JP6128287B1 - Speech recognition apparatus and speech recognition method - Google Patents

Abstract

The speech recognition apparatus according to the present invention corrects the difference between the average vector of the acoustic model and the average vector of the input speech to the feature vector of the input speech, and the input speech is the first utterance. It collates with the acoustic model which adjusted the dispersion value according to whether it is. Therefore, even if there is a difference in the frequency characteristics of the microphone between the learning of the acoustic model and the recognition of the input speech, and there is noise in the surroundings, the accuracy of speech recognition is not delayed for the first utterance without a delay in the recognition end time. Can ensure

Description

  The present invention relates to a speech recognition apparatus and speech recognition method for recognizing input speech.

  There are various types of microphones used for speech recognition, and frequency characteristics are not unified. For example, the difference in frequency characteristics of microphones varies greatly depending on the type of smartphone or mobile phone. If the microphone frequency used for recording the learning data of the acoustic model, which is a standard pattern of sound used for speech recognition, and the frequency characteristics of the microphone used for speech recognition are different, the speech recognition performance generally decreases. In recent speech recognition, a method of performing pattern matching based on a statistical method using a frequency pattern of an input speech as a feature vector is mainly used. This method uses pattern matching between an acoustic model that models the statistical features of this feature vector using frequency vector feature vectors of speech data spoken by a large number of speakers in advance and the feature vector of the input speech. Realize voice recognition. For this reason, it is possible to improve the recognition performance by learning an acoustic model by using various microphones having different frequency characteristics as speech and recording speech of many speakers.

  However, there was a limit to the improvement effect when the frequency characteristics of the microphone at the time of learning and recognition were greatly different. On the other hand, it is known that the performance of speech recognition is improved by a method called CMN (Cepstral mean normalization). The CMN is realized by obtaining an average vector of speech feature vectors during learning and recognition of the acoustic model, and subtracting the average vector from each feature vector during learning and recognition. Since the average vector reflects the frequency characteristic of the microphone, the difference in the frequency characteristic of the microphone can be almost absorbed by subtracting the average vector.

In Patent Document 1, when an HMM (Hidden Markov Model) is used as an acoustic model, a method of performing CMN by approximately obtaining an average vector from parameters of an HMM obtained after learning rather than performing CMN during learning. Disclosure. Disclosed is a technique for obtaining an acoustic model that is robust to both multiplicative distortion such as the difference in frequency characteristics of microphones and additive distortion such as ambient noise by combining this method with acoustic model noise adaptation. . In Patent Document 1, as a method for calculating an average vector of input speech, there is a method of calculating an average vector from one entire utterance for each utterance of input speech or calculating an average vector from feature vectors up to the previous utterance at the time of speech recognition. It is shown.

JP 2006-349723 A

However, since the method of Patent Document 1 cannot calculate the average vector of the entire utterance unless one utterance is completed, the recognition process can be performed only after one utterance is completed, resulting in a slow response speed of recognition. is there.
The present invention aims to solve the above problems. In other words, the object is to ensure the accuracy of speech recognition without delay of the recognition end time for the first utterance even if there is a difference in the frequency characteristics of the microphone or there is noise in the surroundings.

The speech recognition apparatus according to the present invention includes an analysis unit that analyzes input speech and outputs a first feature vector, an acoustic model that models a second feature vector of speech data uttered by a plurality of speakers , wherein an average vector of the average vector or the first feature vector of the input speech acoustic models, the first feature is subtracted from the vector to output the first feature vector after the correction the correction means, said An adjustment unit that adjusts a variance value of the parameter of the acoustic model according to whether or not the speech is the first utterance, and the acoustic model that has adjusted the variance value is compared with the corrected first feature vector. And collating means for outputting the speech recognition result.

The speech recognition apparatus of the present invention subtracts the average vector of the acoustic model or the feature vector of the input speech from the feature vector of the input speech to obtain a corrected feature vector. Then, the corrected feature vector is collated with an acoustic model whose variance value is adjusted according to whether or not the input speech is the first utterance, and a recognition result is output. Therefore, even if there is a difference in the frequency characteristics of the microphone between the learning of the acoustic model and the recognition of the input speech, and there is noise in the surroundings, the accuracy of speech recognition is not delayed for the first utterance without a delay in the recognition end time. Can be secured.

