JP2003345388A - Method, device, and program for voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce recognition error in a device for voice recognition. <P>SOLUTION: A voice signal input means 101 receives a voice signal, and a recognition result candidate generation means 102 outputs N word strings, word duration length sequences respectively corresponding to N word strings, and N scores from preliminarily prepared sound models and a word dictionary. A duration length normalizing means 103 normalizes the duration lengths on the basis of a parameter prescribing linear function mapping, which is estimated by a normalization constant calculation means 105, by mapping prescribed by a linear function in accordance with speaking speed. A recognition result candidate correction means 104 corrects scores of each of the candidates of a recognition result in accordance with likelihoods of arrays of duration lengths of recognition result candidate word strings. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, a voice recognition method, and a voice recognition program, and more particularly to a voice recognition device for recognizing continuously uttered voices and converting them into a symbol string representing the utterance contents. Voice recognition method, and
Regarding speech recognition programs.

[0002]

2. Description of the Related Art Conventionally, a recording medium in which a voice recognition device and a voice recognition program of this kind are recorded is mainly a standard pattern of phonemes (vowels a, i, u, ..., Consonants k, s, t, ...). As a one-way type (left-to-right
Type) Hidden Markov Model (HMM) is prepared, and by repeating matching between these standard patterns and each part of the input speech signal, a phoneme string that best matches the input speech signal is derived and uniquely determined from this phoneme string. A symbol string (for example, a word string) is output as a recognition result.

In the voice recognition technology based on HMM, an arbitrary part of a voice signal can be a target of matching, so that an extremely long time section and a consonant often match, and a short time section and a vowel often match. obtain. However, in many cases, a vowel lasts for a relatively long time and a consonant ends in a short time, so the above phenomenon may cause a recognition error.

Generally, the length of time during which various units such as vowels, consonants, syllables and words continue in a voice signal is called a duration time. HMM-based speech recognition technology does not have a mechanism for explicitly controlling the duration. This is HMM
Was one of the big drawbacks. As one of the solutions to this drawback, there is a method of introducing a duration distribution model as shown in the document "1988, Seiichi Nakagawa, Speech Recognition by Stochastic Model, IEICE".

This is a method of controlling the duration by incorporating the probability distribution model of the duration of speech into the Viterbi search for obtaining the optimum state transition sequence of the HMM and calculating the likelihood. More simply, a plurality of recognition result candidates are calculated by a normal Viterbi search (first pass), and then the plausibility of the sequence of durations of each candidate is calculated based on the probability distribution model of durations. Then
A method of performing a two-step procedure of correcting the score (likelihood) of each candidate (second pass) can also be used instead. As the probability distribution, an average value estimated by learning in advance,
A normal distribution defined by the variance value is used.

FIG. 5 is a block diagram showing an example of an embodiment of a voice recognition apparatus in which the latter method of performing the two-step procedure among the above-mentioned conventional techniques is introduced.

In this example, a voice signal input means 201 for receiving a voice signal inputted from the outside, a time series of voice signals from the voice signal input means 201, and a series of feature quantities (for example, mel cepstrum) useful for recognition. Calculates and prepares an acoustic model (set of standard patterns)
And, from the grammar (word dictionary) prepared as necessary, the standard pattern (phoneme) that best matches the feature quantity sequence.
N sequences are sequentially obtained from the top by a Viterbi search or the like, and N symbol sequences (a sequence of characters and words) uniquely determined from the obtained N phoneme sequences and N symbol sequences are respectively corresponded. A sequence of symbol durations, a recognition result candidate generation unit 202 that outputs N scores (likelihoods),
The symbol string, the duration sequence, and the score are received from the recognition result candidate generating unit 202, and the score representing the likelihood of the duration is calculated based on the duration distribution model.
The recognition result candidate correction means 203 for correcting the score by adjusting the score output from the recognition result candidate generation means 202, and the duration time referred to by the recognition result candidate correction means 203 for calculating the score representing the plausibility of the duration time. Duration length statistic storage means 204 for storing the average length and variance value
Equipped with.

The contents of information received by the recognition result candidate correction means 203 from the recognition result candidate generation means 202 and the processing procedure of the received information will be described in more detail.
The N recognition result candidates generated by the recognition result candidate generating unit 202 are represented by a set of a symbol string (for example, a word sequence), a symbol duration length sequence, and a score (likelihood). _Let one of the sets be (W _{k, 1} , W _{k, 2} , ..., W _{k, n (k)} ; X
_{k, 1, X k, 2} , ..., X k, n (k); represents _{L k)} and is referred to as a k-th candidate. Recognition result candidate correction means 203
Receives N sets (first to Nth candidates) from the recognition result candidate generation means 202. Where k = 1,
2, ..., N.

Also, L₁> L_Two＞… ＞ L_NAnd W
_{k, 1}, W_{k, 2},…, W_{k, n (k )}Is the recognition result
A symbol string (for example, a series of words) that is a substance of the k-th candidate.
X _{k, 1}, X_{k, 2}, ..., X_{k, n (k)}Is the sign
It is a duration sequence corresponding to a column one-to-one. L_kIs
Viewpoint of acoustic model (match between input speech signal and standard pattern)
Is the score or likelihood calculated from
It represents the acoustic plausibility of the kth candidate. for example
For example, refer to the document “Lawrence Rabiner et al. (1995, Furui
(Translated by Teihi), Basics of speech recognition (below), NTT Advanced
Kunoroji Co., Ltd., pp. 188-194 "_kOf calculation
Examples are given.

