JP6466762B2 - Speech recognition apparatus, speech recognition method, and program - Google Patents

Description

  The present invention relates to a technology for automatically determining parameters of a speech recognition model and performing speech recognition using the parameters.

  The speech recognition model expressing the decoding by the GMM-HMM speech recognition device includes one language weight parameter (model parameter). This model parameter is an empirical parameter (heuristic parameter), and it is necessary to set this parameter value for practical use (see, for example, Non-Patent Document 1). Also, an ANN-HMM Hybrid speech recognition device is known that can handle more input information than a GMM-HMM speech recognition device by using a neural network instead of a mixed normal distribution as an acoustic model. The speech recognition model expressing the decoding by the ANN-HMM Hybrid speech recognition device includes two heuristic parameters, and these parameter values need to be set (see, for example, Non-Patent Document 2).

  Further, as a conventional method for automatically setting heuristic parameters in the GMM-HMM speech recognition apparatus, there is a method described in Non-Patent Document 3.

Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Mikio Yamamoto, “IT Text Speech Recognition System”, Ohmsha (2001): 104-105. Dahl, George E., et al, “Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition,” Audio, Speech, and Language Processing, IEEE Transactions on 20.1 (2012): 30-42. Mak, Brian, and Tom Ko, “Min-max discriminative training of decoding parameters using iterative linear programming,” INTERSPEECH, 2008.

  However, in the method of Non-Patent Document 3, the parameter value is set without considering the noise component included in the input acoustic signal. Therefore, in the case of a speech recognition model in which appropriate parameter values differ depending on the noise component, the speech recognition accuracy or speech recognition rate may be reduced due to the influence of the noise component. Although it is conceivable to perform noise suppression processing on the input acoustic signal in advance, noise distortion processing may add inappropriate distortion from the viewpoint of speech recognition.

  An object of the present invention is to automatically set an appropriate model parameter according to a noise component included in an input acoustic signal.

  Select the model parameter of the speech recognition model or the combination of model parameters of the speech recognition model according to the relationship between the magnitude of the speech component and noise component contained in the input acoustic signal, and according to the selected model parameter or combination of model parameters Apply the speech recognition model to the input acoustic signal.

  Thereby, an appropriate model parameter according to the noise component contained in the input acoustic signal can be automatically set.

FIG. 1 is a block diagram illustrating a functional configuration of the learning device according to the embodiment. FIG. 2 is a block diagram illustrating a functional configuration of the speech recognition apparatus according to the embodiment. FIG. 3 is a flowchart for illustrating the correspondence table generation processing according to the embodiment. FIG. 4 is a flowchart for illustrating the speech recognition processing according to the embodiment. 5A and 5C are diagrams for illustrating the correspondence table of the embodiment, and FIG. 5B is a diagram for illustrating the correspondence relationship between the S / N ratio of the correspondence table and the S / N ratio of the input acoustic signal. .

Embodiments of the present invention will be described below.
[Overview]
First, an outline of the embodiment will be described.
In each embodiment, the speech recognition apparatus determines whether the speech recognition model model parameter or the speech recognition model model corresponds to the relationship between the “speech component” and the “noise component” included in the “input acoustic signal” to be speech recognition target. A combination of parameters is selected and a speech recognition model corresponding to the selected model parameter or combination of model parameters is applied to the input acoustic signal. By selecting model parameters according to the “relation between speech and noise components”, appropriate model parameters can be automatically set even for speech recognition models with different appropriate parameter values according to “noise components”. . The “relationship between the size of the speech component and the noise component” is, for example, a relative value between the size of the “speech component” and the size of the “noise component” or a function value of the relative value. The “relationship between the size of the speech component and the noise component” may be, for example, a ratio (S / N ratio) of the size of the “sound component” to the size of the “noise component”. The ratio of the size of the “noise component” to the size of the “noise component” may be the ratio of the “noise component” to the size of the “acoustic signal” including the “voice component” and the “noise component”. It may be the ratio of the size of the “sound component” to the size of the “acoustic signal”, or a value obtained by subtracting the size of the “sound component” from the size of the “acoustic signal”. It may be a value obtained by subtracting the magnitude of the “noise component” from the magnitude of the “acoustic signal”, or may be a function value of such a ratio or value. A specific example of the “speech recognition model” may be a speech recognition model that expresses decoding by the GMM-HMM speech recognition device (see, for example, Non-Patent Document 1), or decoding by the ANN-HMM Hybrid speech recognition device. (For example, refer to Non-Patent Document 2 etc.) or other speech recognition models. The “model parameter” is, for example, a “heuristic parameter”. However, at least a part of the “model parameter” may be a parameter other than the “heuristic parameter”.

