JP4362054B2 - Speech recognition apparatus and speech recognition program - Google Patents

Description

  The present invention relates to speech recognition technology, and more particularly to a speech recognition apparatus and speech recognition program for recognizing input speech in a noisy environment.

  Conventionally, as a technique for performing speech recognition, an acoustic likelihood indicating similarity is obtained from a feature amount (feature vector) obtained by analyzing an acoustic signal of an input speech and an acoustic model obtained by modeling a feature amount of a phoneme in speech. The speech recognition is performed by searching for a word as a recognition candidate from the acoustic likelihood and a language model obtained by modeling the appearance frequency and connection probability of the word (for example, Non-Patent Document 1). ).

  Here, a conventional speech recognition apparatus will be described with reference to FIG. FIG. 4 is a block diagram showing a configuration of a conventional speech recognition apparatus. As shown in FIG. 4, the conventional speech recognition device 20 is based on an acoustic model (stored in the acoustic model storage unit 21) and a language model (stored in the language model storage unit 22). The content is determined by a search algorithm, and the recognition result is output.

  In the speech recognition apparatus 20, first, the acoustic analysis unit 23 extracts a plurality of feature quantities (for example, cepstrum) for each phoneme from the input speech (input speech). Then, based on the plurality of feature quantities (feature vectors) extracted by the acoustic analysis unit 23 by the acoustic likelihood calculation unit 24 and the acoustic model stored in the acoustic model storage unit 21, The acoustic likelihood indicating the similarity with the acoustic model is calculated.

Then, the speech recognition device 20 uses the search means 25 based on the acoustic likelihood calculated by the acoustic likelihood calculation means 24 and the connection probability of the word indicated by the language model stored in the language model storage means 22. Then, the candidate of the connected word is searched, and the search result (search candidate) is notified to the acoustic likelihood calculating means 24. Then, the speech recognition apparatus 20 can calculate the connection probability of the word string when the input speech is a word string by sequentially operating the acoustic likelihood calculating unit 24 and the search unit 25. Then, the search means 25 outputs the word string having the maximum connection probability as the recognition result of the input voice.
Thus, conventionally, a method of performing speech recognition based on an acoustic model and a language model learned from learning data in advance is common.
Imai, Kobayashi, Onoe, Ando, "Speech recognition system for automatic subtitle conversion of news program", IPSJ Spoken Language Processing Technical Report, 23-11, pp. 59-64, Oct. , 1998

However, the speech recognition method described above recognizes the noise as a phoneme when noise is superimposed on the input speech and the noise is similar to the phoneme modeled in the acoustic model, There is a problem that an incorrect voice recognition result is output.
This problem is caused by performing a search similar to that of the speech portion in the noise portion that has not been learned by the acoustic model, although the reliability of the acoustic likelihood is reduced. This problem is also caused by the fact that the acoustic model does not sufficiently learn the features of the input speech on which noise is superimposed.

  To solve this problem, it is conceivable to create an acoustic model from learning data in which noise is superimposed on the speech (noise-superimposed speech). However, since the learning data to be learned by the acoustic model is limited, the acoustic model Is difficult to learn in response to a wide variety of noise superimposed speech.

  The present invention has been made in view of the above problems, and even when noise is superimposed on the input speech, the speech can be obtained without using an acoustic model corresponding to a variety of noise superimposed speech. An object of the present invention is to provide a speech recognition device and a speech recognition program that can reduce recognition errors in recognition.

The present invention has been created to achieve the above object, and first, the speech recognition apparatus according to claim 1 includes an acoustic model and a language model, and a second acoustic model obtained by modeling the acoustic model. A speech recognition device for recognizing an input speech using a noise model obtained by modeling noise, an acoustic analysis means, a reliability calculation means, an acoustic likelihood calculation means, an acoustic likelihood correction means, An output series search means and a recognition result output means are provided.

