JP5233330B2 - Acoustic analysis condition normalization system, acoustic analysis condition normalization method, and acoustic analysis condition normalization program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the following problems: recognition performance gets low when an acoustic analytical condition at the time of learning an acoustic model is different from an acoustic analytical condition executed for an input signal of a recognition object; and a high calculation cost and a large volume of data are required when changing the acoustic analytical condition. <P>SOLUTION: The acoustic analytical condition normalizing system or the like includes at least: an acoustic analytical means (I) for extracting a feature quantity from the input signal of the recognition object, according to the acoustic analytical condition (I); a sound model storage means for preparing a sound model (I), based on the feature quantity extracted according to the acoustic analytical condition (I) from the first learning input signal in the acoustic analytical means (I), and for storing it; a correlation table storage means for storing a correlation table between the sound model (O) prepared based on the feature quantity extracted from the second learning input signal according to the acoustic analytical condition (O) identical to that at the time of preparing the sound model, in the acoustic analytical means (O), and the sound model (I); and an acoustic analytical condition correcting means for correcting the feature quantity of the input signal of the recognition object, using the sound model (I) and the correlation table. <P>COPYRIGHT: (C)2009,JPO&amp;INPIT

Description

  The present invention corrects differences in acoustic analysis conditions in a speech recognition system.

  It is known that the speech recognition system degrades the recognition performance due to the influence of the transfer characteristics of a device that records speech such as background noise and microphone. In order to prevent recognition performance degradation due to the influence of such additive or multiplicative noise, noise suppression is generally performed in an acoustic analysis unit.

  FIG. 12 is a configuration diagram of a speech recognition system that performs general noise suppression in analogy with Patent Document 1. In FIG. In FIG. 12, first, in the input signal acquisition unit 101, a voice input signal recorded using a microphone or the like is converted into digital information, cut out in units of frames, and subjected to FFT (Fast Fourier Transform) or the like to obtain spectral information of the voice. get. When the utterance of the frequency f at time t is St, f, the additive noise is Nt, f, and the multiplying noise is Ht, f, the spectrum Xt, f of the input signal is expressed by the following equation.

  Next, the noise component calculation unit 102 estimates an additive noise component included in the input signal. As a method for calculating the noise component, a method in which a plurality of frames that do not include voice in the input signal are averaged is considered. Next, the estimated speech calculation unit 103 suppresses the noise component using a spectral subtraction method, a Wiener filter method, or the like, and calculates an estimated value of speech. An example of the spectral subtraction method is shown below. In the spectral subtraction method, the calculated noise component is subtracted from the input signal in the spectral domain to suppress the noise component and estimate clean speech.

Here, S ^SS t, f is an estimated value of clean speech calculated by the spectral subtraction method. <Nf> is a noise component calculated by the noise component calculation unit 102. max [] indicates that a larger value of two values separated by a comma is adopted. α is a parameter for controlling the strength and distortion of noise suppression. β is a parameter indicating flooring, and this parameter is also a parameter for controlling the strength and distortion of noise suppression. These parameters are known to have different optimum values depending on the type of noise and SNR. Patent Document 1 discloses a method for changing these parameters in accordance with the SNR. Next, the feature amount extraction unit 104 extracts a cepstrum, logarithmic power, or primary and secondary difference amounts thereof, or a combination thereof from the estimated speech. Here, in order to remove the multiplicative noise, a part related to the multiplicative noise is subtracted or normalized from the feature amount. A cepstrum average normalization method is widely known as a method for removing multiplicative noise. The configuration until the feature amount is extracted from the input signal described above is called an acoustic analysis unit 100, and a plurality of configurations other than the above are proposed. Next, the decoder unit 601 searches for an appropriate word string using the acoustic model 602 and the language model 603 for the feature amount. As an acoustic model, Hidden Markov Model (HMM) or the like is widely used. The acoustic model is created by learning feature quantities extracted in advance for the learning data under the same conditions as the acoustic analysis unit 100. Finally, the output unit 604 outputs the word string.

  FIG. 13 is a configuration diagram of a general distributed speech recognition system inferred from Patent Document 2. In such a distributed speech recognition system, first, an acoustic analysis is performed by the acoustic analysis unit 100 on the terminal side, and a feature amount is calculated. Next, the feature amount transmitting unit 701 transmits the calculated feature amount to the server side. Next, the feature amount receiving unit 702 on the server side receives the feature amount. Next, the decoder unit 601 searches for an appropriate word string using the acoustic model 602 and the language model 603 for the received feature amount. Next, the word string transmission unit 703 transmits the word string to the terminal side. Next, the word string receiving unit 704 on the terminal side receives the word string. Finally, the word string is output from the output unit 705 on the terminal side.