  In each embodiment, speech recognition accuracy or speech recognition when applied to an acoustic signal having a relationship represented by “first index” and “first index” representing “a relationship between the magnitudes of speech components and noise components” A model parameter or a combination of model parameters of a speech recognition model having a maximum rate or a predetermined value or more is associated with the rate. The speech recognition apparatus selects a model parameter or a combination of model parameters associated with the “first index” corresponding to “a relationship between the magnitudes of the speech component and the noise component” included in the “input acoustic signal”, A speech recognition model using the selected model parameter or a combination of model parameters is applied to the “input acoustic signal”. Accordingly, high speech recognition accuracy or speech recognition rate can be realized in accordance with the “relationship between the magnitudes of the speech component and the noise component”. The example of the “first index” is an index that represents the above-described example of “the relationship between the magnitudes of the voice component and the noise component”, and is, for example, a value that represents the S / N ratio.

  More specifically, the speech recognition apparatus corresponds to the “first index” that is closest to the “second index” that represents “the relationship between the magnitudes of the speech component and the noise component” included in the “input acoustic signal”. Select model parameters or model parameter combinations. As a result, even when there is no “first index” that matches the “second index”, it is possible to select a model parameter or a combination of model parameters that can realize high speech recognition accuracy or speech recognition rate.

  Alternatively, the speech recognition apparatus selects a model parameter or a combination of model parameters corresponding to the “first index” that is the same as or close to the “second index” and that maximizes the speech recognition accuracy or the speech recognition rate. May be. The “first index in the vicinity of the second index” means a “first index” whose distance from the “second index” is within a predetermined range. There may be only one “first index in the vicinity of the second index”, two, or three or more. Examples of the “first index in the vicinity of the second index” are the “second index” that is larger than the “second index” and closest to the “second index”, and smaller than the “second index”. The “first index” closest to the “index”, both of them, three or more “first indices” whose distance from the “second index” is within a predetermined value, and the like.

  Alternatively, the speech recognition apparatus maximizes the “weighting value” of the speech recognition accuracy or speech recognition rate among the model parameters or the combination of model parameters corresponding to the “first index” that is the same as or close to the “second index”. You may choose what to do. The “weighting value” is a value obtained by multiplying the voice recognition accuracy or the voice recognition rate by a larger weight as the distance between the “second index” and the “first index” is smaller. Thus, an appropriate model parameter or combination of model parameters can be selected based on both the distance between the “second index” and the “first index” and the indexes of both the speech recognition accuracy and the speech recognition rate.

  Further, not the discrete “first index” and the model parameter or the combination of model parameters, but their complementary values may be used. That is, the speech recognition apparatus determines whether the “first index” or its complementary value matches the “second index”, and the model parameter associated with the “first index” that matches the “second index”. Alternatively, a combination of model parameters, or a complementary value of a model parameter or a combination of complementary values of a model parameter corresponding to the complementary value of the “first index” that matches the “second index” may be selected. In other words, the speech recognition apparatus complements the “first index” or the “first index” that matches the “second index” representing the “relationship between the magnitudes of the speech component and the noise component” included in the input acoustic signal. A model parameter corresponding to the value or a complementary value of the model parameter or a combination of the model parameter or the complementary value of the model parameter may be selected. There is no limitation on the complementing method, and for example, known methods such as linear complementing, polynomial complementing, and spline complementing can be used.

[First Embodiment]
Next, the first embodiment will be described.
<Configuration>
As illustrated in FIG. 1, the learning device 11 of the present embodiment includes a speech signal storage unit 111a, a noise signal storage unit 111b, a correct word string storage unit 111c, a noise-added speech signal storage unit 111e, a speech recognition result storage unit 111f, The component adjustment adding unit 113, the index generating unit 114, the voice recognizing unit 116, the comparing unit 117, and the correspondence table generating unit 118 are included. As illustrated in FIG. 2, the speech recognition apparatus 12 according to the present exemplary embodiment includes a correspondence table storage unit 121a, an input acoustic signal storage unit 121b, a noise component storage unit 121c, an input unit 122, a speech / non-speech segment determination unit 123, an index. A generation unit 124, a selection unit 125, and a voice recognition unit 126 are included. Each device is, for example, a general-purpose or dedicated computer provided with a processor (hardware processor) such as a CPU (central processing unit) and a memory such as RAM (random-access memory) and ROM (read-only memory). It is configured by executing the program. The computer may include a single processor and memory, or may include a plurality of processors and memory. This program may be installed in a computer, or may be recorded in a ROM or the like in advance. In addition, some or all of the processing units are configured using an electronic circuit that realizes a processing function without using a program, instead of an electronic circuit (circuitry) that realizes a functional configuration by reading a program like a CPU. May be. In addition, an electronic circuit constituting one device may include a plurality of CPUs.