  According to such a configuration, the speech recognition apparatus analyzes the acoustic signal of the input speech (input speech) by the acoustic analysis means by spectrum analysis, linear prediction analysis, cepstrum analysis, and the like, and extracts speech feature values. This feature quantity need not be a single feature quantity, but can be a feature vector having a plurality of feature quantities, so that the features of the input speech can be appropriately expressed.

In addition, the speech recognition apparatus calculates the reliability indicating the degree to which the feature amount (feature vector) extracted by the acoustic analysis unit is the feature amount of speech by the reliability calculation unit. In other words, based on a noise model obtained by modeling noise data in advance and a second acoustic model obtained by modeling an acoustic model, an input speech having the feature amount is determined according to a ratio of probability density function values in each model. The degree (reliability) of whether or not it is speech is calculated. Here, the noise model is obtained by learning noise data assumed in advance and modeling the data using, for example, a mixed normal distribution model. The acoustic model is obtained by modeling a feature amount for each phoneme learned in advance from a large amount of speech data using a hidden Markov model. The second acoustic model is generated by modeling the acoustic model with a mixed normalization distribution model, for example. As described above, by using the second acoustic model in which the data amount is reduced, it is possible to suppress the calculation amount when calculating the reliability.

  Then, the speech recognition device calculates an acoustic likelihood indicating the similarity between the input speech and the acoustic model based on the feature amount and the acoustic model by the acoustic likelihood calculating means. That is, in this acoustic likelihood calculating means, the acoustic likelihood indicating whether or not the input speech is speech is calculated only by the acoustic model. However, noise may be superimposed on the input speech, and there is no guarantee that the acoustic likelihood is accurate. Therefore, the speech recognition apparatus corrects the acoustic likelihood calculated by the acoustic likelihood calculating means by the acoustic likelihood correcting means with the reliability calculated by the reliability calculating means, so that the acoustic likelihood considering noise ( Corrected acoustic likelihood) is calculated.

Further, the speech recognition apparatus searches for an output sequence (word, morpheme, phoneme, etc.) having a high connection probability based on the corrected acoustic likelihood and the language model by the output sequence search means. Note that the output sequence candidates (search candidates) searched here are appropriately notified to the acoustic likelihood calculating means so that the acoustic likelihood calculating means can input the sound of the input speech based on the acoustic model of the search candidate. Calculate the likelihood.
Then, the speech recognition apparatus identifies the output sequence having the maximum connection probability among the plurality of output sequences searched by the output sequence search unit by the recognition result output unit as an output sequence in which the input speech is recognized. And output as a speech recognition result.

According to a second aspect of the present invention, there is provided a speech recognition program that uses an acoustic model and a language model, a second acoustic model obtained by modeling the acoustic model, and a noise model obtained by modeling noise data, to input speech. In order to recognize, the computer functions as an acoustic analysis unit, a reliability calculation unit, an acoustic likelihood calculation unit, an acoustic likelihood correction unit, an output sequence search unit, and a recognition result output unit.

According to this configuration, the speech recognition program analyzes the acoustic signal of the input speech (input speech) by the acoustic analysis unit, and extracts the feature amount of the speech.
In addition, the speech recognition program calculates the reliability indicating the degree to which the input speech is speech based on the feature amount extracted by the acoustic analysis unit, the noise model, and the second acoustic model by the reliability calculation unit. .

Then, the speech recognition program calculates an acoustic likelihood indicating the similarity between the input speech and the acoustic model based on the feature amount and the acoustic model by the acoustic likelihood calculating unit, and the acoustic likelihood correcting unit calculates the acoustic likelihood. By correcting the acoustic likelihood calculated by the likelihood calculating means with the reliability calculated by the reliability calculating means, an acoustic likelihood considering noise (corrected acoustic likelihood) is calculated.
Further, the speech recognition program searches for an output sequence (word, morpheme, phoneme, etc.) having a high connection probability based on the corrected acoustic likelihood and the language model by the output sequence search means, and by the recognition result output means, Among the output sequences searched for by the output sequence search means, the output sequence having the maximum connection probability is identified as the output sequence that has recognized the input speech, and is output as a speech recognition result.

