronunciations satisfying the predetermined filter rule from the pronunciations included in the feature classification result.

In one embodiment, the pronunciation filtering module includes:

A candidate pronunciation module, configured to determine the pronunciation as a candidate pronunciation when the probability that any speech signal frame is mapped to the corresponding pronunciation satisfies a predetermined filter rule;

Candidate utterance deletion module, used to classify the candidate utterance from the features when the probability of mapping to the candidate utterance satisfies a predetermined filtering rule in any frame of the predetermined frame number of adjacent voice signal frames of the voice signal frame The pronunciation contained in the result is deleted;

The candidate pronunciation retention module is used for retaining the candidate pronunciation in the feature classification result when the probabilities of the adjacent speech signal frames of the predetermined frame number of the speech signal frame being mapped to the candidate pronunciation do not meet the predetermined filtering rules. Contains in the pronunciation.

Implementing the embodiment provided by the present invention, when recognizing a speech signal, first obtain the feature classification result of the speech signal, and then filter the pronunciation contained in the feature classification result based on the probability contained in the feature classification result, then In the process of recognizing speech signals, there is no need to perform recognition operations related to the filtered pronunciation, such as no need to search for paths related to the filtered pronunciation in the recognition network, so the time spent on the speech recognition process can be effectively reduced , which can improve the efficiency of speech recognition.

Description of drawings

Fig. 1 is the flowchart of the voice recognition method shown in an exemplary embodiment of the present invention;

Fig. 2 is a flow chart of a voice recognition method shown in another exemplary embodiment of the present invention;

Fig. 3 is a logical block diagram of a speech recognition device shown in an exemplary embodiment of the present invention;

Fig. 4 is a logical block diagram of a speech recognition device shown in another exemplary embodiment of the present invention;

Fig. 5 is a hardware structural diagram of a speech recognition device according to an exemplary embodiment of the present invention.

detailed description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with aspects of the invention as recited in the appended claims.

The terminology used in the present invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein and in the appended claims, the singular forms "a", "the", and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise. It should also be understood that the term "and/or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of the present invention, first information may also be called second information, and similarly, second information may also be called first information. Depending on the context, the word "if" as used herein may be interpreted as "at" or "when" or "in response to a determination."

The speech recognition of the embodiment of the present invention involves an acoustic model and a language model in the recognition process, wherein the acoustic model is based on differences in acoustics, phonetics, environmental variables, and the gender and accent of the person who makes the speech. Knowledge representation can use LSTM (Long Short-Term Memory, time recurrent neural network), CTC (Connectionist temporal classification) model, or hidden Markov model HMM to train the speech contained in the training speech library to obtain the acoustics of the speech The mapping of features to utterances, constituting the acoustic model, is associated with the modeling unit. If the modeling unit is a syllable, the pronunciation is a syllable; if the modeling unit is a phoneme, the pronunciation is a phoneme; if the modeling unit is a state constituting a phoneme, the pronunciation is a state.

When training the acoustic model, considering that the pronunciation will vary with the factors affecting pronunciation such as words, speech speed, intonation, severity, and dialect, the training speech library needs to cover different words, speech speed, intonation, and severity. A large number of voices that affect pronunciation, such as phonetics and dialects. In addition, considering the accuracy of speech recognition, smaller pronunciation units such as syllables, phonemes, and states can be selected as modeling units. Therefore, performing model training based on a large number of speeches contained in the training speech library and predetermined modeling units will construct a large number of acoustic models. During the speech recognition process, a large number of acoustic models are used to classify the features of the speech signal to be recognized, and the obtained feature classification results will contain a large number of pronunciations (categories), such as: 3000 to 10000 pronunciations.