JP 2005-165221 A JP 2006-350090 A

  The disclosures of Patent Document 1-2 are incorporated herein by reference. The following analysis is given by the present invention.

  In the conventional method of performing noise suppression in the acoustic analysis unit shown in FIG. 12, if the acoustic analysis condition at the time of learning the acoustic model is different from the acoustic analysis condition performed on the input signal to be recognized, the recognition performance is reduced. There's a problem. For this reason, if an acoustic analysis is performed under an appropriate acoustic analysis condition in accordance with noise or SNR included in the input signal, it is necessary to relearn the acoustic model under the acoustic analysis condition or to adapt to the acoustic analysis condition. . However, since the size of the acoustic model is large, it requires a lot of calculation costs and a lot of data to learn or adapt it under new acoustic analysis conditions. The acoustic analysis conditions referred to here indicate processing, parameters, microphone characteristics, and the like that constitute the acoustic analysis unit 100.

  In the conventional distributed speech recognition system referred to FIG. 13, the above problem becomes more remarkable. In a normal speech recognition system, there is a one-to-one relationship between an acoustic analysis unit and a decoder unit, but in the distributed speech recognition system, a decoder unit of one server is assigned to the acoustic analysis units of a plurality of terminals. Can be considered. Each terminal is different in the noise environment to be used and has the optimum acoustic analysis conditions. In addition, characteristics such as a microphone vary depending on the terminal. For this reason, it is necessary to match the acoustic model side to the acoustic analysis conditions on the terminal side, but it is very difficult to match the conditions of all terminals.

  As described above, the conventional system has the following problems.

  The first problem is that in the conventional method of performing noise suppression by the acoustic analysis unit, if the acoustic analysis conditions at the time of learning the acoustic model and the acoustic analysis conditions performed on the input signal to be recognized are different, the recognition performance is low. That is.

  The second problem is that in the conventional method of performing noise suppression in the acoustic analysis unit, it is necessary to re-learn or adapt the acoustic model when the acoustic analysis condition to be performed on the recognition target input signal is changed. It means that a lot of calculation cost and a lot of data are needed.

  An object of the present invention is to normalize acoustic analysis conditions so that high-performance speech recognition can be performed even when acoustic analysis conditions at the time of learning an acoustic model differ from acoustic analysis conditions performed on an input signal to be recognized. An acoustic analysis condition normalization system and an acoustic analysis condition normalization program are provided.

  Another object of the present invention is to provide an acoustic analysis condition normalization system and an acoustic analysis condition normalization program that normalize acoustic analysis conditions with low calculation cost and small data.

  In order to solve the above problems, the invention disclosed in the present application is generally configured as follows.

  An acoustic analysis condition normalization system according to one aspect of the present invention includes an acoustic analysis means (I) for extracting a feature value from an input signal to be recognized under an acoustic analysis condition (I), and a first learning input. A speech model storage for storing the created speech model (I) by creating a speech model (I) from features extracted under acoustic analysis conditions (I) in the acoustic analysis means (I) for the signal And the feature quantity extracted under the same acoustic analysis condition (O) as the acoustic model used for speech recognition, unlike the acoustic analysis means (I) in the acoustic analysis means (O) for the input signal for the second learning A correspondence table storing means for storing a correspondence table storing a correspondence relationship between the speech model (O) created from the speech model (I), the speech model (I), and the correspondence table. Correct the feature value of the input signal to be recognized and obtain the corrected feature value. Provided acoustic analysis condition correcting means, at least.

  By adopting such a configuration, mismatch due to a difference in acoustic analysis conditions can be corrected, and speech recognition with high performance can be performed. Therefore, the first object can be achieved. Further, since the speech model can be sufficiently small as compared with the method of adapting the acoustic model to the acoustic analysis conditions, the correction can be performed with a small amount of calculation and a small amount of memory. Can be achieved.