<Learning process>
The process of the learning apparatus 11 is demonstrated using FIG. The learning device 11 uses a “relationship between magnitudes of speech components and noise components (for example, S / N ratio)” and model parameters of speech recognition models or a combination of model parameters (for example, ANN-HMM Hybrid model speech recognition devices). Of heuristic parameters) and the influence of the combination on the speech recognition accuracy or speech recognition rate. In other words, the learning device 11 sets a plurality of “first indicators” representing “a relationship between the magnitudes of the speech component and the noise component”, and each has a “relationship between the sizes of the speech component and the noise component”. Correspondence table that obtains a model parameter or a combination of model parameters of a speech recognition model whose speech recognition accuracy or speech recognition rate when applied to a signal is the maximum or a predetermined value or more and associates them with the corresponding “first index” And output.

  As a premise of the learning process, a time-series audio signal is stored in the audio signal storage unit 111a, and a noise signal is stored in the noise signal storage unit 111b. The voice signal may be obtained by recording voice (for example, speech voice) in a silent environment in advance, or may be generated by a voice synthesis technique. The noise signal may be recorded in advance, or may be generated by a noise generation algorithm (for example, white noise). The correct word string storage unit 111c stores a correct word string of the audio signal stored in the audio signal storage unit 111a.

As illustrated in FIG. 3, the component adjustment adding unit 113 reads the audio signal and the noise signal from the audio signal storage unit 111a and the noise signal storage unit 111b, respectively, and “the relationship between the magnitudes of the audio component and the noise component” is α as a _i obtain noise with audio signal X _i is a time-series signal by adding them. However, i = 0,..., I-1, and I is an integer of 2 or more. That is, component adjustment adder 113, a plurality of types of "size of the relationship between speech and noise components" alpha _0, · · ·, adds the audio signals and noise signals in alpha _I-1, a plurality of types of noise Accompanying audio signals X ₀ ,..., X _I-1 are obtained. For example, the component adjustment adding unit 113 adds the audio signal X _i with noise obtained by adding the audio signal and the noise signal so that the S / N ratio is α _i for each of i = 0,. Output. In other words, the component adjustment adding unit 113 adds a sound signal and a noise signal with a plurality of types of S / N ratios α ₀ ,..., Α _I−1 , for example, and adds a plurality of types of sound signals with noise X ₀ ,. ..., X _I-1 is obtained. For example, α ₀ ,..., Α _I-1 are discrete values different from each other. The S / N ratio may be an S / N ratio at each time point, an average S / N ratio in each time interval, or an average S / N ratio in all time intervals. May be. The “relationship between the magnitude of the audio component and the noise component” such as the S / N ratio may be determined from the effective value of the audio signal and the effective value of the noise signal, or an average value or maximum value instead of the effective value. Or you may determine from an absolute value. For example, (the effective value of the audio signal) / (the effective value of the noise signal) may be used as the S / N ratio, or an average value, a maximum value, or an absolute value may be used instead of the effective value. There is no limitation on the generation method of such a noise-added audio signal X _i . For example, the component adjustment adder 113 converts the noise signal read from the noise signal storage unit 111 b into the noise signal read from the audio signal storage unit 111 a. A noise component N _i multiplied by a coefficient corresponding to α _i (where i = 0,..., I−1) is added to obtain a speech signal X _i with noise. The component adjustment adding unit 113 stores the noise-added audio signal X _i in the noise-added audio signal storage unit 111e, and also includes the noise-added audio signal X _i and the noise component N _i (where i = 0,..., I− 1) is sent to the index generation unit 114 (step S113).

The index generation unit 114 receives the noise-added speech signal X _i and the noise component N _i (where i = 0,..., I−1) as input, and the signal execution value of the noise-added speech signal X _i and the noise component N. _From the signal execution value of _i , a “first index” r _i representing “a relationship between the magnitudes of the speech component and the noise component” is newly obtained and output. For example, the index generation unit 114 newly obtains and outputs the S / N ratio r _i from the signal execution value of the noise-added speech signal X _{i and} the signal execution value of the noise component N _i . For example, r _i = (signal execution value of the noise signal X _i with noise−signal execution value of the noise component N _i ) / (signal execution value of the noise component N _i ), or an average instead of this signal execution value A value or a maximum or absolute value may be used. By determining r _i from the noise-added speech signal X _i and the noise component N _i , the “relationship between the size of the speech component and the noise component obtained from the speech interval signal and the non-speech interval signal of the noise-added speech signal X _i Can be obtained (step S114). The obtained r _i is stored in the noisy audio signal storage unit 111e in association with the noisy audio signal X _i (step S111e).