It is a block diagram of the speech recognition apparatus 1 in Embodiment 1 of this invention. It is a hardware block diagram of the speech recognition apparatus 1 in Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the correction | amendment means 4 in Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the correction | amendment means 4 in Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the adjustment means 6 in Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the adjustment means 6 in Embodiment 2 of this invention.

  Embodiments of a speech recognition apparatus and a speech recognition method according to the present invention will be described below in detail with reference to the drawings. Note that the present invention is not limited to the embodiments.

Embodiment 1 FIG.
FIG. 1 is a configuration diagram of a speech recognition apparatus 1 according to Embodiment 1 of the present invention.
In FIG. 1, the speech recognition apparatus 1 outputs a feature vector (first feature vector) obtained by analyzing the input speech (input speech) 2 by the acoustic analysis unit 3 to the correction unit 4. Corrects the feature vector depending on whether or not the input speech 2 is the first utterance. The correction means 4 acquires the initial vector 5 if the input speech 2 is the first utterance, corrects the feature vector with the initial vector 5, and temporarily stores the feature vector if it is the second utterance or later. The feature vector is corrected with an average vector of feature vectors up to the previous utterance. Then, the variance value of the parameter of the acoustic model 7 is adjusted by the adjusting means 6 depending on whether or not the input speech 2 is the first utterance, and the acoustic model 7 having the variance value adjusted by the matching means 8 is corrected. The recognition result 9 of the input speech 2 is output by collating the vector.
The acoustic analysis unit 3 corresponds to an analysis unit.

FIG. 2 is a hardware configuration diagram of the speech recognition apparatus 1 according to Embodiment 1 of the present invention.
The speech recognition apparatus 1 includes a processor 10 and a memory 11. Specifically, the acoustic analysis unit 3, the correction unit 4, the adjustment unit 6, and the collation unit 8 are realized by the processor 10 executing a program stored in the memory 11. The initial vector 5 and the acoustic model 7 are stored in the memory 11.

  The acoustic analysis unit 3 analyzes the input speech 2 and outputs a feature vector obtained by the analysis to the correction unit 4. This feature vector is a vector representing a spectral feature that is a frequency pattern of speech, and is, for example, data of 1 to 12 dimensions of MFCC (Mel Frequency Cepstral Coefficient). For example, the feature vector is divided into sections of every 10 milliseconds called input frames 2 and an acoustic analysis is performed for each frame, and a plurality of feature vectors are obtained from speech data of one utterance. For example, in the case of an utterance of 1 second, since 1 second = 1000 milliseconds, 1000 feature seconds / 10 milliseconds = 100 feature vectors are obtained.

  If the input speech 2 is the first utterance, the correction unit 4 acquires the initial vector 5 stored in the memory 11 of FIG. 2 and corrects the feature vector input from the acoustic analysis unit 3 with the initial vector. The corrected feature vector is output to the matching unit 8. At this time, the initial vector 5 uses an average vector of feature vectors (second feature vectors) of learning data used when learning the acoustic model 7. On the other hand, if the input speech 2 is after the second utterance, the correcting means 4 can calculate the average vector up to the previous utterance, so the input speech 2 obtained by a known method such as Patent Document 1 is used. Is used as a correction vector, the feature vector input from the acoustic analysis means 3 is corrected with the correction vector, and the corrected feature vector is output to the matching means 8.

  The adjusting means 6 adjusts the variance of the parameters of the acoustic model 7 stored in the memory 11 of FIG. 2 if the input speech 2 is the first utterance, and adjusts the first utterance if the input utterance is the second utterance or later. Adjustment is made so that the variance of the parameters of the acoustic model 7 is smaller than that of the utterance.

