JP2014077865A - Speech recognition device, error correction model learning method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To optimize an error correction model using a piece of information picked up from latest utterance content of a speech as an object of speech recognition.SOLUTION: A speech language resource storage 21 stores a result of speech recognition of a voice data of a speech made by a speech recognition section 11 and speech correct word strings maintaining the order of the speech. A model learning section 13 learns in a statistical manner a tendency of recognition failure of a word based on the words included in the speech recognition result, linguistic characteristics according to the order of the speech obtained from the words included in a previous speech close to the speech recognition result in time scale, and linguistic characteristics in the identical speech obtained from the speech recognition result to generate an error correction model for correcting the tendency of the learnt recognition failure. A speech recognition section 14 recognizes the speech included in the voice data and corrects incorrectness in a selection of speech recognition result using the generated error correction model.

Description

  The present invention relates to a speech recognition device, an error correction model learning method, and a program.

  For error correction in speech recognition, statistical error correction models obtained as a result of learning by statistically learning the tendency of speech recognition errors using linguistic features from speech and transcriptions (correct sentences) There is a technology for improving the performance of speech recognition by using (see, for example, Non-Patent Document 1).

Kobayashi et al., “News speech recognition by discriminative scoring based on word error minimization”, IEICE Journal, vol.J93-D no.5, 2010, p. 598-609

In speech recognition for broadcast programs and the like, a plurality of continuous utterances are successively recognized, but the content of the utterance being processed by the speech recognition is often related to the utterance content immediately before speech recognition is finished. For example, in a cooking program, if there is an utterance about the introduction of ingredients, it is expected that the subsequent utterance about the cooking method will continue. That is, there is a high possibility that words related to food ingredients and words related to the cooking method co-occur in adjacent utterances. For example, if the utterance of “pick a pork fin” is followed by the utterance of “next salt and pepper”, there is a relationship between “pig fin” and “salt pepper”, and these are co-occurring It will be easy to do.
However, in conventional model parameter learning of an error correction model, speech data, a speech recognition result thereof, and a correct word string are used, but the speech data speech order is not considered during learning. As described above, in the conventional error correction model, information such as word co-occurrence between utterances related to the utterance order is not taken into consideration, so that it is not an optimal model for correctly predicting the utterance contents.

  The present invention has been made in consideration of such circumstances, and a speech recognition device that optimizes an error correction model using information extracted from past utterance contents from utterances targeted for speech recognition, An error correction model learning method and program are provided.

[1] According to one aspect of the present invention, a speech language resource storage unit that stores a speech recognition result obtained by speech recognition of speech speech data and the correct word string of the speech while maintaining the speech order. The linguistic features according to the order of utterances obtained from the words included in the speech recognition result obtained from the speech data and the words included in the correct word string of utterances in the past than the speech of the speech data And a model learning unit that statistically learns the tendency of recognition errors of words based on the above and generates an error correction model for correcting the learned tendency of recognition errors. is there.
According to this invention, the speech recognition apparatus recognizes speech data, and includes words included in the obtained speech recognition result and words included in correct word strings of utterances utterances earlier than the speech recognition result. Extract linguistic features according to the order of utterances. As the correct word string of the past utterance, for example, the correct word string of the latest past utterance that is temporally adjacent to the speech recognition result is used. The speech recognition apparatus statistically learns the tendency of recognition errors of words based on the extracted linguistic features, and generates an error correction model for correcting the learned tendency of recognition errors.
Thus, an error correction model suitable for correctly predicting the utterance content can be generated using the information extracted from the utterance content before the utterance subject to speech recognition.

[2] One aspect of the present invention is the speech recognition device described above, in which the model learning unit is linguistic in accordance with linguistic features in the same utterance obtained from the speech recognition result and the order of the utterances. A tendency of statistically learning a recognition error tendency of a word based on the characteristic and generating an error correction model for correcting the learned tendency of the recognition error.
According to this invention, the speech recognition apparatus extracts linguistic features corresponding to the order of utterances from the speech recognition result and the correct word string and extracts linguistic features in the same utterance from the speech recognition result. The speech recognition apparatus statistically learns the tendency of recognition errors of words based on these extracted linguistic features, and generates an error correction model for correcting the tendency of learned recognition errors.
As a result, the speech recognition device uses the linguistic features in the same utterance in addition to the information extracted from the utterance contents past the utterance that is the target of speech recognition, and corrects the recognition error with high accuracy. An error correction model can be generated.