  In a preferred embodiment of the present invention, a speech model storage means for storing a speech model (I) created in advance under the same acoustic analysis conditions as performed on an input signal, the speech model (I) and the acoustic model in advance. Correspondence table storage means for storing a correspondence table storing the correspondence relationship with the speech model (O) created under the acoustic analysis condition at the time of learning, and correcting the acoustic analysis condition using the speech model (I) and the correspondence table The sound analysis condition correction means is provided.

  Hereinafter, for the sake of easy understanding, (O) is used for components related to acoustic analysis conditions when learning acoustic models for speech recognition, and (I) is used for components related to acoustic analysis conditions during input signal analysis. It shall be represented with a symbol.

  Next, the form of the 1st Example for inventing is demonstrated.

  FIG. 1 is a diagram showing the configuration of the first exemplary embodiment of the present invention. Referring to FIG. 1, an acoustic analysis condition normalization system according to a first embodiment of the present invention includes an input signal acquisition unit 101, a noise component calculation unit 102, an estimated speech calculation unit 103, a feature amount extraction unit 104, and a speech model. A storage unit 105, a correspondence table storage unit 106, an acoustic analysis condition correction unit 107, an output unit 108, and the like are provided.

  The input signal acquisition unit 101 cuts out audio input signals recorded using a microphone or the like in units of frames, and performs FFT or the like to acquire audio spectrum information.

  The noise component calculation unit 102 calculates a noise component from the input signal.

  The estimated speech calculation unit 103 calculates a temporary estimated speech from the input signal and the noise component.

  The feature amount extraction unit 104 extracts feature amounts from the estimated speech.

  The speech model storage unit 105 stores a speech model (I) created in advance under the same acoustic analysis conditions as those performed on the input signal.

  The correspondence table storage unit 106 stores a correspondence table between the speech model (I) and the speech model (O) prepared under the same acoustic analysis conditions as when learning the acoustic model.

  The acoustic analysis condition correcting unit 107 corrects the feature amount extracted for the estimated speech using the speech model (I) and the correspondence table.

  The output unit 108 outputs the corrected feature amount.

  The input signal acquisition unit 101, the noise component calculation unit 102, the estimated speech calculation unit 103, and the feature quantity extraction unit 104 are collectively referred to as an acoustic analysis unit (I) 100.

  Of course, these units may implement their functions and processing by a program executed on a computer constituting the noise suppression system (the same applies to other embodiments).

  Next, details and a creation method of the speech model (I) and the correspondence table used in this embodiment will be described.

  Figure 2 shows a speech model (O) created under the same acoustic analysis conditions (O) as during acoustic model learning, and a speech model (I) created under the same acoustic analysis conditions (I) as performed on the input signal. ) And a correspondence table of elements constituting the two speech models. As a speech model, VQ (Vector Quantization) codebook, GMM (Gaussian Mixture Model), HMM (Hidden Markov Model), etc. can be considered. Here, an example in which the GMM is a voice model is shown. The GMM is a model represented by a probability density function configured by mixing a plurality of Gaussian distributions. Each Gaussian distribution corresponds to a unit divided into phonemes or speech clusters. The horizontal axis and the vertical axis in FIG. 2 indicate that two dimensions in the feature quantity have been selected. The ellipse in the figure represents a Gaussian distribution, and the center point represents the average value of the Gaussian distribution. An arrow connecting the Gaussian distributions indicates that the two Gaussian distributions are in a correspondence relationship.

  An example of a method for creating the two speech models and the correspondence table will be described with reference to the configuration diagram of FIG. 3 and the flowchart of FIG. Note that first learning data described later corresponds to a second learning input signal in claims, and second learning data corresponds to a first learning input signal in claims.

  First, in the acoustic analysis unit (O) 200, the first learning data is analyzed under the acoustic analysis condition (O), the feature amount (O) is calculated, and an EM (Expectation Maximization) algorithm or the like is used. A voice model (O) is created (step S101). This processing can be performed in advance when creating the acoustic model. In addition, the speech model (O) can be created using a part or all of the Gaussian distribution constituting the acoustic model. The created speech model is stored in the speech model (O) storage unit 201. Here, it is assumed that the speech model (O) is represented by a GMM configured by a mixture of K Gaussian distributions as shown in Equation 3 below.

Here, p (F ^O ) is a probability density function that outputs the feature quantity O. Further, a _i represents the mixing weight of the i-th Gaussian distribution, and N () represents a Gaussian distribution as shown in Equation 4.

  Here, μ represents an average value, Σ represents variance, and D represents a dimension of the feature amount.