  The acoustic model 7 is obtained by modeling a statistical feature of a feature vector using a feature vector (second feature vector) of a frequency pattern of voice data uttered by a plurality of speakers. For example, if the speech recognition apparatus 1 performs word recognition, an acoustic model in units of words obtained by connecting acoustic models in units of phonemes is stored. Assuming that the speech recognition apparatus 1 performs speech recognition using a prefecture name as a recognition target vocabulary, for example, the acoustic model of the word Tokyo is phoneme / t /, / o /, / o. It is assumed that acoustic models of phonemes of /, / k /, / j /, / o /, and / o / are connected in order, and an HMM is used. The learning data of the acoustic model 7 is data recorded by various speakers with various microphones having different frequency characteristics. The acoustic model 7 calculates the average vector of the feature vectors of the learning data, and learns the sound of the phoneme unit in advance using data obtained by subtracting the average vector from each feature vector of the learning data.

  The collating unit 8 collates the corrected feature vector input from the correcting unit 4 with the acoustic model 7 adjusted by the adjusting unit 6 and outputs a recognition result 9 from the speech recognition apparatus 1.

Next, the operation of the speech recognition apparatus 1 will be described.
First, the acoustic analysis means 3 acoustically analyzes the input speech 2 and outputs a feature vector to the correction means 4. Then, the correction unit 4 corrects the feature vector according to whether or not the input speech 2 is the first utterance, and outputs the corrected feature vector to the matching unit 8.

Details of the operation of the correction means 4 will be described below.
FIG. 3 is a flowchart showing the operation of the correction means 4 in Embodiment 1 of the present invention.
The correction unit 4 acquires an initial vector 5 stored in the memory 11 of FIG. 2 (step (hereinafter referred to as ST) 1). Next, it is determined whether or not the current input voice 2 is the first utterance (ST2). It is assumed that the correcting means 4 counts the number of utterances of the input speech 2 and determines whether or not the input speech 2 is the first utterance based on the number of utterances.

If the input speech 2 is the first utterance, the acquired initial vector 5, that is, an average vector of feature vectors of learning data used when learning the acoustic model 7 is used as a correction vector (ST3). If the input speech 2 is after the second utterance, an average vector of feature vectors up to the previous utterance temporarily stored is calculated as a correction vector (ST4).
Then, the corrected feature vector obtained by subtracting the correction vector from the feature vector of the input speech 2 is output to the matching means 8 (ST5). Further, the feature vector of the input speech 2 before correction is temporarily stored (ST6).

  As described above, the initial vector used as the correction vector in the first utterance is a vector that is subtracted from the feature vector during learning of the acoustic model 7. Even during recognition, by subtracting the initial vector as the correction vector from the feature vector of the input speech 2, there is an effect that the feature vector of the input speech 2 and the feature vector calculation method at the time of learning are made the same. However, there is no effect of absorbing the difference in frequency characteristics of the microphone during learning and recognition. In order to obtain the effect, it is necessary to obtain an average vector of feature vectors of the input speech 2 by some method and subtract it from the feature vector of the input speech 2 as will be described later. However, in the first utterance, especially the first part of the utterance, it is difficult to calculate a statistically valid average vector because the number of feature vectors of the input speech 2 that can be used for calculating the average vector is small. An average vector of time feature vectors is used.

  Moreover, since the average vector of the feature vectors of the input speech 2 can be calculated after the second utterance, an effect of absorbing the difference in frequency characteristics between the microphones during learning and recognition can be obtained.

The correction vectors for the second and subsequent utterances are the average vector of feature vectors up to the previous utterance temporarily stored in the correction means 4 and a vector obtained by weighted averaging the correction vectors used before the first utterance. A correction vector may be used.
FIG. 4 is a flowchart showing the operation of the correction means 4 in the first embodiment of the present invention.
4, parts that are the same as those in FIG. 3 are given the same numbers as in FIG. The only difference between FIG. 4 and FIG. 3 is that the process of ST3 of FIG. 3 is replaced with ST3a of FIG. 4, and the process of ST6 of FIG. 3 is replaced with the process of ST6a of FIG.

The operation of ST3a is as described above, and the correction vector 4 temporarily stores the average vector of feature vectors up to the previous utterance and the correction vector used before one utterance to obtain a correction vector.
In ST6a, the correction vector is temporarily stored in addition to the feature vector of the input speech 2 before correction.