The model parameter setting unit 109 sets a plurality of types of model parameters or combinations of model parameters of a predetermined speech recognition model, and outputs them to the speech recognition unit 116. Hereinafter, the combination of the model parameters or model parameters is expressed as h _m. However, m = 0,..., M−1, and M is an integer of 2 or more. If h _m is the model parameter, h _m is a scalar representing the parameter values, if h _m is a combination of the model parameters, h _m is a vector of parameter values as elements.

ここで、ｔは時刻、ｘ_ｔは各時刻ｔの音響特徴量ベクトル、ｓはHMM状態系列、ｓ_ｔは各時刻ｔのHMM状態、ｗは単語列、Ｐ（ｓ_ｔ｜ｘ_ｔ）はニューラルネットワークによる音響モデル、Ｐ（ｓ_ｔ）は各HMM状態に関するユニグラムモデルであるHMM State Unigramモデル、Ｐ（ｓ｜ｗ）は辞書、Ｐ（ｗ）は言語モデルを表している。βおよびγはヒューリスティックパラメータ（モデルパラメータ）である。この例の場合、モデルパラメータ設定部１０９は、βおよびγの組み合わせ（βおよびγを要素とするベクトル）ｈ_ｍを複数種類設定する。例えば、所定の範囲内のβ，γから取り得るすべてのβおよびγの組み合わせｈ_ｍ（ただし、ｍ＝０，・・・，Ｍ−１）を設定する。 A specific example of the speech recognition model is the following speech recognition model that expresses decoding by the ANN-HMM Hybrid speech recognition device (see, for example, Non-Patent Document 2).

Here, t is time, _{x t} is the acoustic feature vector, s is HMM state sequence at each time t, _{s t} is HMM states at each time t, w is the word _{sequence, P (s t | x t} ) is a neural A network acoustic model, P (s _t ) represents a HMM State Unigram model that is a unigram model for each HMM state, P (s | w) represents a dictionary, and P (w) represents a language model. β and γ are heuristic parameters (model parameters). In this example, the model parameter setting unit 109, (vector and the β and gamma components) a combination of β and gamma h _m a multiple type setting. For example, beta within a predetermined range, all the beta and combinations of gamma h _m can take from gamma _(although, m = 0, ···, M -1) to set the.

ここで、Ｐ（ｘ_ｔ｜ｓ_ｔ）は混合正規分布による音響モデルである。この例の場合、モデルパラメータ設定部１０９は、ｈ_ｍ＝βを複数種類設定する。例えば、所定の範囲内のすべてのβをｈ_ｍ（ただし、ｍ＝０，・・・，Ｍ−１）として設定する（ステップＳ１１４）。 As the speech recognition model, the following speech recognition model that expresses decoding by the GMM-HMM speech recognition device may be used (see, for example, Non-Patent Document 1).

Here, P (x _t | s _t ) is an acoustic model with a mixed normal distribution. In the case of this example, the model parameter setting unit 109 sets plural types of h _m = β. For example, all β within a predetermined range are set as h _m (where m = 0,..., M−1) (step S114).

The speech recognition unit 116, from the noise with sound signal storage unit 111e _{r i} and _{X i} (however, i = 0, ···, I -1) reads, _{h m} (but sent from the model parameter setting portion 109 , M = 0,..., M−1) is used for speech recognition of X _i and a word string that is the speech recognition result is output. Speech recognition result is obtained for all combinations of (i, m), the speech recognition result obtained is stored in association with the r _i and (i, m) in the speech recognition result storage unit 111f (step S115 ).

The comparison unit 117 compares the correct word string read from the correct word string storage unit 111c with the speech recognition result read from the speech recognition result storage unit 111f, and the speech recognition accuracy of the speech recognition result for each (i, m). Ask for. Alternatively, the comparison unit 117 may obtain the speech recognition rate for each (i, m) instead of the speech recognition accuracy. The resulting speech recognition accuracy or the speech recognition rate is sent corresponding (i, m) and with _{r i} in the correspondence table generation unit 118 (step S116).