[3] One aspect of the present invention is the speech recognition device described above, in which the model learning unit includes a co-occurrence relationship between a plurality of consecutive words in the same utterance obtained from the speech recognition result, One or more of word co-occurrence relationships, word syntactic information, or word semantic information, and a word included in the speech recognition result and a word included in the correct word string of the past utterance. It is characterized in that the error tendency of words is statistically learned based on the origin relation.
According to the present invention, the speech recognition apparatus provides the word co-occurrence relationship and syntactic and semantic information in the same utterance obtained from the speech recognition result, the word included in the speech recognition result, and the correct word of the past utterance. A word error tendency is statistically learned based on the word co-occurrence relationship obtained from the sequence, and an error correction model for correcting the learned recognition error tendency is generated.
Thereby, the speech recognition apparatus can generate an error correction model that corrects a recognition error with high accuracy.

[4] One aspect of the present invention is the speech recognition device described above, wherein the error correction model is a calculation formula for correcting a speech recognition score using a feature function based on the linguistic feature and its weight. The model learning unit is calculated by an evaluation function determined using the feature function value obtained from the speech recognition result and the correct word string and a recognition error of a word included in the speech recognition result. The weight is statistically calculated based on an evaluation value, and the error correction model is generated using the calculated weight.
According to this invention, the speech recognition apparatus uses the feature function value obtained from the speech recognition result, the weight used by the error correction model defined by the feature function representing the linguistic feature and its weight, and the speech recognition result. An error correction model is generated by determining that the evaluation value calculated by the evaluation function determined using the recognition error obtained by comparing the correct word string with the correct word string is the evaluation value indicating the least recognition error. To do.
Thereby, the speech recognition apparatus can learn the recognition error tendency efficiently and generate an error correction model.

[5] One aspect of the present invention is the speech recognition device described above, which recognizes input speech data and uses the error correction model generated by the model learning unit to input the input speech. It further comprises a voice recognition unit for correcting an error in selecting a voice recognition result obtained from data.
According to this invention, the speech recognition apparatus selects a speech recognition result using an error correction model from among correct answer candidates obtained by speech recognition of speech data.
Thereby, the speech recognition apparatus can obtain a speech recognition result with a good recognition rate.

[6] According to one aspect of the present invention, a speech recognition result obtained by speech recognition of speech speech data and the correct word sequence of the speech are stored in the speech language resource storage unit while maintaining the speech order. In the order of utterances obtained from the spoken language resource storage process, the words included in the speech recognition result obtained from the speech data, and the words included in the correct word sequence of utterances in the past than the speech of the speech data A model learning process for statistically learning a tendency of recognition error of a word based on a corresponding linguistic feature and generating an error correction model for correcting the learned tendency of the recognition error. This is an error correction model learning method.

[7] According to one aspect of the present invention, a speech language resource that stores a speech recognition result obtained by speech recognition of speech speech data and a correct word string of the speech while maintaining a speech sequence in a computer. A language according to the order of utterances obtained from the storage means and the words included in the speech recognition result obtained from the speech data and the words included in the correct word sequence of utterances earlier than the speech of the speech data Model learning means for statistically learning the tendency of recognition errors of words based on typical characteristics and generating an error correction model for correcting the learned tendency of recognition errors It is a program to make it.

  According to the present invention, it is possible to optimize an error correction model using information extracted from past utterance contents rather than utterances that are targets of speech recognition.