In addition, in order to recognize the text information corresponding to the speech signal, the current speech recognition technology needs to search all possible paths in the recognition network for each pronunciation, and an exponential path increment will be generated during this search process. If all possible paths involved in 3,000 to 10,000 pronunciations are searched in the recognition network, the storage resources and calculations required for the search may exceed the limit that the speech recognition system can bear. Therefore, the current speech recognition technology will consume a lot of time and resources, there is a problem of low speech recognition efficiency, and the present invention proposes a solution for how to improve the speech recognition efficiency.

In the solution of the present invention, in order to solve the problem of low speech recognition efficiency, the feature classification results obtained in the speech recognition process are improved, and the filtering rules are set in advance according to the equipment resources involved in speech recognition and the recognition efficiency requirements, and then when recognizing speech signals, First obtain the feature classification result of the speech signal, and then filter the pronunciation contained in the feature classification result based on the probability contained in the feature classification result, then in the process of recognizing the speech signal, there is no need to use the The path related to the filtered pronunciation is searched, so the time spent in the search process can be effectively reduced, and the efficiency of speech recognition can be improved. The voice recognition process of the present invention will be described in detail below in conjunction with the accompanying drawings.

Please refer to FIG. 1. FIG. 1 is a flow chart of a voice recognition method shown in an exemplary embodiment of the present invention. This embodiment can be applied to various electronic devices with voice processing capabilities, and may include the following steps S101-S103:

Step S101. Obtain a feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation feature of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation.

Step S102 , based on the probabilities contained in the feature classification result, filter the pronunciation contained in the feature classification result.

Step S103, identifying the speech signal based on the filtered feature classification result.

In the embodiment of the present invention, the voice signal may be the voice uttered by the user collected in real time by the local voice collection device, or may be the voice transmitted remotely by the voice collection device. When obtaining the feature classification result of the speech signal, the speech signal can be preprocessed by the speech preprocessing module in the art in real time, and the preprocessed speech signal is carried out feature extraction by the feature extraction module, and the extracted features can include PLP ( Perceptual LinearPredictive, perceptual linear prediction), LPCC (Linear Predictive Cepstral Coefficient, linear predictive cepstral coefficient), FBANK (Mel-Scale Filter Bank, Mel scale filter bank), MFCC (Mel-FrequencyCepstral Coefficients, Mel cepstral coefficient number), etc., and then perform corresponding processing on the extracted features through the acoustic model to obtain the feature classification result, and the probability contained in the feature classification result is used to represent the possibility of mapping the speech signal frame to the corresponding pronunciation. In other examples, feature classification results transmitted from other terminal devices may also be directly received.

After obtaining the feature classification result, the scheme of the present invention, considering the partial pronunciation contained in the feature classification result, has a lower correlation with the speech signal frame of the speech signal to be recognized, and has less impact on the accuracy of speech recognition, reducing the When the impact of a large number of pronunciations contained in the feature classification results on the speech recognition efficiency, these pronunciations that have little influence on the speech recognition accuracy can be filtered out from the feature classification results before performing speech recognition based on the feature classification results. The number of pronunciations contained in the feature classification result is reduced, thereby improving the speech recognition efficiency.

Generally, the lower the correlation between an utterance and the speech signal frame to be recognized, the lower the probability that the speech signal frame is mapped to the utterance when an acoustic model is used to classify the acoustic features of the speech signal. Therefore, the pronunciation contained in the feature classification result can be filtered based on the probability that the speech signal frame is mapped to each pronunciation. After filtering, the probability that any speech signal frame is mapped to the filtered pronunciation is less than the probability that the speech signal frame is mapped to other Probability of pronunciation.

In addition, when filtering less relevant pronunciations, considering the requirements of different application scenarios for the accuracy of speech recognition, it is also necessary to measure the impact of the filtered pronunciations on the accuracy of speech recognition. Therefore, according to the accuracy of speech recognition Requirements, pre-set various filtering rules that can limit the influence of filtered pronunciations on recognition accuracy. For various predetermined filtering rules, when filtering the pronunciation contained in the feature classification result, it is judged whether the probability that any speech signal frame is mapped to the corresponding pronunciation satisfies the predetermined filtering rule; if the corresponding pronunciation satisfies the predetermined filtering rule, all The above corresponding pronunciation is filtered out. Filtered pronunciations generally refer to pronunciations that are removed from the feature classification results.