  Next, the acoustic analysis unit (O) 200 analyzes the second learning data under the acoustic analysis condition (O), and calculates the feature amount (O) (step S102). In parallel with this, the second learning data is analyzed by the acoustic analysis unit (I) 100 under the acoustic analysis condition (I), and the feature quantity (I) is calculated (step S103). The second learning data may be the same acoustic environment as the first learning data or may be a different acoustic environment, but the same acoustic as the input signal to be recognized. Sound recorded in the environment is desirable. The acoustic environment here refers to the difference in background noise, speaker, or a combination of both. As the second learning data, the same input signal may be input as shown in FIG. 1, or two input signals of different acoustic environments may be input as shown in FIG. It is necessary to obtain information as to whether the two input signals having different acoustic environments are synchronized in time series or which frames in the time series correspond to each other such as phoneme labels.

  Next, the speech model adaptation unit 202 creates a speech model (I) after adaptation from the speech model (O) using the feature amount (O) and the feature amount (I) (step S104). . Specifically, the time series of the feature amount (O) is classified to which of the Gaussian distributions included in the speech model (O) is close, and the feature amount corresponding to the classified frame ( It is conceivable that a Gaussian distribution is newly constructed from the appearance probability of I) to obtain a speech model (I). In addition, a general adaptive algorithm such as a method using an MLLR (Maximum Likelihood Linear Regression) method or an EM algorithm can be used. The created speech model is stored in the speech model (I) storage unit 105. Equation 5 is a GMM representing the speech model (I).

  Next, the correspondence table creation unit 203 creates a correspondence between the Gaussian distribution included in the speech model (O) and the Gaussian distribution included in the speech model (I) (step S105). Specifically, an index is assigned to the Gaussian distribution included in the speech model (I), and an average value of the Gaussian distribution included in the speech model (O) corresponding to the index is held. The created correspondence table is stored in the correspondence table storage unit 106. Equation 6 shows the correspondence table.

  The configuration from step S102 to S104 is used when the acoustic analysis condition (I) can be set in advance, such as when the noise of the input signal can be predicted to some extent, and when the transfer characteristics of the microphone for acquiring the input signal are known. , Can be done in advance. Alternatively, when the noise of the input signal changes and the acoustic analysis condition (I) needs to be changed accordingly, the first few frames or several utterances of the input signal are used as adaptation data in step S102. To S105, and it can be adapted online. Also, here, the speech model (O) is created first under the same acoustic analysis conditions (O) as the acoustic model, and the speech model (I) is created under the same acoustic analysis conditions (I) as performed for the input signal. However, it is also possible to create a speech model (I) first and associate it with the speech model (O).

  Next, the operation of this embodiment will be described with reference to the flowcharts of FIGS.

  First, the acoustic analysis unit (I) 100 extracts the feature quantity (I) from the input signal (step S201). Thereafter, a cepstrum average normalization method or the like may be performed to normalize the multiplicative noise component.

  Next, the acoustic analysis condition correction unit 107 uses the speech model (I) stored in the speech model storage unit 105 for the feature amount (I) and the correspondence table stored in the correspondence table storage unit 106. Then, correction is made so as to match the acoustic analysis condition (O) in which the acoustic model is created (step S202). Specifically, the following calculation is performed.

Here, FI ^{→ 0} indicates the corrected feature value. table [i] is an audio table represented by Expression 6. Ppost (i | F ^I ) is a posterior probability with respect to the feature amount F _t ^I before correction at time t shown in the following Expression 8.

Here, the weight a _i , average vector μ _i ^I , and variance Σ _i ^I of the i-th Gaussian distribution are values held in the speech model (I) shown in Equation 5.

  Finally, the corrected feature value is output (step S203). Further, speech recognition may be performed using the feature amount.

  Next, the effect of the present embodiment will be described.

  The correspondence between the speech model (I) created under the acoustic analysis condition (I) for the input signal to be recognized and the speech model (O) created under the same acoustic analysis condition (O) as during acoustic model learning By correcting the acoustic analysis conditions using the held correspondence table, it is possible to correct mismatches due to differences in acoustic analysis conditions. Further, since a sufficiently small speech model can be used as compared with a method of adapting an acoustic model to acoustic analysis conditions, correction can be performed with a small calculation amount and a small increase in memory amount.

  FIG. 7 is a diagram showing the configuration of the second embodiment of the present invention.