It is a figure which shows the learning method of the error correction model in the speech recognition apparatus by one Embodiment of this invention. It is a figure which shows the example of the feature function according to the order of the story by the embodiment. It is a functional block diagram which shows the structure of the speech recognition apparatus by the embodiment. It is a figure which shows the whole processing flow of the speech recognition apparatus by the embodiment. It is a figure which shows the model learning process flow of the speech recognition apparatus by the embodiment. It is a figure which shows the learning method of the error correction model by the conventional method.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[1. Overview of this embodiment]
An error correction model that reflects the error tendency of speech recognition has already been devised, but this error correction model uses information based on the relationship between adjacent utterance contents for continuously uttered utterances. It is not a thing. Consecutive utterances often include words related to the word used in the previous utterance. Therefore, if such a word connection between adjacent utterances is used in an error correction model, an improvement in speech recognition is expected.

  Therefore, the speech recognition apparatus according to the present embodiment learns an error correction model in which speech recognition performance is adapted to the utterance content using linguistic features included in the latest utterance content, and applies it to speech recognition. As described above, the performance of speech recognition is improved by the error correction model optimized by the latest utterance content.

[2. Error correction model learning algorithm]
Subsequently, an error correction model learning algorithm applied to the speech recognition apparatus according to the embodiment of the present invention will be described.
As described above, in order to solve the conventional problem, the speech recognition apparatus according to the present embodiment introduces an utterance order relationship into speech data used for learning, and incorporates a relationship between adjacent utterances into an error correction model. . The difference between this embodiment and the conventional method is how to handle data when learning an error correction model.

  FIG. 6 is a diagram illustrating an error correction model learning method according to a conventional method. As shown in the figure, in the conventional method, the learning data composed of a plurality of utterances does not preserve the order relation, and the error tendency of words has been learned using data collectively.

  FIG. 1 is a diagram illustrating an error correction model learning method according to the present embodiment. As shown in the figure, in this embodiment, in consideration of the order relation of each utterance in the learning data, the relation between the temporally adjacent utterances is extracted as a linguistic feature, and the error correction model is learned. Use. As a result, an error correction model reflecting the relationship between adjacent utterances can be obtained, so that speech recognition performance can be improved as compared with the conventional method.

[2.1 Error correction model of conventional method]
According to Bayes' theorem, when speech input x is given, a word string w ^ (“^” represents “hat”) that is likely to be associated with speech input x is expressed by the following equation (1). It can ask for.

The voice input x and the word string w correspond to, for example, an utterance unit, and P (w | x) is a posterior probability that a word string (sentence hypothesis) w is obtained when the voice input x occurs.
P (x | w) is a likelihood indicating acoustic likelihood for the word string w, and the score (acoustic score) is a hidden Markov model (HMM) and a Gaussian mixture distribution (Gaussian Mixture). It is calculated based on a statistical acoustic model (hereinafter referred to as “acoustic model”) typified by Model, GMM). In other words, when an acoustic feature amount is given, a score representing the likelihood of each of a plurality of correct candidate words is an acoustic score.

  On the other hand, P (w) is a linguistic generation probability for the word string w, and the score (language score) is described as a statistical language model (hereinafter, “language model”) such as a word n-gram model. ). In other words, when a word string before or after a speech recognition target word, or both word strings before and after the given word string, a score representing the likelihood of each of a plurality of correct answer word strings is a language score. The word n-gram model gives the occurrence probability of the next word from the history of the word (N-1) based on the statistics of N word chains (N is 1, 2, or 3, for example). It is a model.

  In the following description, HMM-GMM is used for the statistical acoustic model, and n-gram is used for the statistical language model.

  When P (x | w) P (w) in Equation (1) is maximum, the logarithm is also maximum. Therefore, in speech recognition, the evaluation function q (w | x) of the word string w, which is a sentence hypothesis (correct answer candidate) when the speech input x is generated, based on the Bayesian theorem of the above equation (1) is as follows. This is determined as shown in equation (2). Note that κ is a weight of the language score P (w) with respect to the acoustic score P (x | w).

  Then, as shown in the following equation (3), the word having the maximum result of the evaluation function q (w | x) indicated by the equation (2) is selected from the set of word strings w of correct answer candidates for the voice input x. The column。 is selected as the speech recognition result for speech input x.

  In the error correction model in the conventional method, equation (1) is changed to the following equation (4).