According to the first or second aspect of the present invention, which output series the input speech corresponds to, based on a noise model obtained by modeling noise data, even for speech with superimposed noise. The acoustic likelihood shown can be corrected. As a result, when the reliability of acoustic likelihood decreases, it is possible to perform recognition by weighting linguistic constraints such as vocabulary, grammar, and meaning based on the language model. Errors can be reduced.
In addition, it is not necessary to model the acoustic model in correspondence with a wide variety of noise-superimposed speech, and only the noise model needs to be constructed, so that the model can be easily constructed.

According to the invention described in claim 1 or claim 2 , the reliability is calculated using the second acoustic model (voice model) in which the data amount of the acoustic model used for calculating the acoustic likelihood is reduced. Therefore, it is possible to reduce the amount of calculation for the reliability calculation.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Configuration of voice recognition device]
First, the configuration of a speech recognition apparatus according to a reference example of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus.

The noise model storage means 11 is a general storage medium such as a hard disk that stores the noise model 11a. The noise model 11a stored here is obtained by modeling noise data based on a database (not shown) in which data of expected noise is accumulated. As this noise model 11a, for example, a mixed normal distribution model can be used. Incidentally, the noise model 11a may be used single noise model lambda ^N, a plurality of models (cluster model) for each noise type _{^{λ N = {λ 1 N,}} λ 2 N, ..., λ M N } (M is the number of clusters in the noise model).
Here, the noise means a sound other than the voice to be recognized, for example, a sound of an airplane or a siren, a busy sound, a sound of turning a manuscript in a news program, or the like.

The acoustic model storage unit 12 is a storage medium such as a general hard disk that stores the acoustic model 12a. The acoustic model 12a stored here is obtained by modeling a feature amount for each phoneme learned in advance from a large amount of speech data using a “hidden Markov model”. Similarly to the noise model 11a, this acoustic model 12a may use a single acoustic model λ ^S , or a plurality of models (cluster models) λ ^S = {λ for each acoustic type (for example, for each person). ₁ ^S , λ ₂ ^S ,..., Λ _K ^S } (K is the number of clusters in the acoustic model) may be used. Further, for example, when the acoustic model 12a is modeled by using a triphone in which front and rear phonemes are paired with respect to one phoneme as a recognition unit, there are as many cluster models as the number of triphones.

The language model storage means 13 is a general storage medium such as a hard disk that stores the language model 13a. The language model 13a stored here models the appearance frequency, connection probability, and the like of an output sequence (words, morphemes, phonemes, etc.) learned from a large amount of speech data. For example, a general “N-gram language model” or the like can be used.
Here, the noise model storage unit 11, the acoustic model storage unit 12, and the language model storage unit 13 are configured as separate storage units, but each model may be stored in the same storage unit. It is. Each storage unit may be connected via a network.

The acoustic analysis unit 14 analyzes a voice (input voice) input from the outside and extracts a feature amount of the voice. The extracted feature amount of the voice is output to the reliability calculation unit 15 and the acoustic likelihood calculation unit 16 in time series as a feature vector.
The acoustic analysis means 14 extracts a framed waveform by applying a window function (such as a Hamming window) to the voice waveform of the input voice, and performs various frequency analysis on the waveform. Extract the amount. For example, a cepstrum coefficient or the like, which is a value obtained by inverse Fourier transform of the logarithm of the power spectrum of a framed waveform, is used as the feature amount. In addition to the cepstrum coefficient, a general speech feature quantity such as a mel frequency cepstrum coefficient (MFCC), LPC (Linear Predictive Cordind) coefficient, logarithmic power, or the like can be used as the feature quantity.

The reliability calculation means 15 is based on the noise model 11a previously learned and stored in the noise model storage means 11 and the acoustic model 12a previously learned and stored in the acoustic model storage means 12. To calculate a reliability P (S | x _t ) indicating the degree to which the feature vector x _{t at} time t output in time series is speech (not noise). The reliability calculated here is output to the acoustic likelihood correcting means 17.