The recognition result candidate correction means 203 is provided for each recognition result.
Verification of plausibility of fruit candidates from the viewpoint of duration
And the score L of the kth candidate according to the result_kTo correct
It That is, each element W of the k-th candidate symbol string of the recognition result
_{k, i}Duration X corresponding to _{k, i}Is the average value μ
(W_{k, i}) Score value L_kDownward correction
To do. Specifically, the score is modified according to the following formula (Formula 1).
To correct.

[0011]

[Equation 1]

Where μ (W _{k, i} ) and σ
(W _{k, i} ) is the average and variance of the durations of the symbols (eg, words) W _{k, i} that make up the recognition result symbol string, and the average μ (W) and variance σ (W) for every element W. Are stored in the duration statistics storage means 204. Further, α is a weighting coefficient (constant) for the plausibility of the duration. Recognition result candidate correction means 20
3 reads out and uses the values of μ (W) and σ (W) from the duration statistics storage means 204 as needed.
Note that these values are calculated and stored in advance by performing a Viterbi search on large-scale learning data.

However, the above-mentioned conventional technique has a problem that the robustness is lost when the utterance speed is extremely fast or slow, or when sudden noises overlap before and after the utterance.

Generally, the statistics μ (W) and σ (W) stored in the duration statistics storage means 204 are averages calculated from durations extracted from voice data of various speech rates, and Since it is a variance, it is a correctly estimated statistic for utterances with a standard speech rate, but is not an accurate estimator for utterances that are out of the standard. If the score of Formula 1 is corrected using an inaccurate statistic, the score is corrected inappropriately, which causes a recognition error. That is, the statistics μ (W) and σ (W) are appropriately adjusted according to the fluctuation of the speech rate (speech rate normalization).
There is a need to.

Regarding the fluctuation of the speech rate described above, one measure for ensuring robustness is described in "Japanese Patent Laid-Open No. 11-311".
The technology of "information processing device and method, and providing medium" described in Japanese Patent Laid-Open No. 994 "is known. This technique defines "the ratio of the duration of each word and the sum of the duration of each word" to the "normalized duration",
This is to ensure a certain degree of robustness against fluctuations in speech rate.

[0016]

However, in the technique described in the above publication, when sudden noises are overlapped before and after utterance and the duration of some words is extremely disturbed, Robustness can be lost. For example, when noise overlaps immediately before the utterance, only the duration of the first word becomes extremely long, and the “total duration of each word” significantly increases. Then, the normalization is performed by dividing by a value larger than the original value, and the length of the duration of each word after the normalization becomes smaller than the original value. As a result, the normalization of the duration is erroneously performed, and the robustness against the speech rate is impaired. It is clear that a similar problem occurs when the noise overlaps immediately after the utterance. As is well known, this type of problem is especially likely to occur when the utterance contains a small number of words.

On the other hand, in the prior art disclosed in the above publication, it may be a problem that the duration is assumed to be proportional to the speech rate. In this technique, the duration length of each word is normalized by multiplying the duration length of each word by a constant coefficient that is the reciprocal of the total duration length of each word. However, words are generally composed of vowels and consonants. Vowels directly fluctuate and have a proportional relationship with speech rate, but consonants do not depend on speech rate and can maintain a constant value. Are known. That is, in general, the duration is a linear component that changes depending on the speech rate,
It is considered to consist of a bias component that does not depend on the speech rate. Therefore, it is difficult to realize proper normalization of the duration without combining the procedure of multiplying by a constant (multiplication) and the procedure of adding another constant (addition).

As described above, even if the duration of some words is extremely disturbed, such as when noise is overlapped before and after the utterance, the accuracy is not deteriorated, and the part that depends on the utterance speed is not affected. There is a demand for a technique that normalizes the duration time taking into consideration both the non-dependence part and the normalization time, appropriately digitizes the plausibility of the duration time, and reflects it in the speech recognition score.

[0019]

[Means for Solving the Problems] As a general tendency, when noises overlap before and after utterance, the duration of the first word or the last word may increase, but Other than, the duration of each word located in the middle of the utterance is hardly affected by noise. Therefore, if the coefficients for duration normalization are determined so that the deviations from the average of the durations of each word included in the utterance are equally reduced, the durations of some words will be Even if it is incorrect, it is expected that the influence of the incorrect duration will be mitigated by using the duration of the remaining words, and robust normalization can be performed. Furthermore, it is expected that the normalization procedure will be a process of multiplying a certain constant and adding another constant to normalize the duration length in accordance with the phenomenological theory of the fluctuation in speech rate.

Therefore, in order to solve the above-mentioned problems,
The voice recognition device and the recording medium recording the voice recognition program according to the present invention include, in addition to the conventional art, a unit for normalizing the duration and a unit for normalizing the duration normalize the duration. A means for calculating the normalization constants (proportionality constant and bias constant) necessary for the calculation is further provided, and the means for calculating the normalization coefficient is prepared in advance with the recognition result word string, the duration sequence, and Using the duration statistics, the normalization constant is calculated based on the square error minimization so that the difference between the duration of each word and its average value becomes as small as possible.