ExpΣ _i λ _i f _i (w) in equation (4) is a score reflecting the error tendency of the word string w, and acts as a penalty or reward for the word string w. Further, f _i (w) (i = 1,...) Is an i-th feature function, and λ _i is a weight (feature weight) of the feature function f _i (w). The feature function is defined as a function that becomes the number if a linguistic rule is established in a given word string (here, word string w), and is 0 if not established. These rules are, for example, linguistic features such as consecutive words in the same utterance, co-occurrence relationship of two or more words that are not consecutive, syntactic information or semantic information of words. Examples of rules specific feature function f _i in the conventional method, and the like below.

  For example, there are the following (1) and (2) as feature functions based on the co-occurrence relationship of words.

(1) When the word string w includes a continuous word binary set (u, v), a function that returns the number (2) When the word string w includes a non-continuous word binary set (u, v) , A function that returns the number

  For example, the following (3) and (4) are feature functions based on syntax information obtained by replacing each word constituting the word string w with a part-of-speech category (syntax information) such as a noun or a verb. Note that c (•) is a function that maps words to parts of speech.

(3) A function that returns the number of part-of-speech binaries (c (u), c (v)) that are consecutive in the word string w (4) A part-of-speech binary pair that is not consecutive in the word string w u), c (v)), a function that returns the number if it is included

  Alternatively, for example, the following (5) and (6) are feature functions based on semantic information obtained by replacing each word constituting the word string w with a category (semantic category) representing semantic information. is there. The semantic category can be obtained by using a thesaurus stored in a database provided outside or inside the speech recognition apparatus of the present embodiment. Note that s (•) is a function that maps words to semantic categories.

(5) A function that returns the number of semantic category binary groups (s (u), s (v)) that are consecutive in the word string w (6) A semantic category binary group that is not consecutive in the word string w ( a function that returns the number of s (u), s (v))

As described above, an error tendency of speech recognition is expressed as a penalty for a linguistic feature by a feature function and its weight, and is estimated based on an evaluation function that minimizes a word error in learning data. In other words, the conventional learning of error tendency is to obtain the weight λ _i of Equation (4) using the speech recognition result of speech data and the correct word string as learning data.

[2.2 Learning algorithm of error correction model applied to this embodiment]
Now, with respect to the word string w, if the word string obtained from the latest input speech is u, the conditional probability P (w | x,) of the word string w when the speech input x and the word string u are given. u) is expressed by the following equation (5).

  However, the derivation of Equation (5) uses the Bayes' theorem and the fact that the word string u and the speech input x are independent. The word string u is a word string having an arbitrary length, and may be a word string obtained by connecting a plurality of utterance contents.

  However, when an utterance word string u adjacent to the voice input x is given, the most likely word string w ^ for the input is expressed by the following expression (6), and the expression (1) is changed. To do.

  Here, the conditional probability P (w | u) of the word string w when the word string u is given by the latest input speech is assumed as shown in Expression (7).

Note that g _j (w, u) (j = 1,...) Is a feature function representing a linguistic feature with respect to the word string w and the word string u, and φ _j is a weight corresponding to g _j (Feature weight). Examples of the feature function g _j of linguistic features corresponding to the order of such utterances are as follows. Here, v and z are words.

(Example) Function that returns the number of words v included in the word sequence w of the utterance of interest when the word z is included in the word sequence u of the preceding utterance

FIG. 2 is a diagram illustrating an example of the feature function g _j . In the figure, the preceding utterance word string u is a correct word string (or likely recognition result), and the current utterance word string w of interest is the correct candidate word string w ₁ , w ₂ , w _3. It is a set of. In the drawing, the word sequence u of the preceding utterance includes the word z, and the correct candidate word sequence w ₁ includes one word v. In this case, g _j (w ₁ , u) = 1. On the other hand, since the correct candidate word string w3 includes two words v, g _j (w ₃ , u) = 2.

  From Expression (5) and Expression (7), the following Expression (8) is obtained.

  Considering the feature function of the conventional discriminative language model, Expression (8) becomes the following Expression (9).