Hereinafter, a method for calculating the reliability P (S | x _t ) will be specifically described.
First, _let the noise model 11a be λ ^N = {λ ₁ ^N , λ ₂ ^N ,..., Λ _M ^N } (M is the number of clusters of the noise model), and the feature of the noise model 11a in the i-th cluster model λ _i ^N A degree (likelihood) that the vector x _t is a feature amount of noise is defined as a conditional probability P (x _t | λ _i ^N ) (1 ≦ i ≦ M). Then, the degree (noise model likelihood) P (x _t | λ ^N ) that the feature vector x _t is a feature amount of noise is expressed as the likelihood sum in each cluster model λ _i ^N as shown in the equation (1). Can be calculated.

Similarly, the acoustic model 12a is λ ^S = {λ ₁ ^S , λ ₂ ^S ,..., Λ _K ^S } (K is the number of clusters of the acoustic model), and the i-th cluster model λ _i ^{S of the} acoustic model 12a. The degree (likelihood) that the feature vector x _t is the feature amount of speech is defined as a conditional probability P (x _t | λ _i ^S ) (1 ≦ i ≦ K). Then, the degree (acoustic model likelihood) P (x _t | λ ^S ) that the feature vector x _t is the feature amount of speech is expressed as the likelihood sum in each cluster model λ _i ^S as shown in the equation (2). Can be calculated.

As described above, the degree P (x _t | λ ^N ) that the feature vector x _t is a noise feature amount and the degree P (x _t | Based on (λ ^S ), the reliability P (S | x _t ) indicating the degree to which the feature vector x _t is speech (not noise) can be calculated by the equation (3).

In addition, when the reliability of sufficient accuracy cannot be obtained due to insufficient accuracy of the noise model 11a, a lower limit value may be provided for the reliability value P (S | _xt ). For example, the lower limit value as the alpha, (4) the formula, P may calculate the | | (x _t S) as a confidence value P'corrected for (S x _t). As a result, when the accuracy of the noise model is not sufficient, it is possible to avoid the adverse effect of losing information on the acoustic likelihood itself.

Here, as shown in the equations (1) and (2), the degree to which the feature vector x _t is the feature amount of noise or speech is calculated as the likelihood sum in each cluster model. Alternatively, the maximum degree (likelihood) in each of the cluster models λ _i ^N and λ _i ^S may be the degree that the feature vector x _t is a noise or speech feature amount. That is, according to the following formulas (5) and (6), the degree of noise feature (noise model likelihood) P (x _t | λ ^N ), the degree of voice feature (acoustic model likelihood) P (x _t | λ ^S ) may be calculated.

Further, the degree P (x _t | λ ^N ) that the feature vector x _t is a noise feature amount and the degree P (x _t | λ ^S ) that is a speech feature amount are _{expressed by} Equations (7) and (8). As shown in the _figure , a specific time width (moving average window width w) is set for the feature vector xt, and the degree (likelihood) that is the feature amount of noise or speech for each time width is added. May be. Equations (7) and (8) are moving average values obtained by adding an average value for a certain period for each of a plurality of frames in each frame obtained by framing input speech by a window function such as a Hamming window. Thereby, the accuracy of reliability can be improved. Here, the reason why the averaging by the division is not performed is to calculate the reliability as the ratio by the equation (3).

Returning to FIG. 1, the description of the configuration of the speech recognition apparatus 10 will be continued.
The acoustic likelihood calculating means 16 is a similarity between the feature vector extracted by the acoustic analyzing means 14 and inputted in time series and the phoneme modeled by the acoustic model 12 a stored in the acoustic model storage means 12. Is calculated. The acoustic likelihood calculating unit 16 calculates an acoustic likelihood for each output sequence search candidate that is sequentially output from the searching unit 18. The acoustic likelihood calculated here is output to the acoustic likelihood correcting means 17.
For example, if the search candidate acoustic model sequentially output from the search means 18 is λ ^AM = {λ ₁ ^AM , λ ₂ ^AM ,..., Λ _J ^AM } (J is the total number of search candidate models) The likelihood calculating means 16 calculates the degree (acoustic likelihood) that the feature vector x _t is speech with a conditional probability P (x _t | λ _j ^AM ) (1 ≦ j ≦ J).