The correspondence table generation unit 118 selects m (i) ε {0,..., M−1} that maximizes the speech recognition accuracy or speech recognition rate for each i. Alternatively, the correspondence table generation unit 118 may select one m (i) ε {0,..., M−1} for which the speech recognition accuracy or the speech recognition rate is greater than or equal to a predetermined value for each i. . Correspondence table generation unit 118, the correspondence table _{[r i, h m (i} )] that associates _{r i} and _{h m (i)} to generate and output. FIG. 5A is an example of a correspondence table [r _i , h _{m (i)} ] when I = 8, r _i is the S / N ratio, and h _{m (i)} is a combination of model parameters β and γ ( Step S117). Correspondence table _{[r i, h m (i} )] are stored in the correspondence table storage unit 121a of the voice recognition device 12 (FIG. 2).

<Voice recognition processing>
The process of the speech recognition apparatus 12 will be described with reference to FIG. The speech recognition device 12 selects a model parameter of the speech recognition model or a combination of model parameters of the speech recognition model according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal. That is, a model parameter or a combination of model parameters _{hm (i)} corresponding to the “first index” r _i corresponding to the magnitude relationship between the speech component and the noise component included in the input acoustic signal is selected. The speech recognition device 12 performs speech recognition by applying a speech recognition model corresponding to the selected model parameter or combination of model parameters to the input acoustic signal.

  First, an input acoustic signal is input to the input unit 122 and stored in the input acoustic signal storage unit 121b. The input acoustic signal is a time-series signal, for example, an audio signal on which a noise component is superimposed (step S121). The voice / non-speech section discriminating unit 123 reads the input acoustic signal from the input acoustic signal storage unit 121b and discriminates between the voice section and the non-speech section of the input acoustic signal. For this determination, for example, Reference 1 (Jongseo Sohn, Nam Soo Kim, Wonyong Sung, “A Statistic Model-Based Voice Activity Detection,” IEEE SIGNAL PROCESSING LETTERS, VOL.6, NO.1, 1999.) etc. A well-known method is used. The signal in the non-voice section is stored as a noise component in the noise component storage unit 121c, and the input acoustic signal is sent to the index generation unit 124 (step S122).

  The index generation unit 124 obtains a “second index” u representing “a relationship between the magnitudes of the voice component and the noise component” by using the input acoustic signal and the signal of the non-voice section read from the noise component storage unit 121c. Output. The “relationship between the size of the speech component and the noise component” in the speech recognition process is preferably based on the same standard as the “relationship between the size of the speech component and the noise component” in the learning process described above. That is, when the S / N ratio is used in the learning process as “the relationship between the magnitudes of the voice component and the noise component”, it is desirable that the S / N ratio is also used in the voice recognition process. U may be obtained at each time point, u may be obtained for each predetermined time interval, or u may be obtained for all time intervals of the input acoustic signal. It may be determined from the effective value of the input acoustic signal and the effective value of the signal in the non-speech section, or may be determined from an average value, a maximum value, or an absolute value instead of the execution value. For example, u = (signal effective value of input acoustic signal−signal effective value of signal in non-speech section) / (signal effective value of signal in non-speech section), or an average value or It may be determined from a maximum value or an absolute value. The obtained u is sent to the selection unit 125 (step S123).

Selecting unit 125, the correspondence table _{[r i, h m (i} )] stored in the correspondence table storage unit 121a with reference to, corresponding to the "second index" closest "first index" to u _{r i} Model A parameter or model parameter combination hm _(i) is selected. For example, in the example of FIGS. 5A and 5B, since r ₂ is closest to u, a combination of model parameters h _{m (2)} = (γ ₅ , β ₃ ) corresponding to r ₂ is selected. If u is an intermediate value of the two r _i adjacent, so that the closest r _i in u there are two. In such a case, for example, hm _(i) corresponding to one of the predetermined r _i is selected. Since r _i is obtained from the noise-added speech signal X _i and the noise component N _i , it is obtained from the input acoustic signal (corresponding to the noise-added speech signal) and the non-speech interval signal (corresponding to the noise component). An appropriate _{hm (i)} can be selected for u. The selected _{hm (i)} is sent to the speech recognition unit 126 (step S124).

The speech recognition unit 126 performs speech recognition by applying the speech recognition model using the sent model parameter or model parameter combination hm _(i) to the input acoustic signal read from the input acoustic signal storage unit 121b, The voice recognition result is output (step S125).

<Features of this embodiment>
In this embodiment, since the model parameter of the speech recognition model or the combination of model parameters of the speech recognition model is selected according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, it is included in the input acoustic signal. Appropriate model parameters can be automatically set according to noise components.