A first voice recognition apparatus of the present invention includes a voice signal input means for receiving a voice signal inputted from the outside, a set of standard patterns prepared in advance for receiving the voice signal from the voice signal input means, and Based on the grammar, a plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and Recognition result candidate generating means for outputting a score corresponding to each word string, and a word string, a duration length series and a score from the recognition result candidate generating means, a word string,
Duration length normalizing means for normalizing the duration time using the normalization constant from the duration time, and the average output of the recognition result candidate generating means, the correction of the score based on the variance, Recognition result candidate correction means for outputting, and a normalization constant calculation means for calculating a normalization constant used by the continuation time length normalization means for normalization of the continuation time length based on the average and variance of the continuation time length, Duration duration statistic storage means for storing the average and variance of durations used when the recognition result candidate modifying means calculates a score and when the normalization constant calculating means calculates a normalization constant; , Are provided.

A second voice recognition apparatus of the present invention is a voice signal input means for receiving a voice signal inputted from the outside, a set of standard patterns prepared in advance for receiving the voice signal from the voice signal input means, and Based on the grammar, a plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and Recognition result candidate generating means for outputting a score corresponding to each word string, and a word string, a duration length series and a score from the recognition result candidate generating means, a word string,
Duration length normalizing means for normalizing the duration time using the normalization constant from the duration time, and the average output of the recognition result candidate generating means, the correction of the score based on the variance, The recognition result candidate correction means for outputting, a first type normalization constant for multiplying the duration time by a constant, and a second type normalization constant for performing a constant addition of the duration time are the duration time normalization means. Is an average of the duration as a normalization constant used for normalization of the duration, the normalization constant calculating means for calculating based on the variance, when the recognition result candidate correction means calculates the score, and, And a duration constant statistic storage means for storing the average and the variance of the duration used when the normalization constant calculation means calculates the normalization constant.

A third voice recognition apparatus of the present invention comprises a voice signal input means for receiving a voice signal input from the outside, a set of standard patterns prepared in advance by receiving the voice signal from the voice signal input means, and Based on the grammar, a plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and Recognition result candidate generating means for outputting a score corresponding to each word string, and a word string, a duration length series and a score from the recognition result candidate generating means, a word string,
Duration length normalizing means for normalizing the duration time using the normalization constant from the duration time, and the average output of the recognition result candidate generating means, the correction of the score based on the variance, The recognition result candidate correction means for outputting and a normalization constant used by the duration length normalization means for normalizing the duration length have a minimum squared error between the duration length series and the average of the duration lengths. Normalization constant calculation means for calculating the average length and variance of the durations so that when the recognition result candidate correction means calculates the score, and the normalization constant calculation means calculates the normalization constant And a duration length statistic storage means for storing the average and the variance of the duration lengths used when performing.

The first voice recognition method of the present invention comprises a first procedure for receiving a voice signal input from the outside, and the first procedure described above.
, A sequence of standard patterns prepared in advance and a grammar are used to find a plurality of word strings with a high score that match the voice signal in order from the top. A second sequence for outputting a word duration sequence corresponding to each of the word strings and a score corresponding to each of the plurality of word sequences, and a word sequence, a duration sequence, and a score from the second procedure. A third procedure for receiving and normalizing the duration using a normalization constant from the word string and the duration, and correcting the score output by the second procedure based on the average and variance of the durations And outputting a fourth procedure, a fifth procedure for calculating a normalization constant used by the third procedure for normalizing the duration, based on the average and variance of the duration, and Score of 4 steps When computing, and, characterized in that it comprises a sixth step of storing duration mean, the variance used in the fifth step calculates a normalization constant.

A second voice recognition method of the present invention comprises a first procedure for receiving a voice signal input from the outside, and the first procedure described above.
, A sequence of standard patterns prepared in advance and a grammar are used to find a plurality of word strings with a high score that match the voice signal in order from the top. A second sequence for outputting a word duration sequence corresponding to each of the word strings and a score corresponding to each of the plurality of word sequences, and a word sequence, a duration sequence, and a score from the second procedure. A third procedure for receiving and normalizing the duration using a normalization constant from the word string and the duration, and correcting the score output by the second procedure based on the average and variance of the durations Then, the fourth procedure for outputting, the first type normalization constant for multiplying the duration time by a constant, and the second type normalization constant for performing the constant addition of the duration time are added by the third procedure. Normalization of duration The fifth procedure for calculating a score based on the average and variance of the durations as the normalization constant used for the above, when the fourth procedure calculates the score, and the fifth procedure for calculating the normalization constant. And a sixth step of storing the average of the duration lengths used at the time and the variance.

A third voice recognition method of the present invention comprises a first procedure for receiving a voice signal input from the outside, and the first procedure described above.
, A sequence of standard patterns prepared in advance and a grammar are used to find a plurality of word strings with a high score that match the voice signal in order from the top. A second sequence for outputting a word duration sequence corresponding to each of the word strings and a score corresponding to each of the plurality of word sequences, and a word sequence, a duration sequence, and a score from the second procedure. A third procedure for receiving and normalizing the duration using a normalization constant from the word string and the duration, and correcting the score output by the second procedure based on the average and variance of the durations Then, the fourth procedure for outputting and the normalization constant used by the third procedure for normalizing the duration length are the minimum squared error between the duration sequence and the average duration. When to continue to be When the average length, the fifth step of calculating, based on the dispersion, the fourth step calculates the score,
And, the fifth procedure includes a sixth procedure of storing an average and a variance of the durations used when calculating the normalization constant.