Several ways to filter the pronunciation contained in the feature classification result are listed below:

Filtering method 1: filter out low-probability utterances by a predetermined number, the predetermined number may refer to the number of utterances that are retained in the feature classification results in the utterances corresponding to the speech signal frame; it may also refer to the predetermined ratio threshold corresponding to the speech signal frame The product of the total number of pronunciations. When filtering, if the probability that any speech signal frame is mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to each utterance in the predetermined number of utterances, then it is determined that the corresponding utterance satisfies a predetermined filtering rule.

Wherein, the predetermined ratio threshold can be set by the designer of the present invention according to the required speech recognition accuracy, for example, set to 1/4, which refers to the ratio of the retained pronunciation to all pronunciations.

In an example, during actual filtering, pronunciations can be deleted from the feature classification results in order of increasing probability, and when the ratio of the number of retained pronunciations to the number of all original pronunciations satisfies a predetermined ratio threshold, the feature classification is completed. Filtering of classification results.

In other examples, the predetermined ratio threshold may refer to the ratio of the number of unfiltered utterances to the number of filtered utterances. In actual filtering, pronunciations can be selected from the feature classification results in order of probability from large to small. When the ratio of the number of selected pronunciations to the number of remaining pronunciations meets the predetermined ratio threshold, the feature classification results are completed. filter.

In practical applications, when the predetermined number refers to the number of pronunciations that are retained in the feature classification result in the pronunciation corresponding to the frame of the speech signal frame, the designer of the present invention can set the predetermined number according to the speech recognition accuracy that needs to be achieved, for example , set it to any value from 2000 to 9000. When filtering, the pronunciations mapped to each speech signal frame can be arranged in order of probability from small to large, and then the pronunciations arranged in the front predetermined digits can be deleted from the feature classification results to complete the classification of the feature classification results For filtering, the predetermined number of digits is equal to the predetermined number of values.

In other examples, the predetermined number may refer to the number of utterances that are not filtered out, for example, set to 1000. In the actual filtering, the pronunciations mapped to each speech signal frame can be arranged in order of probability from large to small, and then the pronunciations arranged in the first predetermined digits are kept in the feature classification results, and other pronunciations are separated from the feature The classification result is deleted to complete the filtering of the feature classification result, and the predetermined number of digits is equal to the numerical value of the quantity threshold. In other embodiments, other technical means may also be used to filter the feature classification results in a filtering manner, which is not limited in the present invention.

Filtering method two: filter out low-probability pronunciations according to a predetermined difference threshold, which can be set by the designer of the present invention according to the speech recognition accuracy required to be achieved, for example, set to -0.5, means to be filtered The probability difference between the probability of the dropped utterance and the most probable utterance to which the same speech signal frame is mapped. When filtering, if any speech signal frame is mapped to the probability of the corresponding pronunciation, and the probability difference between the maximum mapping probability of the speech signal frame is within a predetermined difference range, then it is determined that the corresponding pronunciation satisfies the predetermined filtering rule, the corresponding pronunciation can be filtered.

In an example, during actual filtering, the pronunciations mapped to each speech signal frame may be arranged in descending order of probability, and the speech signal frame may be mapped to the probability of the pronunciation arranged first, It is determined as the maximum probability, and then starting from the last pronunciation, the difference between the probability that the speech signal frame is mapped to each pronunciation and the maximum probability is obtained sequentially. If the difference is less than -0.5, the pronunciation is classified from the feature Deleted from the results. In other embodiments, other technical means may also be used to filter the feature classification results according to the second filtering method, which is not limited in the present invention.