Assuming that the likelihood of the acoustic model is a logarithmic acoustic score by HMM is h ₀ (x, w), and a logarithmic language score by the n-garam language model is h ₁ (w), the equation (9) is expressed by the following equation (10): Can be rewritten as

Here, κ is a weighting factor for the language score. Z (Λ, Φ) is a normalization constant for satisfying the probability condition, and is represented by the following expression (11). The word string w ′ in Equation (11) is a plurality of speech recognition results obtained by speech recognition from the speech input x. The model parameter Λ is (λ ₁ , λ ₂ ,...), And the model parameter Φ is (φ ₁ , φ ₂ ,...).

  The learning of the error correction model by the speech recognition apparatus of the present embodiment is to estimate the model parameters Λ and Φ used for the error correction model shown in Expression (10) from the learning data.

  Here, when learning data composed of M utterances is given, an objective function L (Λ, Φ) for model parameter estimation is expressed by the following equation (12).

P (w _{m, n} | x _m , w _m−1 ^ref ) in the equation (12) is calculated as in the following equation (13).

m indicates the order of utterances, N _m is the total number of sentence hypotheses w _{m, 1} , w _{m, 2} ,... generated by speech recognition on the learning data of the m-th utterance, sentence hypothesis w _{m, n} ( n is an integer of 1 or more) is a word string of the nth correct answer candidate of learning data of the mth utterance. w _m ^ref is a correct word string of learning data of the m-th utterance, and R (·, ·) is a function that returns an edit distance between two word strings. The edit distance between two word strings can be efficiently obtained by dynamic programming. Since the edit distance is equivalent to the number of error words in the speech recognition result for the correct word string (operation of substitution, insertion, omission error), the objective function L (Λ, Φ) in Expression (12) is the expected word. It represents the number of errors. If the model parameter Λ and the model parameter Φ are estimated so as to minimize the objective function L (Λ, Φ), an error correction model that minimizes the number of expected word errors can be obtained. Improvement can be expected. This is because if the model parameters Λ and Φ are estimated so as to minimize the objective function L (Λ, Φ), the recognition error expected for the correct candidate word string is minimized, and the unknown input is different from the learning data. This is because also in speech recognition for speech, recognition errors are similarly minimized by the model parameters Λ and Φ. That is, the objective function of Expression (12) is used as an evaluation function for calculating an evaluation value as to whether or not the model parameters Λ and Φ are appropriate because the recognition error expected for the correct candidate word string is minimized.

  In order to estimate the parameters, the gradients ΔΛ and ΔΦ with respect to the model parameters Λ and Φ of the objective function are obtained from the following expressions (14) and (15).

The gradient ΔΛ is (∂L (Λ, Φ) / ∂λ ₁ , ∂L (Λ, Φ) / ∂λ ₂ , ∂L (Λ, Φ) / ∂λ ₃ ,...), And the gradient ΔΦ is (∂L (Λ, Φ) / ∂φ ₁ , ∂L (Λ, Φ) / ∂φ ₂ , ∂L (Λ, Φ) / ∂φ ₃ ,...).

If the model parameters Λ ^t and Φ ^t are learned by iterative updating, and the model parameters Λ ^t−1 and Φ ^t−1 are obtained after the ^t− 1th ^iteration , the following equation (16), Expression (17) is a parameter update expression.

Here, η _Λ and η _Φ are the coefficients of the gradient ΔΛ and the gradient ΔΦ obtained by the equations (14) and (15), respectively.

[3. Configuration of voice recognition device]
FIG. 3 is a functional block diagram showing the configuration of the speech recognition apparatus 1 according to an embodiment of the present invention, and only functional blocks related to the invention are extracted and shown.
The speech recognition apparatus 1 is realized by a computer device, and as shown in the figure, a speech recognition unit 11, a feature amount extraction unit 12, a model learning unit 13, a speech recognition unit 14, a speech language resource storage unit 21, and an acoustic model storage A unit 22, a language model storage unit 23, and an error correction model storage unit 24.

  The spoken language resource storage unit 21 stores learning data. The acoustic model storage unit 22 stores an acoustic model. The language model storage unit 23 stores a language model. The error correction model storage unit 24 stores an error correction model.