A first voice recognition program of the present invention includes a first procedure for receiving a voice signal input from the outside, a set of standard patterns prepared in advance by receiving the voice signal from the first procedure, and , Based on the grammar
A plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and a plurality of word strings are respectively corresponded. The second step of outputting the score,
A third procedure for receiving a word string, a duration sequence and a score from the second procedure and normalizing the duration using the normalization constant from the word sequence and the duration;
Of the procedure of (4), which corrects the score output based on the average and variance of the duration, and outputs the fourth procedure, and the normalization constant used by the third procedure for normalizing the duration. A fifth step of calculating based on the mean and variance of the length,
A sixth step of storing an average and a variance of durations used when the fourth procedure calculates a score and when the fifth procedure calculates a normalization constant. It is characterized by

A second voice recognition program of the present invention includes a first procedure for receiving a voice signal input from the outside, a set of standard patterns prepared in advance by receiving the voice signal from the first procedure, and , Based on the grammar
A plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and a plurality of word strings are respectively corresponded. The second step of outputting the score,
A third procedure for receiving a word string, a duration sequence and a score from the second procedure and normalizing the duration using the normalization constant from the word sequence and the duration;
The score output by the procedure of (1) is corrected based on the average and variance of the duration, and is output, and the first type of normalization constant that performs a constant multiple of the duration and the duration A fifth procedure for calculating a second type normalization constant for performing constant addition as a normalization constant used by the third procedure for normalization of the duration, based on the average and variance of the duration, A sixth step of storing an average and a variance of durations used when the fourth procedure calculates a score and when the fifth procedure calculates a normalization constant. It is characterized by

A third voice recognition program of the present invention includes a first procedure for receiving a voice signal input from the outside, a set of standard patterns prepared in advance by receiving the voice signal from the first procedure, and , Based on the grammar
A plurality of word strings with a high score that match the voice signal are obtained in order from the top, and a plurality of word strings, a word duration sequence corresponding to each of the plurality of word strings, and a plurality of word strings are respectively corresponded. The second step of outputting the score,
A third procedure for receiving a word string, a duration sequence and a score from the second procedure and normalizing the duration using the normalization constant from the word sequence and the duration;
The procedure for correcting the output score of the procedure (1) on the basis of the average and variance of the duration and outputting the fourth procedure, and the normalization constant used by the third procedure for normalizing the duration A fifth step of calculating based on the average and variance of the durations so that the squared error between the time series and the average of the durations is minimized, and the fourth procedure when calculating the score. , And the sixth step of storing the average and variance of the duration used in the calculation of the normalization constant in the fifth procedure, and the sixth procedure.

[0030]

BEST MODE FOR CARRYING OUT THE INVENTION Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of the first embodiment of the present invention.

Referring to FIG. 1, in the first embodiment of the present invention, an audio signal input means 101 for receiving an audio signal inputted from the outside and a time series of the audio signal are received from the audio signal input means 101, A series of feature quantities (eg, mel cepstrum) useful for recognition is calculated, and a feature quantity series is created from a prepared acoustic model (set of standard patterns) and a grammar (word dictionary) prepared as necessary. A sequence of standard patterns (phonemes) that best matches is found in order from the top by a Viterbi search or the like, and it corresponds to each of the N word strings and N word strings that are uniquely determined from the N phoneme strings obtained. Recognition result candidate generation means 102 for outputting a word duration sequence and a score (likelihood) of N pieces, and recognition result candidate generation means 102.
From the word string and the duration length sequence, the score is received from, and the duration length normalizing means 103 for normalizing (numerical alignment) the duration length from the word string and the duration length, and the continuation based on the duration length distribution model Recognition result candidate correction means 10 for calculating a score representing the plausibility of the time length and correcting the score by adding or subtracting the score generated by the recognition result candidate generation means 102.
4, the normalization constant calculation means 105 for calculating a normalization constant required when the duration length normalization means 103 normalizes the duration time, and the recognition result candidate correction means 104 When calculating the score that expresses likeness,
Further, the normalization constant calculation means 105 is composed of a continuous time length statistic storage means 106 (for example, a memory) for storing an average value or a dispersion value of the continuous time length referred to when calculating the normalization constant.

Each means can be realized by operating as a program stored in a computer.

Next, the operation of the first embodiment of the present invention will be described with reference to the drawings.

The audio signal input means 101 is provided with an input device such as a microphone, and fetches and stores time series data of the input audio signal.

The recognition result candidate generating means 102 is an acoustic model (a set of standard pattern HMMs) learned in advance using appropriate voice data and a correct answer word string corresponding thereto.
And a grammar (dictionary that defines the arrangement of words) prepared according to the recognition task, receives time-series data of a voice signal from the voice signal input means 101, and uses this for a feature vector useful for recognition (for example, mel Cepstral) sequence, and using the acoustic model and the grammar, a sequence of standard patterns (for example, phonemes) that best matches the feature vector sequence and matches the grammar is sequentially searched from the top by Viterbi search or the like. A plurality (here, N) is obtained, N word strings uniquely determined from the obtained N phoneme strings, word duration sequences corresponding to the N word strings, and N word strings. The score (likelihood) of is output.