Filtering method three: filter the pronunciation distributed outside the beam width according to the histogram distribution of probability. During actual filtering, you can first obtain the histogram distribution of the probability that any speech signal frame is mapped to each pronunciation; The beam width corresponding to the graph distribution; then the pronunciation whose probability distribution is outside the beam width is determined as the pronunciation satisfying the predetermined filtering rule; finally, the pronunciation satisfying the predetermined filtering rule is determined from the feature classification result The included pronunciation is deleted. In practical applications, the beam width can be determined by the designer of the present invention according to the speech recognition accuracy required to be achieved and the distribution of the histogram, such as: presetting needs to filter out 8000 low-probability pronunciations, which can be obtained from the histogram The low-probability side starts to search for 8000 pronunciations, and determines the position of the 8000th pronunciation as the beam width boundary. In other embodiments, other technical means may also be used to filter the feature classification results according to the third filtering method, which is not limited in the present invention.

After filtering the pronunciation contained in the feature classification result according to any of the above filtering methods, the predetermined recognition network can be directly called to search for paths related to the pronunciation contained in the filtered feature classification result to find the best Based on this path, the text information corresponding to the speech signal to be recognized is output with the maximum probability to complete the speech recognition. The recognition network mentioned here can refer to the decoder according to the trained acoustic signal for the speech signal to be recognized. The recognition network built by model, language model and dictionary.

When looking for the best path, the probability (acoustic score) contained in the feature classification result can be converted to a similar probability (language score) between words (such as words, words, phrases) contained in the speech model Numerical space, and weighted addition, as the comprehensive score of the path search process, each speech signal frame will be limited by a preset threshold value, and the difference with the best path is greater than this threshold value, then the The path is discarded, otherwise it is kept; after the search of each voice signal frame is completed, all paths will be sorted according to the preset maximum number of paths, and only the optimal path of this number will be kept until the last frame is completed, thus the final road map.

In some examples, the modeling unit of the acoustic model that outputs the feature classification results is relatively small, such as using the state as the modeling unit. Since a single phoneme can be composed of three to five states, the speech signal formed by the pronunciation of a phoneme can be It is divided into multiple speech signal frames. Therefore, it is easy to have a situation where the acoustic features of multiple consecutive speech signal frames are relatively similar. Then, the pronunciation of each frame in these continuous speech signal frames is described in the feature classification result, and a similar situation is likely to occur. . For this situation, if the present invention is based on the probability contained in the feature classification result and the predetermined filtering rules, respectively filter the pronunciations mapped to each frame of the speech signal frame, it is easy to filter out the pronunciations that have a greater impact on the recognition accuracy, in order to To avoid filtering this type of pronunciation by mistake, when filtering the feature classification results, the filtering status of continuous speech signal frames can be considered comprehensively. The specific implementation process can refer to the method shown in Figure 2, including the following steps S201-S205:

Step S201. Obtain the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation feature of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation.

Step S202, if any speech signal frame is mapped to the probability of the corresponding utterance and satisfies a predetermined filtering rule, determine the utterance as a candidate utterance.

Step S203, if the probability of mapping to the candidate pronunciation in any frame of the predetermined number of adjacent speech signal frames of the speech signal frame satisfies a predetermined filtering rule, then select the candidate pronunciation from the feature classification results contained in the Pronunciation is deleted.

Step S204, if the probabilities of the predetermined number of adjacent speech signal frames of the speech signal frame being mapped to the candidate pronunciation do not satisfy the predetermined filtering rule, then keep the candidate pronunciation in the pronunciation contained in the feature classification result .

Step S205, identifying the speech signal based on the filtered feature classification result.

In the embodiment of the present invention, the predetermined filtering rule may be any rule involved in the above-mentioned filtering method 1 to filtering method 4, and may also be other filtering rules that can limit the impact of the filtered pronunciation on the recognition accuracy.