  The voice recognition unit 11 recognizes voice data to generate learning data. The voice data indicates a feature amount obtained by performing short-time spectrum analysis on the voice waveform of the utterance. In the present embodiment, broadcast audio / caption data D1 is used as audio data. The speech recognition unit 11 associates speech data of speech, speech recognition result data D2 obtained by speech recognition of speech data, and correct word sequence data D3 indicating a correct word sequence of the content of speech to learn data. Is written in the spoken language resource storage unit 21. At this time, the speech recognizing unit 11 also holds the order of utterances when performing speech recognition in the spoken language resource storage unit 21.

The feature quantity extraction unit 12 obtains linguistic features in the same utterance and linguistic features according to the utterance order from the speech recognition result data D2 and correct word string data D3 of the learning data arranged in the utterance order. Extract. The feature quantity extraction unit 12 outputs feature function data D4 indicating feature functions f _i and g _j using the obtained linguistic features as rules.

  The model learning unit 13 receives the feature function data D4 output from the feature quantity extraction unit 12 and the learning data stored in the spoken language resource storage unit 21, and inputs model parameters Λ and Φ of the error correction model by statistical means. learn. The model learning unit 13 writes the error correction model using the learned model parameters Λ and Φ into the error correction model storage unit 24.

  The speech recognition unit 14 refers to the acoustic model stored in the acoustic model storage unit 22 and the language model stored in the language model storage unit 23, and stores the error correction model stored in the error correction model storage unit 24. Is used to perform speech recognition of the input speech data D5, and speech recognition result data D6 is output.

[4. Processing procedure of voice recognition device]
FIG. 4 is a diagram showing an overall processing flow of the speech recognition apparatus 1 according to the present embodiment. Hereinafter, processing of each step shown in FIG.

[4.1 Step S1]
In this embodiment, in order to generate an error correction model, a speech recognition result of an utterance and a correct word string that is a transcription of the utterance content are required as learning data. Therefore, the voice recognition unit 11 collects the broadcast voice / caption data D1 as a set of voice data and correct word string data, and recognizes the voice data included in the broadcast voice / subtitle data D1. The voice recognition unit 11 sets the caption data acquired from the broadcast voice / subtitle data D1 or text data obtained by manually correcting the result of voice recognition as correct word string data D3. The speech recognition unit 11 stores learning data in which speech data of each utterance, speech recognition result data D2 indicating a speech recognition result, and correct word string data D3 are associated with each other in the speech language resource storage unit 21. At this time, the voice recognition unit 11 stores and stores the order of utterances when voice recognition is performed. The speech recognition result data D2 of the m-th (m = 1, 2,...) learning data includes sentence hypotheses w _{m, n} (n = 1) that are correct candidates obtained by speech recognition of the m-th speech data. , 2,..., And the correct word string data D3 of the m-th learning data includes the correct word string w _m ^ref of the m-th speech data.

[4.2 Step S2]
The model learning unit 13 extracts a feature function based on linguistic features used for error tendency learning from the learning data stored in the spoken language resource storage unit 21.

First, the model learning unit 13 obtains continuous words, non-consecutive two or more words, syntactic information of words or semantics from each of the speech recognition result data D2 and the correct word string data D3 included in the learning data. All feature functions based on linguistic features in the same utterance such as information are extracted. Further, the model learning unit 13 indicates the correct candidate sentence hypothesis w _{m, n} (corresponding to the word strings w ₁ , w ₂ ,... Shown in FIG. 2) indicated by the speech recognition result data D2 and the correct word string data D3. With reference to the correct word string w _m−1 ^ref preceding the utterance (corresponding to the preceding word string u shown in FIG. 2), all feature functions based on the linguistic features corresponding to the utterance order are extracted. The correct word string w _m−1 ^ref is a correct word string of a past utterance that is temporally adjacent to the utterance of the sentence hypothesis w _{m, n} . The model learning unit 13 counts the frequency at which these extracted feature functions appear. The model learning unit 13 performs error tendency learning on a feature function based on a linguistic feature in the same utterance whose counted appearance frequency is equal to or higher than a predetermined threshold, and a feature function based on a linguistic feature corresponding to the order of utterances. The feature function f _{i to be} used and the feature function g _j are determined. The model learning unit 13 outputs the feature function data D4 in which the determined feature functions f _i and g _j are set to the model learning unit 13.