In this way, when the average vector of feature vectors up to the previous utterance temporarily stored and the correction vector used before one utterance are weighted and averaged to obtain a correction vector, the previous utterance is more emphasized. become. Therefore, even when the speaker changes in the middle, the correction vector can be updated quickly, and the recognition rate can be increased.
The above is a detailed description of the operation of the correction means 4.

Next, the operation of the adjusting means 6 will be described.
FIG. 5 is a flowchart showing the operation of the adjusting means 6 in Embodiment 1 of the present invention.
The adjusting means 6 acquires the acoustic model 7 stored in the memory 11 of FIG. 2 (ST21). Next, it is determined whether or not the current input speech 2 is the first utterance (ST22). If the first utterance is the first utterance, the variance value, which is a parameter of the acoustic model 7, is multiplied by α and adjusted. Is output to the matching means 8 as an acoustic model (ST23). Here, α is a constant (first constant) larger than 1, and a specific value is determined in advance by a recognition experiment or the like. Thus, by multiplying a value larger than 1 to increase the variance value of the acoustic model, even when the feature vector of the input speech 2 fluctuates due to a difference in the frequency characteristics of the microphone, a decrease in acoustic likelihood is suppressed, and the microphone A sound model that is robust against the difference in frequency characteristics is obtained. On the other hand, there is a possibility that the discrimination performance between acoustic models may be lowered. Therefore, it is better to set the constant α by verifying the recognition performance in advance in the recognition experiment.

On the other hand, if the input speech 2 is after the second utterance, the adjusting means 6 multiplies the variance value, which is a parameter of the acoustic model 7, by β, and outputs it to the matching means 8 as an adjusted acoustic model (ST24). Here, β is a constant not less than 1 and not more than α (second constant), or is not less than α and is one or more variables that monotonously decrease according to the number of utterances of the input speech 2. If it is a constant, it is set to 1, for example. If it is a variable, the value is reduced according to the number of utterances. For example, the value of β is calculated by equation (1).

β = MAX (1, α−c * n) (Formula 1)

  In Expression (1), MAX (1, α-c * n) is an operator that selects the larger of 1 and α-c * n. C is an experimentally determined constant, and n is the number of input voices.

When the input speech 2 is the second utterance or later, the variance value of the acoustic model is made the same as that of the original acoustic model stored in the memory 11 or the variance value is made closer to the number of utterances. The discrimination performance is the same as that of the original acoustic model stored in the memory 11 or can be close to the discrimination performance of the acoustic model according to the number of utterances.
Note that, as the variance value of the acoustic model 7 is reduced in this way, the robustness against the difference in the frequency characteristics of the microphones is lowered. However, after the second utterance, the correction means 4 performs learning as described above. No problem arises because processing is performed to obtain the effect of absorbing the difference in microphone frequency characteristics at the time of recognition.
The above is the description of the operation of the adjusting unit 6.

Next, the operation of the verification unit 8 will be described.
When the corrected feature vector is input, the matching unit 8 performs pattern matching on the corrected feature vector using the adjusted acoustic model input from the adjusting unit 6, and the acoustic having the highest likelihood is obtained. The vocabulary of the model is output from the speech recognition apparatus 1 as the recognition result 9. For example, a Viterbi algorithm is used as a pattern matching method.

  As described above, in the speech recognition apparatus 1 of the present embodiment, for the first utterance, the average vector of the first feature vector obtained by analyzing the input speech 2 and the average vector during learning of the acoustic model Is corrected to the first feature vector, and pattern matching is performed on the corrected first feature vector using the acoustic model having the variance value increased by the adjusting means 6, so that the recognition end time for the first utterance Even if the frequency characteristics of the microphone are different between the learning and recognition of the acoustic model without delay, robust speech recognition is possible. Furthermore, for the second and subsequent utterances, the speech recognition performance utilizing the learning effect of the acoustic model can be maintained by correcting the difference between the average vectors during learning and recognition.