The predetermined number of consecutive speech signal frames can be set by the designers of the present invention according to the required speech recognition accuracy, for example, set to 6, the first three adjacent frames and the next three adjacent frames.

As can be seen from the foregoing embodiments: when the speech recognition method of the present invention recognizes a speech signal, the feature classification result of the speech signal is first obtained, and then based on the probability contained in the feature classification result, the pronunciation contained in the feature classification result is Filtering, then in the process of recognizing speech signals, there is no need to perform recognition operations related to the filtered pronunciation, such as no need to search for paths related to the filtered pronunciation in the recognition network, so it can effectively reduce the voice The time consumed by the recognition process can improve the efficiency of speech recognition.

Furthermore, the speech recognition method of the embodiment of the present invention can be applied to human-computer interaction software of various electronic devices, for example: voice dialing, voice control, and voice search in smart phones. When applied to voice search in smart phones, if If the user sends a voice within a predetermined range from the smart phone, the voice recognition method applied to the voice search can first obtain the feature classification result of the voice after receiving the user voice collected by the voice collection device, and then based on the described Probability contained in the feature classification results, filter the pronunciation contained in the feature classification results, and then only search for paths related to the pronunciation that have not been filtered out in the recognition network, and quickly identify the text information corresponding to the user's voice through path search , so that the voice assistant can quickly respond to the user based on the recognition result.

Corresponding to the foregoing method embodiments, the present invention also provides device embodiments.

Referring to FIG. 3 , FIG. 3 is a logic block diagram of a speech recognition device according to an exemplary embodiment of the present invention. The device may include: a classification result acquisition module 310 , a pronunciation filtering module 320 and a speech recognition module 330 .

Wherein, the classification result acquisition module 310 is used to obtain the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation characteristics of each speech signal frame and the mapping of each speech signal frame to the corresponding pronunciation. probability.

The pronunciation filtering module 320 is configured to filter the pronunciation contained in the feature classification result based on the probability contained in the feature classification result.

A speech recognition module 330, configured to recognize the speech signal based on the filtered feature classification result.

In some examples, the pronunciation filtering module 320 may include:

The first filtering module is used for mapping the probability of any speech signal frame to the corresponding pronunciation, and the probability difference between the maximum mapping probability of the speech signal frame, when within a predetermined difference range, the corresponding pronunciation to filter.

The second filtering module is configured to filter the corresponding utterance when the probability that any speech signal frame is mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to each utterance in the predetermined number of utterances.

In other examples, the pronunciation filtering module 320 may also include:

The probability distribution module is used to obtain the histogram distribution of the probability that any speech signal frame is mapped to each pronunciation.

A beam width determining module, configured to acquire the beam width corresponding to the histogram distribution.

An utterance determining module, configured to determine an utterance whose probability distribution is outside the beam width as an utterance satisfying the predetermined filter rule.

A pronunciation deletion module, configured to delete pronunciations satisfying the predetermined filter rule from the pronunciations included in the feature classification result.

Referring to FIG. 4 , FIG. 4 is a logic block diagram of a speech recognition device according to another exemplary embodiment of the present invention. The device may include: a classification result acquisition module 410 , a pronunciation filtering module 420 and a speech recognition module 430 . The pronunciation filtering module 420 may include a candidate pronunciation determination module 421 , a candidate pronunciation deletion module 422 and a candidate pronunciation retention module 423 .

Wherein, the classification result acquisition module 410 is used to obtain the feature classification result of the speech signal to be recognized; the feature classification result includes the pronunciation used to describe the pronunciation characteristics of each speech signal frame and the mapping of each speech signal frame to the corresponding pronunciation. probability.

The candidate utterance determining module 421 is configured to determine the utterance as a candidate utterance when the probability that any speech signal frame is mapped to the corresponding utterance satisfies a predetermined filter rule.