The acoustic likelihood correcting unit 17 corrects the acoustic likelihood for each search candidate calculated by the acoustic likelihood calculating unit 16 based on the reliability calculated by the reliability calculating unit 15 to obtain a corrected acoustic likelihood. Is to be generated. The corrected acoustic likelihood calculated here is output to the search means 18.
For example, when the acoustic likelihood P (x _t | λ _j ^AM ) (1 ≦ j ≦ J) of the search candidate at time t is output from the acoustic likelihood calculating unit 16, the acoustic likelihood correcting unit 17 9) As shown in the equation, the acoustic likelihood P (x _t | λ _j ^AM ) is raised to the power by using the reliability P (S | x _t ) calculated by the reliability calculation means 15 as a power, and the sound A corrected acoustic likelihood P ′ (x _t x _t λ _j ^AM x _t ) obtained by correcting the likelihood is generated.

In Equation (9), since the reliability is normalized by 0 ≦ P (S | x _t ) ≦ 1, the corrected acoustic likelihood P ′ (x _t | λ _j ^AM ) is the acoustic likelihood P ( x _t | λ _j ^AM ). However, in the comparison of the speech likelihood (corrected acoustic likelihood) between the contending candidates when searching for a search candidate by the search means 18 described later, if the difference in log likelihood obtained by taking the logarithm of the acoustic likelihood is taken, The dynamic range of the log-likelihood difference is only P (S | x _t ) times, and there is no effect even if the acoustic likelihood increases.

  Based on the corrected acoustic likelihood generated by the acoustic likelihood correcting unit 17, the searching unit 18 searches the language model 13 a for a candidate of an output series to be connected, and sets the search candidate that is the search result as the acoustic likelihood. While outputting to the calculation means 16, the output series with the largest connection probability is output to the outside as a recognition result for the input speech. Here, the search means 18 is comprised by the output series search part 18a and the recognition result output part 18b.

  The output sequence search unit (output sequence search means) 18 a has a connection probability based on the corrected acoustic likelihood generated by the acoustic likelihood correction means 17 and the language model 13 a stored in the language model storage means 13. This searches for an output sequence that increases. Note that since the corrected acoustic likelihood is low for speech with noise superimposed thereon, the output sequence search unit 18a searches for an output sequence based on linguistic restrictions (for example, vocabulary, grammar, meaning, etc.) of the language model 13a. Increase the degree. In other words, for speech with superimposed noise (noise superimposed speech), the recognition accuracy of speech distorted by noise can be improved by weakening acoustic constraints and relatively increasing the weight of linguistic constraints. .

  The recognition result output unit (recognition result output unit) 18b outputs an output sequence having the maximum connection probability among the plurality of output sequences searched by the output sequence search unit 18a. The output sequence search unit 18a searches for a plurality of output sequences, but the recognition result output unit 18b uses the plurality of output sequences as a voice (output sequence) having the maximum connection path probability. The recognition result.

Having described the structure of the speech recognition apparatus 10 according to a reference example of the present invention, the present invention provides reliability calculation means 15, using the speech models modeled by reducing the amount of data of the acoustic model 12a, The reliability was calculated .
That is, according to the present invention, as shown in the block diagram of the speech recognition device 10B shown in FIG. 2, the speech model storage means 19 is added to the speech recognition device 10 (see FIG. 1) . The voice recognition device 10B is configured by a reliability calculation unit 15B in which a voice model storage unit 19 storing a voice model 19a is added to the voice recognition device 10 and the function of the reliability calculation unit 15 is changed. . Other configurations are the same as those of the speech recognition apparatus 10 shown in FIG.