  Further, in Patent Document 1, when the input utterance length is short, the accuracy of the average vector is lowered and the recognition performance is lowered. However, in the speech recognition apparatus according to the present invention, the input speech is analyzed according to the learning of the acoustic model. If the feature vector is corrected and the first utterance, the variance value, which is a parameter of the acoustic model, is increased. Therefore, when the acoustic model is learned and recognized, the frequency characteristics of the microphone are different and the input utterance length is short. However, robust speech recognition is possible. Furthermore, for the second and subsequent utterances, the speech recognition performance utilizing the learning effect of the acoustic model can be maintained by correcting the difference between the average vectors during learning and recognition.

  In Patent Document 1, since the average vector is obtained for the entire learning data, the CMN does not work in the first utterance, and the recognition performance deteriorates when the frequency characteristics of the microphone are different between learning and recognition. In the speech recognition apparatus according to the invention, if the first utterance, the variance value that is a parameter of the acoustic model is increased. Therefore, the frequency characteristics of the microphone are different between learning and recognition of the acoustic model, and the first input speech Robust voice recognition is possible even for utterances. Furthermore, for the second and subsequent utterances, the speech recognition performance utilizing the learning effect of the acoustic model can be maintained by correcting the difference between the average vectors during learning and recognition.

In the first embodiment, an average vector of feature vectors of learning data is calculated for the acoustic model 7 and the acoustic model in units of phonemes is learned in advance by data obtained by subtracting the average vector from each feature vector of the learning data. In the above description, the case where CMN is performed has been described.
In the case of the acoustic model 7 that does not perform CMN at the time of learning, since the acoustic model 7 does not subtract the average vector of the feature vectors of the learning data, the correcting means 4 is the feature of the input speech 2 for the first utterance. The average vector of input speech 2 is not subtracted from the vector. Furthermore, for the second and subsequent utterances, the average vector of the feature vectors from the feature vector of the input speech 2 to the previous utterance is subtracted, and the average vector of the learning data of the acoustic model 7 is added.
As described above, the correction unit 4 corrects the difference between the average vector during learning of the acoustic model 7 and the average vector of the feature vector of the input speech 2 to the feature vector of the input speech 2 in accordance with the learning of the acoustic model 7. To do.

The acoustic model 7 includes learning data for each speaker created by subtracting an average vector of all the second feature vectors for each speaker from the second feature vector for each speaker, and all speakers. Both the learning data of all speakers created by subtracting the average of all the second feature vectors of all speakers from the second feature vector may be used as learning data.
The learning data for each speaker has an effect of improving the recognition performance by suppressing the variation of the feature vector due to the difference between the microphone and the speaker and learning the acoustic model with high accuracy. On the other hand, the all-speaker learning data has the same features as the original learning data because it is simply a uniform subtraction of the feature vectors of the learning data. Since the original learning data includes speaker feature vectors recorded using microphones having various frequency characteristics, there is an effect of constructing an acoustic model that is robust against differences between microphones and speakers. Also, by subtracting the average vector of all speakers from the feature vectors of all speakers, the feature vectors of all speakers can be roughly aligned with the feature vectors of the learning data for each speaker. It is possible to learn an acoustic model that has both.

When speech recognition is performed, as described above, by using the acoustic model learned using both the learning data for each speaker and the learning data for all speakers, the correcting means 4 is the first utterance. Because the average vector of all the feature vectors of all speakers is used as the correction vector, the acoustic model with the variance value adjusted is compared with the corrected feature vector obtained by subtracting the correction vector from the input speech feature vector, so the recognition ends. The effect of learning the acoustic model using the learning data of all the speakers created by subtracting the average of all the second feature vectors of all the speakers without any time delay It becomes robust and can secure the accuracy of voice recognition. Further, for the second and subsequent utterances, the correction means 4 uses the average vector of the feature vectors up to the utterance before the input speech 2 as a correction vector, so that all the second feature vectors for each speaker are used. By learning the acoustic model using the learning data for each speaker created by subtracting the average vector of the speakers, learning was performed so that the CMN effect could be sufficiently exerted on the feature vectors resulting from differences in microphones and speakers. Since the dispersion value is adjusted so that the acoustic model is the same as or close to the acoustic model, the recognition performance can be improved.