The candidate utterance deletion module 422 is configured to remove the candidate utterance from the feature when the probability of mapping to the candidate utterance satisfies a predetermined filtering rule in any frame of the predetermined frame number of adjacent voice signal frames of the voice signal frame. Pronunciations included in the classification results are deleted.

The candidate utterance retention module 423 is configured to retain the candidate utterance in the feature classification result when the probabilities of the predetermined number of frames of the speech signal frame being mapped to the candidate utterance do not satisfy the predetermined filtering rules. included in the pronunciation.

A speech recognition module 430, configured to recognize the speech signal based on the filtered feature classification result.

For the implementation process of the functions and effects of each unit (or module) in the above device, please refer to the implementation process of the corresponding steps in the above method for details, and will not be repeated here.

As for the device embodiment, since it basically corresponds to the method embodiment, for related parts, please refer to the part description of the method embodiment. The device embodiments described above are only illustrative, and the units or modules described as separate components may or may not be physically separated, and the components shown as units or modules may or may not be physical units Or modules, which can be located in one place, or can be distributed to multiple network units or modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. It can be understood and implemented by those skilled in the art without creative effort.

Embodiments of the speech recognition device of the present invention can be applied to electronic equipment. Specifically, it can be realized by computer chips or entities, or by products with certain functions. In a typical implementation, the electronic device is a computer, and the specific form of the computer may be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email sending and receiving device , gaming consoles, tablet computers, wearables, internet-connected TVs, smart bikes, driverless cars, smart refrigerators, other smart home devices, or any combination of these.

The device embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by reading the corresponding computer program instructions in a readable medium such as a non-volatile memory into the memory for execution by the processor of the electronic device where it is located. From the hardware level, as shown in Figure 5, it is a hardware structural diagram of the electronic device where the speech recognition device of the present invention is located, except for the processor, memory, network interface, and non-volatile memory shown in Figure 5 The electronic device where the device in the embodiment is located usually may also include other hardware according to the actual function of the electronic device, which will not be repeated here. The memory of the electronic device can store program instructions executable by the processor; the processor can be coupled to the memory for reading the program instructions stored in the memory, and in response, perform the following operations: obtain the feature classification result of the speech signal to be recognized The feature classification result includes the pronunciation used to describe the pronunciation features of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation; based on the probability contained in the feature classification result, the feature classification result contains filtering the pronunciation of containing; and identifying the speech signal based on the filtered feature classification result.

In other embodiments, for operations performed by the processor, reference may be made to relevant descriptions in the foregoing method embodiments, and details are not repeated here.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the present invention. within the scope of protection.

Claims (10)

1. a kind of audio recognition method, it is characterised in that including step：

Obtain the tagsort result of voice signal to be identified；The tagsort result is comprising for describing each voice signal The pronunciation of the pronunciation character of frame and each speech signal frame are mapped to the probability of corresponding pronunciation；

Based on the probability contained by the tagsort result, the pronunciation contained by the tagsort result is filtered；

The voice signal is recognized based on the tagsort result after filtering.

2. method according to claim 1, it is characterised in that the probability based on contained by the tagsort result, Pronunciation contained by the tagsort result is filtered, including：

Judge that whether any speech signal frame is mapped to the probability of corresponding pronunciation and meets predetermined filtering rule；

If the corresponding pronunciation meets predetermined filtering rule, the corresponding pronunciation is filtered.

3. method according to claim 2, it is characterised in that：

If any speech signal frame is mapped to the probability of corresponding pronunciation, and the maximum mapping probabilities of the speech signal frame between Probability difference, in predetermined difference range, it is determined that it is described it is corresponding pronunciation meet predetermined filtering rule；

If any speech signal frame is mapped to the probability of corresponding pronunciation, predetermined number is mapped to less than the speech signal frame The probability of each pronunciation in pronunciation, it is determined that the corresponding pronunciation meets predetermined filtering rule.