In the following, N sets of these word strings, word duration sequences, and scores will be represented by (W _{k, 1} , W _{k, 2} , ...,
W _{k, n (k)} ; X _{k, 1} , X _{k, 2} , ..., X
It is represented by the symbol _{k, n (k)} ; L _k ). Where k =
1, 2, ..., N, a word string, a word duration sequence,
It is assumed that each set of scores is sorted in descending order of the score L _k . n (k) is the number of words in the kth candidate.

The continuation time length normalizing means 103 has the upper N recognition result candidates (W) than the recognition result candidate generating means 102.
_{k, 1} , W _{k, 2} , ..., W _{k, n (k)} ; X _{k, 1} , X
_{k, 2} , ..., _{Xk, n (k)} ; _Lk ), k = 1,2,
, N is received, and the duration is normalized so that the variation factor caused by the speech rate becomes smaller one by one.

The normalization here is a mapping by a linear function defined by two kinds of parameters (normalization constants) κ and λ as in the following expression (Expression 2). κ is a normalization constant that scales the duration (constant multiple), and λ is
It is a constant normalization constant that shifts the duration (constant addition).

[0040]

[Equation 2]

The values of the normalization constants κ and λ are determined by the normalization constant calculation means 105 prior to normalization. The procedure is as follows.

The duration normalizing means 103 is a k-th candidate word string W _{k, 1} , W _{k, 2} , ..., W _{k, n (k)} ,
And the duration length sequence X _{k, 1} , X _{k, 2} , ..., X
_{k, n (k)} is sent to the normalization constant calculation means 105. On the other hand, the average duration duration and variance of all words are prepared and stored in advance in the duration statistics storage means 106.

The normalization constant calculating means 105 receives the above-mentioned word string and duration length sequence from the duration length normalizing means 103, and further, duration length statistic storage means 106.
From the word sequence W _{k, 1} , W _{k, 2} , ..., W _{k, n (k)}
Of the average duration μ (W _{k, 1} ), μ (W
_{k, 2} ), ..., μ (W _{k, n (k)} ) and variance σ (W _k,
1 ), σ (W _{k, 2} ), ..., σ (W _{k, n (k)} ) are read. Then, the normalization constant calculation means 105 calculates the normalization constants κ and λ according to the following expressions (Expression 3 and Expression 4).

[0044]

[Equation 3]

[0045]

[Equation 4]

These calculation formulas are used for the duration length series X
_{k, 1} , X _{k, 2} , ..., X _{k, n (k)} and the average sequence μ (W _{k, 1} ), μ (W _{k, 2} ), ...
The square error between (W _{k, n (k)} ) is set to be minimum. More precisely, it is designed so that the amount representing the plausibility of the duration time (the above-mentioned mathematical expression 1 or the second term on the right side of the following mathematical expression 5) becomes maximum. The normalization constant calculation means 105 passes the normalization constants κ and λ calculated according to these calculation formulas to the continuation time length normalization means 103, and the continuation time length normalization means 103 continues according to the above-mentioned mathematical expression 2. Normalize the time length.

The recognition result candidate correction means 104 receives N recognition result candidates (W
_{k, 1} , W _{k, 2} , ..., W _{k, n (k)} ; X _{k, 1} , X
_{k, 2} , ..., _{Xk, n (k)} ; _Lk ), k = 1,2,
, N is received (note that L _k is already normalized), and the words W _{k, 1} , W _{k, 2} , ..., W _{k, n (} It reads the duration of the mean and variance corresponding to _k), the score L _k of this, modified based on the plausibility of the normalized duration.

The correction method is the same as the mathematical expression 1 (the mathematical expression 5 described later) shown in the prior art. Further, the recognition result candidate correction means 104 re-sorts the N recognition result candidates in the descending order of the corrected score L _k , and ranks the first recognition result candidate (or a plurality of higher recognition result candidates as necessary). Is output as the final recognition result.

Next, the operation of the first embodiment of the present invention will be described in more detail with reference to the drawings.

FIG. 2 is an explanatory view showing an example of correction of recognition result candidates by score correction based on the word duration length.

Referring to FIG. 2, such as the address part of an address
An arbitrary number of digits that often appear in recognition and "no" are displayed alternately.
It is the recognition result when utterances that
The actual utterance content and the recognition result candidate generation means 102
Output the recognition result candidates before correction (up to the 4th candidate),
Corrected recognition result output by the knowledge result candidate correction unit 104
It is a candidate (the same). Each recognition result candidate has its
Core L_kWith the word W_{k, 1}, W_{k, 2},…, W
_{k, n (k)}, And duration X_{k, 1}，
X _{k, 2}, ..., X_{k, n (k)}[Msec unit] pair
It is written.

For example, the 1st place candidate of the recognition result candidate (before correction) has the word [duration length] of [1] of [n] [667], Ms. [444], [466], and [436].
89] and 1 [421].

FIG. 3 is an explanatory diagram showing an example of the word duration statistics. Referring to FIG. 3, there are shown average and variance values of the duration length of each word that the duration statistic storage means 106 has. In this example, the units of mean and variance are msec and square msec.