  In particular, ANN-HMM Hybrid speech recognition has the property of being influenced by noise components included in the input acoustic signal when determining parameter values. Therefore, it is difficult to apply the same method to the ANN-HMM Hybrid speech recognition as the conventional heuristic parameter automatic determination method that does not consider the noise component in the conventional GMM-HMM speech recognition (see Non-Patent Document 3, for example). It was necessary to set the parameter value. Although noise suppression processing may be performed on the input acoustic signal in advance, generally, distortion that is not suitable from the viewpoint of speech recognition is added to the speech by these noise suppression processing. Therefore, it is considered that learning a neural network that directly discriminates the HMM state from an input acoustic signal including a noise component is more suitable in terms of performing processing considering speech recognition. The method of this embodiment can automatically determine the most suitable heuristic parameters for the speech recognition accuracy or speech recognition rate from the input acoustic signal, eliminates manual setting work, and reduces the speech recognition accuracy or speech recognition rate due to noise components. Decline can be prevented.

[Second Embodiment]
Next, a second embodiment will be described. The second embodiment is a modification of the first embodiment. In the present embodiment, among the model parameters or combinations of model parameters corresponding to the “first index” that is the same as or close to the “second index” that represents the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal Select the one that maximizes speech recognition accuracy or speech recognition rate. In the following, differences from the items described so far will be mainly described, and the items already described will be simplified by quoting the same reference numerals.

<Configuration>
As illustrated in FIG. 1, the learning device 21 according to the present embodiment includes a speech signal storage unit 111a, a noise signal storage unit 111b, a correct word string storage unit 111c, a speech signal storage unit with noise 111e, a speech recognition result storage unit 111f, A component adjustment adding unit 113, an index generating unit 114, a speech recognition unit 116, a comparing unit 117, and a correspondence table generating unit 218 are included. As illustrated in FIG. 2, the speech recognition apparatus 22 according to the present exemplary embodiment includes a correspondence table storage unit 221a, an input acoustic signal storage unit 121b, a noise component storage unit 121c, an input unit 122, a speech / non-speech segment determination unit 123, an index. A generation unit 124, a selection unit 225, and a voice recognition unit 126 are included.

<Learning process>
The only difference from the first embodiment is that the process of step S217 is performed instead of step S117 of FIG. In step S217, the correspondence table generation unit 218 selects m (i) ε {0,..., M−1} that maximizes the speech recognition accuracy or speech recognition rate for each i, and this speech recognition accuracy or Let the maximum value of the speech recognition rate be a _i . Alternatively, the correspondence table generation unit 218 selects one m (i) ε {0,..., M−1} for which the voice recognition accuracy or the voice recognition rate is greater than or equal to a predetermined value for each i, and this voice Let a _i be the recognition accuracy or speech recognition rate. Correspondence table generation unit 118, the correspondence table that associates _{a i} and _{r i} and _{h m (i) [r i} , h m (i), a i] to generate and output. Figure 5C, I = 8, _{r i} is the S / N _ratio, the combination of _{h m (i)} is the model parameters β and gamma, the correspondence table when _{a i} is a speech recognition accuracy _{[r i, h m (i ),} it is _{a i]} example (step S217). Correspondence table _{[r i, h m (i} ), a i] is stored in the correspondence table storage unit 221a of the voice recognition device 22 (FIG. 2).

<Voice recognition processing>
The only difference from the first embodiment is that the process of step S224 is performed instead of step S124 of FIG. In step S224, the selection unit 225 refers to the correspondence table [r _i , hm _(i) , a _i ] stored in the correspondence table storage unit 221a, and the “second index” u is the same as or near the “second index”. Among the model parameters or model parameter combinations hm _(i) corresponding to “1 index” r _i , the one that maximizes the speech recognition accuracy or speech recognition rate a _i is selected. For example, in the example of FIG. 5B and FIG. 5C, when r _i near u is r ₂ , r ₂ , r ₃ , a ₂ , a ₂ , a ₃ corresponding to r ₂ , r ₂ , r ₃ , respectively. _{Hm (i)} corresponding to the largest a _i is selected. The selected _{hm (i)} is sent to the speech recognition unit 126. When a plurality of a _i are the same as each other, _{hm (i)} corresponding to any of them may be selected. For example, if a plurality of a _i are identical to each other, of those a _i, h _{m (i)} is may be selected associated with the corresponding a _i closest r _i in u (step S224).

[Modification of Second Embodiment]
Of the model parameters or combinations of model parameters corresponding to the “first index” that is the same as or close to the “second index”, the one that maximizes the weight value of the speech recognition accuracy or the speech recognition rate may be selected. However, the “weighting value” is a value obtained by multiplying the voice recognition accuracy or the voice recognition rate by a larger weight as the distance between the “second index” and the “first index” is smaller.