Embodiment 2. FIG.
The speech recognition apparatus 1 according to Embodiment 2 of the present invention is the same as the functional configuration of the speech recognition apparatus 1 of FIG. 1 according to Embodiment 1 and the hardware configuration of FIG. A configuration different from the first embodiment is the operation of the adjusting means 6. The operation of the other constituent means is the same as in the first embodiment.

The operation of the adjusting means 6 in this embodiment will be described.
FIG. 6 is a flowchart showing the operation of the adjusting means 6 in Embodiment 1 of the present invention.
The adjusting means 6 acquires the acoustic model 7 (ST21). Next, it is determined whether or not the current input voice 2 is the first utterance (ST22).
When the input speech 2 is the first utterance, among the variance values that are the parameters of the acoustic model 7, some of the n1 to n2 dimensional values are multiplied by α, and the adjusted acoustic model is input to the matching unit 8. Output (ST23a).

Here, n1 and n2 are constants satisfying 0 <= n1 <= n2 <N. N is the number of dimensions of the feature vector. α is a constant larger than 1, and a specific value is determined in advance by a recognition experiment or the like. In this way, by multiplying the low-dimensional variance value by a value larger than 1 to increase the variance value of the acoustic model 7, the acoustic likelihood even when the feature vector of the input speech 2 fluctuates due to a difference in the frequency characteristics of the microphone or the like. Can be suppressed, and an acoustic model robust to the difference in the frequency characteristics of the microphone can be obtained. On the other hand, there is a possibility that the discrimination performance between acoustic models may be lowered. Therefore, it is better to set the constant α by verifying the recognition performance in a prior recognition experiment.

The effect of limiting the dimension for adjusting the dispersion value from n1 to n2 is as follows. When the frequency characteristics of the microphones are different between the learning and recognition of the acoustic model, the frequency characteristic difference between the microphones often affects the low-dimensional value of the feature vector MFCC, and the high-dimensional MFCC does not change much. There are many cases. In this case, the high-dimensional variance value is not adjusted, and the variance value of the original acoustic model is used as it is, so that the acoustic model identification performance is prevented from being lowered more than necessary.

Next, the operation of the adjusting unit 6 after the input voice 2 is the second utterance will be described.
After the second utterance, the adjusting unit 6 multiplies a part of n1 to n2 dimensional values among the variance values that are parameters of the acoustic model 7 and outputs the result to the matching unit 8 as an adjusted acoustic model 7. (ST24a).
Here, β is a constant not less than 1 and not more than α, or is not less than α and one or more variables that are not more than α and monotonously decrease according to the number of input utterances. If it is a constant, it is set to 1, for example. If it is a variable, the value is reduced according to the number of utterances. For example, as in the first embodiment, the value of β is calculated by equation (1).

As described above, in the second and subsequent utterances, by making the variance value of the acoustic model 7 the same as the original acoustic model 7 obtained by learning, or by making the variance value closer according to the number of utterances, the acoustic model 7 can be made closer to the original acoustic model 7 obtained by learning the discrimination performance or according to the number of utterances.

  If the variance value of the acoustic model is reduced in this way, the robustness against the difference in the frequency characteristic of the microphone is lowered. However, in the second and subsequent utterances, as described above, the correction means 4 performs learning and recognition. Since the effect of absorbing the difference in the microphone frequency characteristics can be obtained, there is no problem.

  As described above, in the speech recognition device 1 of the present embodiment, the low-dimensional variance value among the variance values that are parameters of the acoustic model 7 depending on whether or not the input speech 2 is the first utterance. Is adjusted, and the adjusted acoustic model 7 and the input speech 2 are collated with the feature vector corrected in accordance with the learning of the acoustic model 7, and the recognition result is output. The dispersion value is increased in the case of, and the adjustment of the dispersion value is not performed in a high dimension where there is not much frequency characteristic difference between the microphones. Therefore, it is possible to absorb the difference in the microphone frequency characteristics between learning and recognition without lowering the identification performance of the acoustic model more than necessary, and there is an effect of improving the recognition performance.