4. method according to claim 3, it is characterised in that the predetermined number is following any：

The quantity of the pronunciation being retained in the corresponding pronunciation of frame speech signal frame in tagsort result；

The product of the total number of predetermined proportion threshold value pronunciation corresponding with the frame speech signal frame.

5. method according to claim 1, it is characterised in that the probability based on contained by the tagsort result, Pronunciation contained by the tagsort result is filtered, including：

Obtain the histogram distribution that any speech signal frame is mapped to the probability of each pronunciation；

Obtain beamwidth corresponding with the histogram distribution；

By pronunciation of the probability distribution outside the beamwidth, it is defined as meeting the pronunciation of predetermined filtering rule；

The pronunciation of predetermined filtering rule will be met, deleted from the pronunciation contained by the tagsort result.

6. method according to any one of claim 1 to 5, it is characterised in that described based on the tagsort result Contained probability, filters to the pronunciation contained by the tagsort result, including：

If the probability that any speech signal frame is mapped to corresponding pronunciation meets predetermined filtering rule, the pronunciation is defined as to wait Publish sound；

If any frame in the adjacent speech signal frame of the predetermined frame number of the speech signal frame, the general of candidate pronunciation is mapped to Rate meets predetermined filtering rule, then candidate pronunciation is deleted from the pronunciation contained by the tagsort result；

If the adjacent speech signal frame of the predetermined frame number of the speech signal frame, the probability for being mapped to candidate pronunciation is unsatisfactory for Predetermined filtering rule, then in pronunciation candidate pronunciation being retained in contained by the tagsort result.

7. a kind of speech recognition equipment, it is characterised in that including：

Classification results acquisition module, the tagsort result for obtaining voice signal to be identified；The tagsort result Pronunciation and each speech signal frame comprising the pronunciation character for describing each speech signal frame are mapped to the general of corresponding pronunciation Rate；

Voice synthesizing filter module, for based on the probability contained by the tagsort result, to contained by the tagsort result Pronunciation is filtered；

Sound identification module, for recognizing the voice signal based on the tagsort result after filtering.

8. device according to claim 7, it is characterised in that the voice synthesizing filter module also includes：

First filtering module, the probability for being mapped to corresponding pronunciation in any speech signal frame, with the speech signal frame Probability difference between maximum mapping probabilities, when in predetermined difference range, filters to the corresponding pronunciation；

Second filtering module, the probability for being mapped to corresponding pronunciation in any speech signal frame, less than the speech signal frame When being mapped to the probability of each pronunciation in the pronunciation of predetermined number, the corresponding pronunciation is filtered.

9. device according to claim 7, it is characterised in that the voice synthesizing filter module includes：

Probability distribution module, for obtaining the histogram distribution that any speech signal frame is mapped to the probability of each pronunciation；

Beamwidth determining module, for obtaining beamwidth corresponding with the histogram distribution；

Pronunciation determining module, for the pronunciation by probability distribution outside the beamwidth, is defined as meeting the predetermined filtering rule Pronunciation then；

Pronunciation removing module, for the pronunciation contained by the pronunciation from the tagsort result of the predetermined filtering rule will to be met Middle deletion.

10. the device according to any one of claim 7 to 9, it is characterised in that the voice synthesizing filter module includes：

Candidate's pronunciation determining module, the probability for being mapped to corresponding pronunciation in any speech signal frame meets predetermined filtering rule When then, the pronunciation is defined as candidate's pronunciation；

Candidate's pronunciation removing module, for any frame in the adjacent speech signal frame of the predetermined frame number of the speech signal frame, When the probability for being mapped to candidate pronunciation meets predetermined filtering rule, by candidate pronunciation from contained by the tagsort result Deleted in pronunciation；

Candidate's pronunciation reservation module, for the adjacent speech signal frame of the predetermined frame number in the speech signal frame, is mapped to the time When the probability for publishing sound is unsatisfactory for predetermined filtering rule, candidate pronunciation is retained in the hair contained by the tagsort result In sound.