  The second example is provided with a voice model storage unit 301 and a voice model selection unit 302 in contrast to the configuration of the first example.

  The speech model storage unit 301 stores a plurality of speech models corresponding to a plurality of acoustic analysis conditions.

The speech model selection unit 302 selects a speech model from the plurality of speech models from the feature amounts extracted by the feature amount extraction unit 104.
It has.

  Next, the operation of this embodiment will be described.

  In this embodiment, after the feature amount is calculated in step S201 of the first embodiment, the speech model selection unit 302 uses the calculated feature amount to satisfy a plurality of acoustic analysis conditions prepared in advance. A speech model is selected from a plurality of corresponding speech models (i, ii,...), And acoustic analysis conditions are corrected using the selected speech model in step S202 and subsequent steps. As a criterion for selection, a method of selecting a method with the highest likelihood using the likelihood of each speech model with respect to the feature amount may be considered.

  Next, the effect of the present embodiment will be described.

  By preparing speech models corresponding to a plurality of acoustic analysis conditions in advance and correcting the acoustic analysis conditions, it is possible to cope with a case where the acoustic analysis conditions change due to a change in noise conditions or the like.

  Further, in the present embodiment, in addition to using the feature amount extracted by the feature amount extraction unit 104 as a reference for selecting a speech model, the input signal acquired by the input signal acquisition unit 101 and the noise component calculation unit 102 are used. A method of selecting a speech model using the noise component calculated in step S1 or the estimated speech calculated in the estimated speech calculation unit 103 is also conceivable.

  FIG. 8 is a diagram showing the configuration of the third embodiment of the present invention.

  The third example includes an SNR calculation unit 401, an estimated speech calculation unit 402, and a speech model selection unit 403, as compared to the configuration of the second example.

  The SNR calculation unit 401 calculates a ratio between the input signal acquired by the input signal acquisition unit 101 and the noise component calculated by the noise component calculation unit 102, and calculates a signal-to-noise ratio (SNR).

  In place of the estimated speech calculation unit 103, the estimated speech calculation unit 402 converts the estimated speech from the input signal acquired by the input signal acquisition unit 101, the noise component calculated by the noise component calculation unit 102, and the calculated SNR. calculate.

  The voice model selection unit 403 selects the voice model from the calculated SNR instead of the voice model selection unit 302.

  By adopting such a configuration, it is possible to change the parameters related to the noise suppression processing method and the noise suppression processing in the estimated speech calculation unit according to the SNR, and to cope with the acoustic analysis conditions thus changed. A voice model can be selected and corrected.

  FIG. 9 is a diagram showing the configuration of the fourth embodiment of the present invention.

  The fourth example includes a second estimated speech calculation unit 501, a second feature amount extraction unit 502, and an output unit 503 with respect to the configuration of the first example.

  The second estimated speech calculation unit 501 includes the estimated speech corrected by the acoustic analysis condition correction unit 107, the input signal acquired by the input signal acquisition unit 101, and the noise component calculated by the noise component calculation unit 102. The estimated speech is calculated again.

  The second feature amount extraction unit 502 extracts feature amounts from the second estimated speech.

  The output unit 503 outputs the second feature amount.

  Next, the operation of this embodiment will be described.

In this embodiment, after performing step S202 of the first embodiment, the feature amount extracted from the input signal to be recognized is corrected to the acoustic analysis condition (O) at the time of learning the acoustic model, and then the second estimated speech. The calculation unit 501 calculates a filter for noise suppression using the corrected feature value and the noise component, and applies it to the input signal. For specific processing, a method using a Wiener filter will be described here. Subsequent processing can be performed independently for each frequency band. The corrected feature value F _t ^{I → 0} obtained using Equation 7 needs to be inversely transformed back to the spectrum amount S _t ^{I → 0} . When a cepstrum is used as a feature quantity, inverse transformation as shown in the following Expression 9 is performed.

  Here, IDCT () indicates inverse discrete cosine transform. Alternatively, it is also conceivable to store an average spectrum corresponding to each Gaussian distribution of the speech model (O) as shown in the following Expression 10 instead of Expression 6 as a correspondence table.

  In this case, Equation 9 is changed to Equation 11 below.

  In addition, a method of using a logarithmic spectrum instead of storing the spectrum in the correspondence table as in the above example can be naturally considered.