The normalization constant calculation means 105 calculates the normalization constants κ and λ for each recognition result candidate based on Expressions 3 and 4.

For example, in the first candidate, κ =
{(1/10898 + 1/5254 + 1/1526 + 1
/ 1526 + 1/5123) × (667 × 442/10
898 + 466 x 160/1526 + 436 x 285 /
4530 + 189 × 160/1526 + 421 × 351
/ 5123)-(667/10898 + 444/525)
4 + 466/1526 + 436/4530 + 189/1
526 + 421/5123) x (442/10898 +
294/5254 + 160/1526 + 285/453
0 + 160/1526 + 351/5123)} / {(1
/ 10898 + 1/5254 + 1/1526 + 1/15
26 + 1/5123) x (6672/10898 + 44
Four^Two/ 5254 + 466^Two/ 1526 + 436^Two/ 45
30 + 189 ^Two/ 1526 + 421^Two/ 5123)-
(667/10898 + 444/5254 + 466/1
526 + 436/4530 + 189/1526 + 421
/ 5123)^Two} = 0.334.

Further, λ = (442/10898 + 294)
/ 5254 + 160/1526 + 285/4530 + 1
60/1526 + 351/5123) / (1/1089)
8 + 1/5254 + 1/1526 + 1/1526 + 1 /
5123) -0.334 x (667/10898 + 44
4/5254 + 466/1526 + 436/4530 +
189/1526 + 421/5123) / (1/108
98 + 1/5254 + 1/1526 + 1/1526 + 1
/5123)=92.7.

Therefore, the normalization constants are (κ, λ) = (0.334, 92.7), (0.99) in order from the first candidate.
4, -221), (0.614, 46.3), (0.6
78, 31.9), and so on.

The duration length normalizing means 103 normalizes (corrects) the duration length of each recognition result candidate based on the mathematical expression 2. The corrected duration time is described in the corrected recognition result candidate in FIG.

For example, the duration time of the word "Nisen" of the first-ranked candidate (second-ranked candidate before correction) of the recognition result candidate (after correction) is 0.334 × 667 + 92.7 = 315.

The recognition result candidate correction means 104 corrects the score L _k of the k-th candidate using the duration length normalized by the duration length normalization means 103 based on the following equation (Equation 5). .

[0061]

[Equation 5]

The corrected score value is written in the corrected recognition result candidate in FIG. However, here, the calculation is performed with α = 1.0.

For example, the score value of the second candidate (first candidate before modification) of the recognition result candidates (after modification) is 86.0−.
^{(315-442) 2/10898 + (240-29
⁴⁾ 2/5254 + ^(248-160) 2/1526 +
^(238-285) 2/4530 + (155-160)
^2/1526 + ⁽²³³ + 351) 2/5123} = 8
It is 6.0-10.3 = 75.3. The recognition result candidate correction means 104 further re-sorts the recognition result candidates based on the corrected score value and re-ranks them. as a result,
The candidates who were in 1st and 2nd place before the correction had a large drop in score due to the correction (mainly due to the unnaturally long duration of the word "no"), and some candidates were in 4th place before the correction. The score goes down (mainly the word "Issen")
Due to the unnaturally small duration of the), the candidate (correct answer), which was in 3rd place before the correction, went up to 1st place.

Here, an example of the procedure for calculating the duration statistics (average and variance) prepared in advance, which is illustrated in FIG. 3, will be described. As learning samples, voice data and correct word strings corresponding to them are prepared. For this voice data, the same acoustic model as that used by the recognition result candidate generation means 102 is used to obtain the optimum alignment of words (start time and end time of each word) by the Viterbi search.

As a result, the durations of various words can be known, so that the averages and variances of the durations can be calculated by aggregating each word. The larger the learning sample, the better. However, it is desirable that one type of word appears several times or more.

Incidentally, if the normalization constant calculation means 105
In order to reduce the cost of the calculation of the normalization constant in, the difference in the variance of each word is ignored, and in many cases, the calculation amount of Equations 3 and 4 is reduced without significantly reducing the effectiveness. Is possible.
The calculation formula is the following formula (Formula 6 and Formula 7).

[0067]

[Equation 6]

[0068]

[Equation 7]

In the above, the procedure of normalizing the duration time, that is, converting the duration time so as to be closer to its average has been described, but the reverse procedure is also possible. In other words, even in the procedure of converting the average of the durations so as to be closer to the actually obtained duration, the same processing can be effectively realized. In that case, the duration length normalizing means 1
The duration length is not normalized in 03. Instead, the recognition result candidate correction means 104 corrects the likelihood using the normalized duration length statistic. The method is represented by the following equation (Equation 8).

[0070]

[Equation 8]

Further, the calculation of the normalization constant used in the above equation is given by the following equations (Equations 9 and 10).

[0072]

[Equation 9]

[0073]

[Equation 10]

Next, a second embodiment of the present invention will be described with reference to the drawings.