<Configuration>
As illustrated in FIG. 2, the speech recognition device 22 ′ of the present modification includes a correspondence table storage unit 221 a, an input acoustic signal storage unit 121 b, a noise component storage unit 121 c, an input unit 122, and a speech / non-speech section determination unit 123. , An index generation unit 124, a selection unit 225 ′, and a voice recognition unit 126.

<Learning process>
The same as in the second embodiment.

<Voice recognition processing>
The difference from the second embodiment is only that step S224 ′ is performed instead of step S224 in FIG. ', The selector 225' step S224 is, the correspondence table _{[r i, h m (i} ), a i] stored in the correspondence table storage unit 221a refers to the "second index" u identical or near Of the model parameters or model parameter combinations hm _(i) corresponding to the “first index” r _i , the one that maximizes the weighting value obtained by multiplying the speech recognition accuracy or speech recognition rate a _i by the weight c _i is selected. To do. However, c _i is a positive value, and is larger as the distance | u−r _i | between u and r _i is smaller. For example, in the example of FIGS. 5B and 5C, when r _i near u is r ₂ , r ₂ , r ₃ , c ₁ , a ₂ , c ₂ corresponding to r ₂ , r ₂ , r ₃ , respectively. · _{aa 2,} of c 3 · aa ₃ selects _{h m (i)} corresponding to the maximum of _{c i} · _a i. In this example, c ₂ > c ₃ > c ₁ is satisfied. The selected _{hm (i)} is sent to the speech recognition unit 126 (step S224 ′).

[Third Embodiment]
Next, a third embodiment will be described. The third embodiment is a modification of the first embodiment. In this embodiment, the model parameter or the complement value of the model parameter or the combination of the model parameter or the complement value of the model parameter corresponding to the “first index” or the complement value of the “first index” that matches the “second index” is selected. To do.

<Configuration>
As illustrated in FIG. 1, the learning device 31 of the present embodiment includes a speech signal storage unit 111a, a noise signal storage unit 111b, a correct word string storage unit 111c, a noise-added speech signal storage unit 111e, a speech recognition result storage unit 111f, A component adjustment adding unit 113, an index generating unit 114, a voice recognition unit 116, a comparing unit 117, and a correspondence table generating unit 318 are included. As illustrated in FIG. 2, the speech recognition device 32 according to the present modification includes a correspondence table storage unit 221a, an input acoustic signal storage unit 121b, a noise component storage unit 121c, an input unit 122, a speech / non-speech segment determination unit 123, An index generation unit 124, a selection unit 325, and a voice recognition unit 126 are included.

<Learning process>
The only difference from the first embodiment is that the process of step S317 is performed instead of step S117 of FIG. In step S317, the correspondence table generation unit 318 sets m (i) ε {0,..., M−1} (where i = 0,...) That maximizes the speech recognition accuracy or the speech recognition rate for each i. ., I-1) is selected. Alternatively, the correspondence table generation unit 318 may select one m (i) ε {0,..., M−1} for which the voice recognition accuracy or the voice recognition rate is greater than or equal to a predetermined value for each i. . Further correspondence table generation unit _318, r 0, · · _{·, r} a _I-1 supplemented by linear interpolation or the _like, r 0, · · _{·, r I-1} and the continuous value r consisting of complementary value _'0 ,..., R ′ _Z−1 (where Z is an integer greater than I). Also, the correspondence table generation unit 318 complements _{hm (0)} ,..., Hm _(I-1) by linear interpolation or the like, and hm ₍₀₎ ,..., Hm _(I-1). And continuous values h ′ _{m (0)} ,..., H ′ _{m (Z−1)} composed of their complementary values. The correspondence table generation unit 318 associates r ′ _z with h ′ _{m (z)} [r ′ _z , h ′ _{m (z)} ] (where z = 0,..., Z−1). ) Is generated and output. The correspondence table [r ′ _z , h ′ _{m (z)} ] is stored in the correspondence table storage unit 321a of the speech recognition device 32 (FIG. 2).

<Voice recognition processing>
The only difference from the first embodiment is that the process of step S324 is performed instead of step S124 of FIG. In step S324, the selection unit 325 refers to the correspondence table [r ′ _z , h ′ _{m (z)} ] stored in the correspondence table storage unit 121a, and is associated with r ′ _z that matches the input u. A model parameter or a combination of model parameters or a complementary value h ′ _{m (z)} thereof is selected. The selected h ′ _{m (z)} is sent to the speech recognition unit 126 (step S324). The subsequent processing is the same as that of the first embodiment except that h ′ _{m (z)} is used instead of h _{m (i)} .

[Modification of Third Embodiment]
Correspondence table generated by the learning process of the first embodiment _{[r i, h m (i} )] was used, the correspondence table during the speech recognition process _{[r i, h m (i} )] correspondence table complements [r ' _z , h ′ _{m (z)} ] may be generated, and the process of step S324 may be executed.