The voice model (second acoustic model) 19a stored in the voice model storage unit 19 is a model in which the data amount is reduced from that of the acoustic model 12a. The speech model 19a is generated by modeling the acoustic model 12a with a mixed normalization distribution model, for example. The speech model 19a may be modeled by learning from a large amount of speech data, but since the purpose is to calculate the reliability by the reliability calculation means 15B, the acoustic model 12a is generated. It is desirable to learn from the speech data used when doing this.
As the acoustic model 12a, the speech model 19a may use a single speech model λ ^S , or a plurality of models λ ^S = {λ ₁ ^S , λ ₂ ^S ,. _K ^S } (K is the number of clusters in the speech model) may be used.

The reliability calculation means 15B is output in time series from the acoustic analysis means 14 based on the noise model 11a stored in the noise model storage means 11 and the speech model 19a stored in the speech model storage means 19. The reliability P (S | x _t ) indicating the degree to which the feature vector x _{t at the} time t is a speech is calculated in this manner by using the speech model 19a with the data amount of the acoustic model 12a reduced. The amount of calculation for recognition can be reduced.

The speech recognition apparatuses 10 and 10B described above can be realized by causing a general computer to execute a program and operating an arithmetic device or a storage device in the computer. This program (voice recognition program) can be distributed via a communication line, or can be distributed by writing on a recording medium such as a CD-ROM.

[Operation of voice recognition device]
Next, the operation of the speech recognition apparatus 10 will be described with reference to FIG. FIG. 3 is a flowchart showing the operation of the speech recognition apparatus 10.

(Acoustic analysis step)
Speech recognition device 10, by the acoustic analysis means 14 analyzes the voice (input voice) input, extracts a feature quantity of the speech in time series as a feature vector x _t (Step S1).

(Reliability calculation step)
Then, the speech recognition apparatus 10 calculates the reliability indicating the degree to which the input speech indicated by the feature vector x _t is speech by the reliability calculation means 15. That is, the reliability calculation means 15 refers to the noise model 11a as shown in the equation (1), and the likelihood sum P (x _t | λ ^N ) indicating the degree to which the feature vector x _t is noise. Is calculated (step S2). Then, the reliability calculation means 15 refers to the acoustic model 12a as shown in the equation (2), and the likelihood sum P (x _t | λ ^S ) indicating the degree to which the feature vector x _t is speech. Is calculated (step S3). The reliability calculation means 15 then adds the noise likelihood sum P (x _t | λ ^N ) calculated in step S2 and the speech likelihood sum P (x _t | λ ^S ) calculated in step S3. Based on the equation (3), the reliability P (S | x _t ) indicating the degree to which the feature vector x _t is speech is calculated (step S4).

(Acoustic likelihood calculation step)
In addition, the speech recognition apparatus 10 outputs the output sequence of the search unit 18 based on the feature vector x _t that is the feature amount of the speech extracted by the acoustic analysis unit 14 by the acoustic likelihood calculation unit 16 and the acoustic model 12a. The acoustic likelihood is calculated for each output sequence search candidate searched by the search unit 18a (step S5). That is, when the search candidate acoustic model is λ ^AM = {λ ₁ ^AM , λ ₂ ^AM ,..., Λ _J ^AM } (J is the total number of search candidate models), the degree to which the feature vector x _t is speech (sound Likelihood) is calculated as a conditional probability P (x _t | λ _j ^AM ) (1 ≦ j ≦ J).

(Acoustic likelihood correction step)
Then, the speech recognition apparatus 10 uses the acoustic likelihood correcting unit 17 to calculate the acoustic likelihood P (x _t | λ _{j of} the search candidate calculated by the acoustic likelihood calculating unit 16 as shown in the equation (9). ^AM ) (1 ≦ j ≦ J) is corrected based on the reliability P (S | x _t ) calculated by the reliability calculation means 15, and the corrected acoustic likelihood P ′ (x _t | λ _j ^AM ) is corrected. Generate (step S6).

(Search step)
Then, the speech recognition apparatus 10 applies the corrected acoustic likelihood P ′ (x _t | λ _j ^AM ) generated by the acoustic likelihood correcting unit 17 and the language model 13 a by the output sequence searching unit 18 a of the searching unit 18. Based on this, an output sequence with a high connection probability is searched, and search candidates that are search results are output to the acoustic likelihood calculating means 16 (step S7). This search result is used when the acoustic likelihood calculating means 16 calculates the acoustic likelihood in step S5.