  The filter gain of the Wiener filter is reduced by using the corrected characteristic amount obtained by the above procedure into the amount of the spectral region and the amount of the spectral region of the noise component calculated by the noise component calculation unit 102. The calculation is performed according to Equation 12 below.

  A second estimated speech is obtained by multiplying the input signal by the filter gain.

  Next, the second feature quantity extraction unit 502 extracts a feature quantity from the second estimated speech. The same feature quantity as the feature quantity extraction unit 104 or a different one may be used as the feature quantity, but it is necessary to match it with the acoustic model learning.

  Finally, as in step S203 of the first embodiment, the output unit 503 outputs the feature amount extracted by the second feature amount extraction unit. Further, speech recognition may be performed using the feature amount.

  Next, the effect of the present embodiment will be described.

  The feature amount corrected up to step S202 of the first embodiment is estimated as an excessively smoothed value, particularly when the number of Gaussian distributions constituting the speech model (I) is small. By creating a noise suppression filter using the corrected estimated speech and the noise component and applying it to the original input signal as shown in Equation 13, the information of the speech possessed by the original input signal is obtained. Only the difference due to the difference in the noise suppression method can be corrected without loss.

  Further, in the present embodiment, the second feature quantity extraction unit 502 uses a feature quantity different from that of the feature quantity extraction unit 104 to reduce the difference in noise suppression method for additive noise and multiplicative noise included in the input signal. In addition to correction, differences in feature values can also be corrected.

  The first to fourth embodiments may be used together with a speech recognition system as shown in FIG.

  This embodiment is provided with a decoder section 601 and an output section 604 in contrast to the configurations of the first to fourth embodiments.

  The decoder unit 601 searches the word string using the acoustic model 602 and the language model 603 for the feature amount extracted by the feature amount extraction unit 104 or the second feature amount extraction unit 502.

  The output unit 604 outputs the searched word string.

  By adopting such a configuration, the acoustic analysis conditions performed on the input signal can be matched with the acoustic analysis conditions at the time of learning the acoustic model used in the decoder unit, and speech recognition with high recognition performance can be performed. . In the example of FIG. 10, an example in which there is one acoustic model is shown, but it is also possible to have a plurality of acoustic models.

  The first to fourth embodiments may be used together with distributed speech recognition as shown in FIG.

  The present embodiment is different from the first to fourth embodiments in the feature amount transmitting unit 701, the feature amount receiving unit 702, the decoder unit 601, the word string transmitting unit 703, the word string receiving unit 704, and the output unit 705. It has.

  The feature amount transmission unit 701 transmits the feature amount extracted by the feature amount extraction unit 104 or the second feature amount extraction unit 502 to the server side.

  The feature amount receiving unit 702 receives the transmitted feature amount on the server side.

  The decoder unit 601 searches for a word string using the acoustic model 602 and the language model 603 for the received feature amount.

  The word string transmission unit 703 transmits the searched word string to the terminal side.

  The word string receiving unit 704 receives the transmitted word string on the terminal side.

  The output unit 705 outputs the received word string.

  By adopting such a configuration, the acoustic model is corrected so that the acoustic analysis environment on the terminal side is the same as when learning the acoustic model, so the acoustic model held on the server side is created only under a certain acoustic analysis condition. You should do it. Of course, it is also possible to prepare a plurality of acoustic models. Further, in the example shown in FIG. 11, the acoustic analysis correction unit is arranged on the terminal side, but it is of course possible to arrange it on the server side.

  In the distributed speech recognition system, it is conceivable that the decoder unit of one server is assigned to the acoustic analysis units of a plurality of terminals. By correcting the feature value using the configuration of the present embodiment, even when the acoustic analysis conditions are different on the plurality of terminals, the decoder of one server can cope with the plurality of terminals.

  The present invention can be applied to uses such as speech recognition under noise.

It is a block diagram which shows the structure of the pressure processing system for noise concerning the 1st Example of this invention. It is a conceptual diagram which shows the relationship between the audio | voice model concerning a 1st Example of this invention, and a corresponding | compatible table. It is a block diagram which shows the structure of the speech model preparation method concerning 1st Example of this invention. It is a block diagram which shows another structure of the audio | voice model creation method concerning the 1st Example of this invention. It is a flowchart which shows the procedure of the audio | voice model creation method concerning the 1st Example of this invention. It is a flowchart which shows the procedure of the noise suppression processing system concerning 1st Example of this invention. It is a block diagram which shows the structure of the noise suppression system concerning the 2nd Example of this invention. It is a block diagram which shows the structure of the noise suppression system concerning the 3rd Example of this invention. It is a block diagram which shows the structure of the noise suppression system concerning the 4th Example of this invention. It is a block diagram which shows the structure of the speech recognition system concerning the 5th Example of this invention. It is a block diagram which shows the structure of the distributed speech recognition system concerning the 6th Example of this invention. It is a block diagram which shows the structure of the speech recognition system which performs noise suppression processing in the conventional speech analysis part. It is a block diagram which shows the structure of the conventional distributed speech recognition system.