[4.3 Step S3]
Subsequently, the model learning unit 13 learns model parameters of the error correction model.
FIG. 5 is a diagram illustrating a processing flow of the error correction model update process executed by the model learning unit 13 in step S3.

(Step S31: Model parameter initialization process)
The model learning unit 13 sets appropriate initial values for the model parameters Λ and Φ. In this embodiment, the initial value is Λ = Φ = 0.

(Step S32: Edit distance calculation process)
In order to calculate the objective function of equation (12), it is first necessary to calculate the edit distance between the speech recognition result and the corresponding correct word string. Therefore, the model learning unit 13 reads the learning data stored in the spoken language resource storage unit 21 as learning data _, and the correct word shown by the sentence hypothesis w _{m, n} indicated by the speech recognition result data D2 and the correct word string data D3. An edit distance R (w _m ^ref , w _{m, n} ) is calculated from the column w _m ^ref . Note that these edit distances are treated as constants when learning the error correction model.

(Step S33: Objective function calculation process)
The model learning unit 13 calculates the value of the objective function L (Λ, Φ) according to the equation (12) using the edit distance R (w _m ^ref , w _{m, n} ) obtained in step S32. Therefore, the model learning unit 13 uses the conditional probability P (w _{m, n} | x _m , w _m−1 ^ref ) in Equation (12) as the acoustic model score h ₀ (x _m | w _{m, n} ), The language model score h ₁ (w _{m, n} ) and the current model parameters Λ and Φ are calculated using the equation (13). The model learning unit 13 stores the acoustic model score h ₀ (x _m | w _{m, n} ) of each sentence hypothesis w _{m, n in} the acoustic model storage unit 22 and the m-th learning data. Using the voice data. The model learning unit 13 acquires the language model score h ₁ (w _{m, n} ) of the sentence hypothesis w _{m, n} using the language model stored in the language model storage unit 23.

(Step S34: gradient calculation process)
The model learning unit 13 obtains the gradients ΔΛ and ΔΦ related to the model parameters Λ and Φ in Expression (12) using Expression (14) and Expression (15) using the current model parameters Λ and Φ. The model learning unit 13 sets the edit distance R (w _m ^ref , w _{m, n} ) and the conditional probability P (w _{m, n} | x _m , w _m−1 ^ref ) in the equations (14) and (15). The value when the objective function L (Λ, Φ) is calculated in step S33 is used. Further, the model learning unit 13 uses the sentence hypothesis w _{m, n} and the correct word string indicated by the speech recognition result data D2 as the value of the feature function g _j (w _{m, n} , w _m−1 ^ref ) in the equation (14). Obtained from the correct word string w _m-1 ^ref indicated by the data D3. The model learning unit 13 acquires the value of the feature function f _i (w _{m, n} ) in Expression (15) from the sentence hypothesis w _{m, n} indicated by the speech recognition result data D2. The model learning unit 13 obtains a feature function f _i and a feature function g _j from the feature function data D4.

The model learning unit 13 updates the model parameters Λ and Φ by the equations (16) and (17) using the obtained gradients ΔΛ and ΔΦ. Note that predetermined values are used as the coefficients η _Λ and η _Φ in the equations (16) and (17).

(Step S35: End determination process)
The model learning unit 13 compares the value of the objective function obtained by the gradient calculation process in step S34 with the value of the objective function before update, and if the change in value is equal to or greater than a predetermined value, the model learning unit 13 performs the process from step S33. Repeatedly, if it is smaller than the predetermined value, it is regarded that the update has converged, the update of the model parameters Λ and Φ is aborted, and the process of step S36 is executed.

(Step S36: Error correction model output process)
The model learning unit 13 converts the error correction model using the model parameters Λ = (λ ₁ , λ ₂ ,...) And Φ = (φ ₁ , φ ₂ ,. 24.