In the second embodiment of the present invention, the first procedure (corresponding to the voice signal input means 101) for receiving a voice signal input from the outside and the voice signal from the first procedure are received and prepared in advance. Based on the set of standard patterns and the grammar, a plurality of word strings having a high score that match the voice signal are obtained in order from the top, and a plurality of word strings,
A second procedure (corresponding to recognition result candidate generating means 102) for outputting a word duration sequence corresponding to each of the plurality of word strings, and a score corresponding to each of the plurality of word strings, and the second procedure. A third procedure (corresponding to the duration length normalizing means 103) for receiving a word string, a duration length sequence and a score from, and normalizing the duration length from the word string and duration length using a normalization constant; A fourth procedure for correcting the score output by the second procedure based on the average and variance of the durations and outputting it (recognition result candidate correction means 10
4)) and a normalization constant used by the third procedure for normalization of the duration length so that the squared error between the duration length sequence and the average of the duration lengths is minimized. A fifth procedure (corresponding to the normalization constant calculation means 105) that calculates based on the average and variance of lengths, when the fourth procedure calculates a score, and when the fifth procedure calculates a normalization constant. A sixth procedure (corresponding to the duration statistical storage means 106) for storing the average and variance of durations used when performing
It is a method including and.

Next, a third embodiment of the present invention will be described with reference to the drawings.

In the third embodiment of the present invention, a first procedure (corresponding to the voice signal input means 101) for receiving a voice signal input from the outside and a voice signal from the first procedure are received and prepared in advance. Based on the set of standard patterns and the grammar, a plurality of word strings having a high score that match the voice signal are obtained in order from the top, and a plurality of word strings,
A second procedure (corresponding to recognition result candidate generating means 102) for outputting a word duration sequence corresponding to each of the plurality of word strings, and a score corresponding to each of the plurality of word strings, and the second procedure. A third procedure (corresponding to the duration length normalizing means 103) for receiving a word string, a duration length sequence and a score from, and normalizing the duration length from the word string and duration length using a normalization constant; A fourth procedure for correcting the score output by the second procedure based on the average and variance of the durations and outputting it (recognition result candidate correction means 10
4)) and a normalization constant used by the third procedure for normalization of the duration length so that the squared error between the duration length sequence and the average of the duration lengths is minimized. A fifth procedure (corresponding to the normalization constant calculation means 105) that calculates based on the average and variance of lengths, when the fourth procedure calculates a score, and when the fifth procedure calculates a normalization constant. A sixth procedure (corresponding to the duration statistical storage means 106) for storing the average and variance of durations used when performing
Is a program that causes a computer to execute.

Next, a fourth embodiment of the present invention will be described with reference to the drawings.

FIG. 4 is a block diagram showing the configuration of the fourth embodiment of the present invention.

Referring to FIG. 4, the fourth embodiment of the present invention is a recording medium 301 in which a voice recognition program is recorded.
Equipped with. The recording medium 301 may be a CD-ROM, a magnetic disk, a semiconductor memory or the like, including a case of being distributed via a network.

The voice recognition program is read from the recording medium 301 to the data processing device 302, and the data processing device 3
02 operation is controlled.

As an example of the fourth embodiment of the present invention, the data processing device 302 controls the voice recognition program to control the voice signal input means 1 of the first embodiment.
01, the recognition result candidate generation means 102, the duration length normalization means 103, the recognition result candidate correction means 104, and the normalization constant calculation means 105 are executed, and the duration time in the first embodiment is executed. Long statistic storage means 10
Duration duration statistic recording medium 30 having information equivalent to 6
By referring to 3, the speech recognition result is output.

[0083]

The effect of the present invention is that the information on the duration can be used more effectively, and as a result, the accuracy of voice recognition can be improved.

The reason is that, in the technique of improving the accuracy of speech recognition by using the information on the duration time, even if the duration time is disturbed by noise or the like, the fluctuation of the speech speed can be estimated with higher accuracy. This is because the fluctuation factors can be absorbed in more detail.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is an explanatory diagram showing an example of correction of recognition result candidates by score correction based on word duration length.

FIG. 3 is an explanatory diagram showing an example of word duration statistics.

FIG. 4 is a block diagram showing a configuration of a fourth exemplary embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a conventional technique.

[Explanation of symbols]

101 audio signal input means 102 recognition result candidate generating means 103 Duration normalization means 104 recognition result candidate correction means 105 Normalization constant calculation means 106 duration duration statistic storage means 201 Audio signal input means 202 recognition result candidate generating means 203 Recognition result candidate correction means 204 Duration duration statistics storage means 301 recording medium 302 Data processing device 303 Duration Length Statistics Recording Medium

Claims (9)