Explanation of symbols

100 Acoustic Analysis Department (I)
DESCRIPTION OF SYMBOLS 101 Input signal acquisition part 102 Noise component calculation part 103 Estimated speech calculation part 104 Feature-value extraction part 105 Speech model storage part (speech model (I) storage part)
106 Corresponding table storage unit 107 Acoustic analysis condition correction unit 108 Output unit 200 Acoustic analysis unit (O)
201 Speech model (O) storage unit 202 Speech model adaptation unit (O → I)
DESCRIPTION OF SYMBOLS 203 Correspondence table preparation part 301 Several speech model storage part 302 Speech model selection part 401 SNR calculation part 402 Estimated speech calculation part 403 Speech model selection part 501 2nd estimation speech calculation part 502 2nd feature-value extraction part 503 Output part 601 Decoder unit 602 Acoustic model 603 Language model 604 Output unit 701 Feature amount transmission unit 702 Feature amount reception unit 703 Word string transmission unit 704 Word string reception unit 705 Output unit

Claims (25)

An acoustic analysis condition normalization system for speech recognition,
Acoustic analysis means (I) for extracting feature values from the input signal to be recognized under acoustic analysis conditions (I);
A speech model (I) configured by mixing a plurality of Gaussian distributions of feature quantities extracted under acoustic analysis conditions (I) in the acoustic analysis means (I) with respect to the first learning input signal is created. Voice model storage means for storing the created voice model (I);
Different from the acoustic analysis means (I) in the acoustic analysis means (O) for the second learning input signal, a plurality of feature quantities extracted under the same acoustic analysis conditions (O) as in the creation of the acoustic model used for speech recognition Correspondence table storage means for storing a correspondence table storing correspondence relationships between each Gaussian distribution included in the speech model (O) configured by mixing Gaussian distributions and each Gaussian distribution included in the speech model (I). When,
For each Gaussian distribution constituting the speech model (I), calculate a posterior probability for the feature quantity of the input signal to be recognized,
By correcting the average value of the corresponding Gaussian distribution using the correspondence table by weighted averaging using the posterior probability, the feature quantity of the input signal is corrected to match the acoustic analysis condition (O), An acoustic analysis condition normalizing system comprising: an acoustic analysis condition correcting unit for obtaining a corrected feature quantity. The correspondence table stores a set of an index of each Gaussian distribution constituting the speech model (I) and an average value of each Gaussian distribution constituting the speech model (O) corresponding to the index. The acoustic analysis condition normalization system described in 1. 3. The correspondence relationship between the respective Gaussian distributions is set by using the same learning data when the speech model (I) and the speech model (O) are created, and stored in the correspondence table. The acoustic analysis condition normalization system according to any one of the above. At the time of creating the speech model (I) and the speech model (O), by creating each Gaussian distribution in units of phonemes, the correspondence relationship between the respective Gaussian distributions is set and stored in the correspondence table. acoustic analysis conditions normalization system according to any one of claims 1 to 2. Characteristics of a microphone that acquires an input signal, a noise estimation method that estimates a noise component from an input signal, a noise suppression method that suppresses a noise component of an input signal to obtain an estimated speech, a feature amount calculation method, or one or more of these The acoustic analysis condition is a combination of
The acoustic analysis condition normalization system according to any one of claims 1 to 4 . The speech model storage means stores a plurality of speech models (i, ii,...) Created under a plurality of different acoustic analysis conditions (i, ii,...)
A speech model selection means for selecting a speech model (I) using the input signal, the noise component, the estimated speech, the feature amount, or a combination thereof;
The acoustic analysis condition correcting unit corrects the feature quantity of the input signal to be recognized using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalization system according to claim 5 . The speech model storage means stores a plurality of speech models (i, ii,...) Created under a plurality of different acoustic analysis conditions (i, ii,...)
SNR estimation means for estimating SNR from the input signal;
Voice model selection means for selecting a voice model (I) from the estimated SNR,
The acoustic analysis condition correcting unit corrects the feature quantity of the input signal to be recognized using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalization system according to any one of claims 1 to 5 . A distributed speech recognition system that extracts feature values from input speech to be recognized in an acoustic analysis unit on a terminal side, and converts the feature values into a word string using the acoustic model in a decoder unit on a server side,