[Other variations]
The present invention is not limited to the embodiment described above. For example, instead of each device exchanging information via a network, at least some of the devices may exchange information via a portable recording medium. Alternatively, at least some of the devices may exchange information via a non-portable recording medium. That is, the combination which consists of a part of these apparatuses may be the same apparatus.

  The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  When the above configuration is realized by a computer, the processing contents of the functions that each device should have are described by a program. By executing this program on a computer, the above processing functions are realized on the computer. The program describing the processing contents can be recorded on a computer-readable recording medium. An example of a computer-readable recording medium is a non-transitory recording medium. Examples of such a recording medium are a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, and the like.

  This program is distributed, for example, by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, this computer reads a program stored in its own recording device and executes a process according to the read program. As another execution form of the program, the computer may read the program directly from the portable recording medium and execute processing according to the program, and each time the program is transferred from the server computer to the computer. The processing according to the received program may be executed sequentially. The above-described processing may be executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. Good.

  In the above embodiment, the processing functions of the apparatus are realized by executing a predetermined program on a computer. However, at least a part of these processing functions may be realized by hardware.

Learning device 11, 21, 31
Voice recognition device 12, 22, 22 ', 32

Claims (5)

A model parameter of the speech recognition model or a combination of model parameters of the speech recognition model is selected according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, and the selected model parameter or the combination of model parameters is selected. A speech recognition apparatus that applies the speech recognition model according to the method to the input acoustic signal ,
A voice having a maximum voice recognition rate or a voice recognition rate equal to or greater than a predetermined value when applied to an acoustic signal having a relationship represented by the first index and the first index representing the magnitude relationship between a voice component and a noise component The model parameter or combination of model parameters of the recognition model is associated with
Among the model parameters or combinations of model parameters corresponding to the first index that is the same as or close to the second index that represents the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, the speech recognition accuracy Or a speech recognition device that selects the one that maximizes the speech recognition rate. A model parameter of the speech recognition model or a combination of model parameters of the speech recognition model is selected according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, and the selected model parameter or the combination of model parameters is selected. A speech recognition apparatus that applies the speech recognition model according to the method to the input acoustic signal,
A voice having a maximum voice recognition rate or a voice recognition rate equal to or greater than a predetermined value when applied to an acoustic signal having a relationship represented by the first index and the first index representing the magnitude relationship between a voice component and a noise component The model parameter or combination of model parameters of the recognition model is associated with
Among the model parameters or combinations of model parameters corresponding to the first index that is the same as or close to the second index that represents the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, the speech recognition accuracy Or select the one that maximizes the weight of the speech recognition rate,
The weight recognition value is a voice recognition device, which is a value obtained by multiplying the voice recognition accuracy or the voice recognition rate by a larger weight as the distance between the second index and the first index is smaller. A model parameter of the speech recognition model or a combination of model parameters of the speech recognition model is selected according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, and the selected model parameter or the combination of model parameters is selected. A speech recognition method that applies the speech recognition model according to the method to the input acoustic signal ,
A voice having a maximum voice recognition rate or a voice recognition rate equal to or greater than a predetermined value when applied to an acoustic signal having a relationship represented by the first index and the first index representing the magnitude relationship between a voice component and a noise component The model parameter or combination of model parameters of the recognition model is associated with
Among the model parameters or combinations of model parameters corresponding to the first index that is the same as or close to the second index that represents the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, the speech recognition accuracy Or a speech recognition method that selects the one that maximizes the speech recognition rate.   A model parameter of the speech recognition model or a combination of model parameters of the speech recognition model is selected according to the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, and the selected model parameter or the combination of model parameters is selected. A speech recognition method that applies the speech recognition model according to the method to the input acoustic signal,
  A voice having a maximum voice recognition rate or a voice recognition rate equal to or greater than a predetermined value when applied to an acoustic signal having a relationship represented by the first index and the first index representing the magnitude relationship between a voice component and a noise component The model parameter or combination of model parameters of the recognition model is associated with
  Among the model parameters or combinations of model parameters corresponding to the first index that is the same as or close to the second index that represents the relationship between the magnitudes of the speech component and the noise component included in the input acoustic signal, the speech recognition accuracy Or select the one that maximizes the weight of the speech recognition rate,
  The speech recognition method, wherein the weighting value is a value obtained by multiplying the speech recognition accuracy or speech recognition rate by a greater weight as the distance between the second index and the first index is smaller. Program for causing a computer to function as a speech recognition apparatus according to claim 1 or 2.

Images