The speech recognition device 10, recognition by the results output section 18b, among the plurality of output sequences that have been searched in step S7, and outputs an output series connection probability is maximized as the speech recognition result (step S8). In this operation, the operation after step S1 is repeatedly executed while the input voice is input.

  Here, in step S4, the reliability is calculated using the degree that the feature vector is the feature quantity of noise and the degree that the feature vector is the feature quantity of speech. However, the above formulas (5) and (6) The reliability may be calculated using the maximum likelihood in the cluster model shown or the moving average values shown in the equations (7) and (8).

Through the above operation, the speech recognition apparatus 10 performs a search based on the linguistic restriction of the language model 13a by reducing the reliability P (S | _xt ) when noise is superimposed on the input speech. The degree can be increased. As a result, even if a mismatch occurs between the acoustic model 12a and the input speech due to noise, recognition errors can be reduced.

It is the block diagram which showed the structure of the speech recognition apparatus which concerns on the reference example of this invention. It is a block diagram showing a configuration of a speech recognition apparatus according to the present invention. It is a flowchart which shows operation | movement of the speech recognition apparatus which concerns on the reference example of this invention. It is the block diagram which showed the structure of the conventional speech recognition apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10, 10B Speech recognition apparatus 11 Noise model storage means 11a Noise model 12 Acoustic model storage means 12a Acoustic model 13 Language model storage means 13a Language model 14 Acoustic analysis means 15 Reliability calculation means 16 Acoustic likelihood calculation means 17 Acoustic likelihood correction Means 18 Search means 18a Output series search section (output series search means)
18b Recognition result output unit (recognition result output means)
19 Voice model storage means 19a Voice model (second acoustic model)

Claims (2)

A speech recognition device that recognizes an input speech using an acoustic model and a language model, a second acoustic model obtained by modeling the acoustic model, and a noise model obtained by modeling noise data,
An acoustic analysis means for analyzing an acoustic signal of the input speech and extracting a feature quantity of the input speech;
A reliability calculation means for calculating a reliability indicating the degree that the input voice is a voice based on the feature amount extracted by the acoustic analysis means, and the noise model and the second acoustic model;
Acoustic likelihood calculating means for calculating an acoustic likelihood indicating a similarity between the input speech and the acoustic model based on the feature amount and the acoustic model;
Acoustic likelihood correction means for correcting the acoustic likelihood calculated by the acoustic likelihood calculation means with the reliability to generate a corrected acoustic likelihood;
Based on the corrected acoustic likelihood generated by the acoustic likelihood correcting means and the language model, an output series searching means for searching for an output series candidate constituting the input speech;
A recognition result output means for outputting, as a speech recognition result of the input speech, an output sequence having a maximum connection probability among a plurality of output sequences searched by the output sequence search means;
A speech recognition apparatus comprising: A computer for recognizing input speech using an acoustic model and a language model, a second acoustic model that models the acoustic model, and a noise model that models noise data,
An acoustic analysis means for analyzing an acoustic signal of the input voice and extracting a feature quantity of the input voice;
A reliability calculation means for calculating a reliability indicating the degree that the input voice is a voice based on the feature amount extracted by the acoustic analysis means, and the noise model and the second acoustic model;
An acoustic likelihood calculating means for calculating an acoustic likelihood indicating the similarity between the input speech and the acoustic model based on the feature quantity and the acoustic model;
An acoustic likelihood correcting means for generating a corrected acoustic likelihood by correcting the acoustic likelihood calculated by the acoustic likelihood calculating means by the reliability;
Output sequence search means for searching for output sequence candidates constituting the input speech based on the corrected acoustic likelihood generated by the acoustic likelihood correction means and the language model;
A recognition result output means for outputting, as a speech recognition result of the input speech, an output sequence having a maximum connection probability among the plurality of output sequences searched by the output sequence search means;
A voice recognition program characterized by functioning as

Images