  As described above, the speech recognition apparatus according to the present invention calculates the difference between the average vector during learning of the acoustic model and the average vector of the input speech with respect to the feature vector of the first utterance of the input speech. Various microphones with different frequency characteristics are corrected because the acoustic model with the dispersion value adjusted according to whether or not the input speech is the first utterance is compared with the feature vector of the input speech after the correction. On the other hand, the speech recognition performance that improves the speech recognition performance of the first utterance and makes use of the learning effect of the acoustic model by correcting the difference between the average vectors during learning and recognition after the second utterance Can keep.

DESCRIPTION OF SYMBOLS 1 Speech recognition apparatus, 2 input speech, 3 acoustic analysis means, 4 correction means, 5 initial vector, 6 adjustment means, 7 acoustic model, 8 collation means, 9 recognition result, 10 processor, 11 memory

Claims (8)

Analyzing means for analyzing the input speech and outputting a first feature vector;
An acoustic model that models the second feature vector of speech data uttered by a plurality of speakers ;
Mean vector or an average vector of the first feature vector of the input speech, the first of said corrected by subtracting from the feature vectors a first feature vector correction means for outputting the acoustic model,
Adjusting means for adjusting a variance value of a parameter of the acoustic model according to whether or not the voice is the first utterance;
A speech recognition apparatus comprising: collation means for collating the acoustic model with the variance value adjusted and the corrected first feature vector to output the speech recognition result.   The speech recognition apparatus according to claim 1, wherein the adjustment unit multiplies the variance value by a first constant larger than 1 if the speech is the first utterance.   3. The adjustment unit according to claim 2, wherein if the voice is after the second utterance, the adjustment unit multiplies the variance value by a second constant that is greater than or equal to 1 and less than or equal to the first constant. Voice recognition device.   3. The adjustment means multiplies the variance value by a variable that is not less than 1 and not more than the first constant and monotonously decreases according to the number of input speech if the speech is after the second utterance. The speech recognition apparatus described in 1. The acoustic model is multi-dimensional data,
The speech recognition apparatus according to claim 1, wherein the adjustment unit adjusts the variance value of a part of the dimensions of the acoustic model. The acoustic model is data learned using data obtained by subtracting an average of all the second feature vectors of all speakers from the second feature vector,
If the speech is the first utterance, the correction means uses the average vector of the second feature vectors as a correction vector, subtracts the correction vector from the first feature vector, and corrects the first The first feature vector before correction is temporarily stored, and if the voice is after the second utterance, the pre-correction before the temporarily stored utterance is stored. An average vector of the first feature vectors is used as a correction vector, the correction vector is subtracted from the first feature vector, the corrected first feature vector is output, and the first feature vector before correction is output. Is temporarily stored, The speech recognition apparatus according to any one of claims 1 to 5. The acoustic model includes learning data for each speaker learned by data obtained by subtracting an average vector of all the second feature vectors for each speaker from the second feature vector for each speaker, Learning data of all speakers learned from data obtained by subtracting the average of all the second feature vectors of all speakers from the second feature vector,
If the voice is the first utterance, the correction means uses an average vector of all the second feature vectors of all speakers as a correction vector, and subtracts the correction vector from the first feature vector. The first feature vector after correction is output, the first feature vector before correction is temporarily stored, and if the speech is after the second utterance, the temporarily stored previous utterance The average vector of the first feature vectors before correction up to is used as a correction vector, the correction vector is subtracted from the first feature vector, and the corrected first feature vector is output. The speech recognition apparatus according to claim 1, wherein the first feature vector is temporarily stored. In a speech recognition method of a speech recognition device that performs speech recognition of input speech,
An analysis step of analyzing the input speech and outputting a first feature vector;
An average vector of an acoustic model obtained by modeling a second feature vector of speech data uttered by a plurality of speakers or an average vector of the first feature vector of the input speech is subtracted from the first feature vector. And a correction step of outputting the first feature vector after correction,
An adjustment step of adjusting a variance value of a parameter of the acoustic model according to whether or not the voice is the first utterance;
A speech recognition method, comprising: a collation step of collating the acoustic model with the adjusted dispersion value and the corrected first feature vector.

Images