[Claims] 1. A voice signal input means for receiving a voice signal input from the outside, a voice signal from the voice signal input means, and a voice signal based on a set of standard patterns prepared in advance and a grammar. Obtain multiple word strings with high matching scores in order from the top, and output multiple word strings, word duration series corresponding to each word string, and scores corresponding to each word string. A recognition result candidate generating unit that receives the word string, the duration sequence and the score from the recognition result candidate generating unit, and the duration time for normalizing the duration length from the word string and the duration length using a normalization constant. Long normalization means,
The recognition result candidate correction means for correcting and outputting the score output by the recognition result candidate generation means based on the average and variance of the durations, and the duration normalization means for normalizing the duration A normalization constant calculation means for calculating the normalization constant to be used based on the average of the durations, the variance, when the recognition result candidate correction means calculates a score, and the normalization constant calculation means calculates the normalization constant. A voice recognition device comprising: a duration statistics storage means for storing an average and a variance of durations used for calculation. 2. A voice signal input means for receiving a voice signal input from the outside, a voice signal from the voice signal input means, and a voice signal based on a set of standard patterns prepared in advance and a grammar. Obtain multiple word strings with high matching scores in order from the top, and output multiple word strings, word duration series corresponding to each word string, and scores corresponding to each word string. A recognition result candidate generating unit that receives the word string, the duration sequence and the score from the recognition result candidate generating unit, and the duration time for normalizing the duration length from the word string and the duration length using a normalization constant. Long normalization means,
Recognition result candidate correction means for correcting and outputting the score output by the recognition result candidate generation means based on the average and variance of the duration time, and a first type normalization constant for multiplying the duration time by a constant, and , A second-type normalization constant for performing constant addition of durations is calculated based on the average and variance of durations as a normalization constant used by the duration normalizing means for normalization of durations. The normalization constant calculating means and the recognition result candidate correcting means store the average and variance of the durations used when calculating the score and when the normalization constant calculating means calculates the normalization constant. A speech recognition apparatus comprising: a duration statistics storage unit. 3. A voice signal input means for receiving a voice signal input from the outside, a voice signal from the voice signal input means, and a voice signal based on a set of standard patterns prepared in advance and a grammar. Obtain multiple word strings with high matching scores in order from the top, and output multiple word strings, word duration series corresponding to each word string, and scores corresponding to each word string. A recognition result candidate generating unit that receives the word string, the duration sequence and the score from the recognition result candidate generating unit, and the duration time for normalizing the duration length from the word string and the duration length using a normalization constant. Long normalization means,
The recognition result candidate correction means for correcting and outputting the score output by the recognition result candidate generation means based on the average and variance of the durations, and the duration normalization means for normalizing the duration Normalization constant calculating means for calculating the normalization constant to be used based on the average of the durations and the variance so that the squared error between the duration sequence and the average of the durations is minimized, and the recognition result And a duration length statistic storage means for storing an average and a variance of the duration length used when the candidate correction means calculates the score and when the normalization constant calculation means calculates the normalization constant. A voice recognition device characterized by the above. 4. A first procedure for receiving a voice signal input from the outside, and a voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a fifth step in which the third procedure calculates a normalization constant used for normalization of the duration based on the average and variance of the duration, and the fourth procedure calculates a score. When, and the fifth Mean procedure duration length used in calculating the normalization constant, speech recognition method, which comprises the steps of the sixth, the storing the dispersion. 5. A first procedure for receiving a voice signal input from the outside, and a voice signal for receiving the voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a normalization constant of the first kind for multiplying the duration length by a constant, and a second kind of normalization constant for adding a constant of the duration time, in order to normalize the duration time by the third procedure. Continue as a normalization constant used for A fifth step of calculating based on the mean and variance of interval lengths, a duration used when the fourth step calculates a score, and when the fifth step calculates a normalization constant. And a sixth step of storing the average and variance of the speech recognition method. 6. A first procedure for receiving a voice signal input from the outside, and a voice signal for receiving the voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a normalization constant used in the third procedure for normalizing the duration length so that the squared error between the duration sequence and the average duration is minimized so as to minimize the mean squared duration. , Calculated based on variance That a fifth procedure, when the fourth procedure for calculating the scores, and the average of the duration of the fifth step is used in calculating a normalization constant,
A sixth step of storing the variance; and a voice recognition method. 7. A first procedure for receiving a voice signal input from the outside, and a voice signal for receiving the voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a fifth step in which the third procedure calculates a normalization constant used for normalization of the duration based on the average and variance of the duration, and the fourth procedure calculates a score. When, and the fifth Mean procedure duration length used in calculating the normalization constant, speech recognition program, characterized in that to execute a sixth step of storing the dispersion, to the computer. 8. A first procedure for receiving a voice signal input from the outside, and a voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a normalization constant of the first kind for multiplying the duration length by a constant, and a second kind of normalization constant for adding a constant of the duration time, in order to normalize the duration time by the third procedure. Continue as a normalization constant used for A fifth step of calculating based on the mean and variance of interval lengths, a duration used when the fourth step calculates a score, and when the fifth step calculates a normalization constant. And a sixth procedure for storing the average and variance of the voice recognition program. 9. A first procedure for receiving a voice signal input from the outside, and a voice signal from the first procedure,
Based on a set of standard patterns and grammar prepared in advance, a plurality of word strings with high scores that match the audio signal are obtained in order from the top, and a plurality of word strings and words corresponding to each of the plurality of word strings are obtained. A second procedure for outputting a duration sequence and a score corresponding to each of a plurality of word strings, and a word sequence, a duration sequence and a score are received from the second procedure, and the word sequence and the duration length are received. From the third procedure for normalizing the duration using the normalization constant, and the fourth procedure for correcting and outputting the score output by the second procedure based on the average and variance of the duration. And a normalization constant used in the third procedure for normalizing the duration length so that the squared error between the duration sequence and the average duration is minimized so as to minimize the mean squared duration. , Calculated based on variance That a fifth procedure, when the fourth procedure for calculating the scores, and the average of the duration of the fifth step is used in calculating a normalization constant,
A sixth step of storing the variance, and a voice recognition program causing a computer to execute the sixth step.