[4.4 Step S4]
When the input speech data D5 is input as speech recognition target speech data, the speech recognition unit 14 stores the error correction model stored in the error correction model storage unit 24 and the acoustic stored in the acoustic model storage unit 22. Using the model and the language model stored in the language model storage unit 23, a word string of correct answer candidates of the input speech data D5 is obtained, and their scores are calculated. At the time of learning, the utterance sequence preceding the utterance currently being processed is a correct word sequence, but since the correct word sequence is not obtained at the time of speech recognition, the speech recognition unit 14 utters before the utterance currently being processed. Is used as a correct word string. The voice recognition unit 14 outputs the voice recognition result data D6 in which the correct candidate word string having the best score is set as the correct word string in real time. By using this error correction model, the speech recognition unit 14 corrects an error in the selection of the speech recognition result obtained from the input speech data D5.

[5. effect]
According to the present embodiment, the speech recognition apparatus 1 can configure an error correction model that reflects the content of the immediately preceding utterance, and recognition errors are reduced compared to conventional speech recognition.

[6. Others]
The voice recognition device 1 described above has a computer system inside. The operation process of the speech recognition apparatus 1 is stored in a computer-readable recording medium in the form of a program, and the above processing is performed by the computer system reading and executing this program. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

DESCRIPTION OF SYMBOLS 1 Speech recognition apparatus 11 Speech recognition part 12 Feature-value extraction part 13 Model learning part 14 Speech recognition part 21 Spoken language resource storage part 22 Acoustic model storage part 23 Language model storage part 24 Error correction model storage part

Claims (7)

A spoken language resource storage unit that stores a speech recognition result obtained by speech recognition of speech data of speech and a correct word string of the speech while maintaining the order of speech;
A linguistic feature according to the order of utterances obtained from words included in the speech recognition result obtained from the speech data and words included in the correct word string of utterances earlier than the speech of the speech data. A model learning unit that statistically learns the tendency of recognition errors of words based on the error and generates an error correction model for correcting the learned tendency of recognition errors;
A speech recognition apparatus comprising: The model learning unit statistically learns the tendency of recognition errors of words based on linguistic features in the same utterance obtained from the speech recognition result and linguistic features according to the order of the utterances, Generating an error correction model for correcting the learned tendency of recognition errors;
The speech recognition apparatus according to claim 1. The model learning unit includes a co-occurrence relationship between a plurality of consecutive words in the same utterance obtained from the speech recognition result, a co-occurrence relationship between a plurality of non-consecutive words, syntactic information of words, or a word semantic Statistically learning the error tendency of the word based on one or more of the information and the co-occurrence relationship of the word included in the speech recognition result and the word included in the correct word sequence of the past utterance,
The speech recognition apparatus according to claim 2. The error correction model is a calculation formula that corrects a speech recognition score using a feature function based on the linguistic feature and its weight,
The model learning unit calculates an evaluation value calculated by an evaluation function determined using the feature function value obtained from the speech recognition result and the correct word string and a word recognition error included in the speech recognition result. Statistically calculating the weight based on, and generating the error correction model using the calculated weight,
The speech recognition apparatus according to any one of claims 1 to 3, wherein A speech recognition unit that recognizes input speech data and corrects an error in selection of a speech recognition result obtained from the input speech data using the error correction model generated by the model learning unit; In addition,
The voice recognition device according to claim 1, wherein the voice recognition device is a voice recognition device. A spoken language resource storage step of storing a speech recognition result obtained by speech recognition of speech data of speech and a correct word string of the speech in a speech language resource storage unit while maintaining the order of speech;
A linguistic feature according to the order of utterances obtained from words included in the speech recognition result obtained from the speech data and words included in the correct word string of utterances earlier than the speech of the speech data. A model learning process for statistically learning the tendency of recognition errors of words based on the generated error correction model for correcting the learned tendency of recognition errors;
An error correction model learning method characterized by comprising: Computer
A spoken language resource storage means for storing a speech recognition result obtained by speech recognition of speech data of speech and a correct word sequence of the speech while maintaining the order of speech;
A linguistic feature according to the order of utterances obtained from words included in the speech recognition result obtained from the speech data and words included in the correct word string of utterances earlier than the speech of the speech data. Model learning means for statistically learning a tendency of recognition error of a word based on the error and generating an error correction model for correcting the learned tendency of the recognition error;
A program for causing a voice recognition apparatus to function.

Images