The feature amount extracted under the acoustic analysis condition (I) in the acoustic analysis unit on the terminal side is corrected using the means according to any one of claims 1 to 7 on the terminal side or the server side. A featured distributed speech recognition system. A terminal of a distributed speech recognition system that extracts a feature amount from input speech to be recognized in an acoustic analysis unit of a terminal, and converts the feature amount into a word string using the acoustic model in a server-side decoder unit,
A terminal of a distributed speech recognition system, wherein the feature amount extracted under the acoustic analysis condition (I) is corrected using the means according to any one of claims 1 to 7 . An acoustic analysis condition normalization method for speech recognition,
A first step of extracting a feature value with respect to an input signal to be recognized under an acoustic analysis condition (I);
A speech model (I) configured by mixing a plurality of Gaussian distributions of feature quantities extracted under acoustic analysis conditions (I) in the acoustic analysis means (I) with respect to the first learning input signal is created. A second step of storing the created speech model (I);
Different from the acoustic analysis means (I) in the acoustic analysis means (O) for the second learning input signal, a plurality of feature quantities extracted under the same acoustic analysis conditions (O) as in the creation of the acoustic model used for speech recognition A third step of storing a correspondence table storing correspondence relationships between each Gaussian distribution included in the speech model (O) configured by mixing Gaussian distributions and each Gaussian distribution included in the speech model (I); ,
For each Gaussian distribution constituting the speech model (I), calculate a posterior probability for the feature quantity of the input signal to be recognized,
By averaging the corresponding Gaussian distribution average value using the correspondence table using the posterior probability,
And correcting the feature quantity of the input signal so as to match the acoustic analysis condition (O), and obtaining a corrected feature quantity. The correspondence table includes an index of each Gaussian distribution constituting the speech model (I), and
The acoustic analysis condition normalization method according to claim 10 , wherein a pair of the index and the average value of each Gaussian distribution constituting the speech model (O) corresponding to the index is stored. Wherein when creating a speech model (I) and said speech model (O), by using the same training data, sets the corresponding relationship between the Gaussian distributions of the respective claims 10 to 11, stored in the correspondence table The acoustic analysis condition normalization method according to any one of the above. At the time of creating the speech model (I) and the speech model (O), by creating each Gaussian distribution in units of phonemes, the correspondence relationship between the respective Gaussian distributions is set and stored in the correspondence table. The acoustic analysis condition normalizing method according to claim 10 . Characteristics of a microphone that acquires an input signal, a noise estimation method that estimates a noise component from an input signal, a noise suppression method that suppresses a noise component of an input signal to obtain an estimated speech, a feature amount calculation method, or one or more of these The acoustic analysis condition is a combination of
The acoustic analysis condition normalizing method according to claim 10 . In the second step, a plurality of speech models (i, ii, ...) created under a plurality of different acoustic analysis conditions (i, ii, ...) are stored,
The input signal, the noise component, the estimated speech, or the feature amount,
Alternatively, the method includes a step of selecting the speech model (I) using a combination of these,
In the fourth step, the feature amount of the input signal to be recognized is corrected using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalization method according to claim 14 . In the second step, a plurality of speech models (i, ii, ...) created under a plurality of different acoustic analysis conditions (i, ii, ...) are stored,
Estimating an SNR from the input signal;
Selecting a speech model (I) from the estimated SNR,
In the fourth step, the feature amount of the input signal to be recognized is corrected using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalizing method according to claim 10 . When the feature value is extracted from the input speech to be recognized in the acoustic analysis unit on the terminal side, and the feature value is converted into a word string using the acoustic model in the decoder unit on the server side,
The feature amount extracted under the acoustic analysis condition (I) in the acoustic analysis unit on the terminal side is corrected using the method according to any one of claims 10 to 16 on the terminal side or the server side. A characteristic acoustic analysis condition normalization method. When the feature amount is extracted from the input speech to be recognized in the acoustic analysis unit of the terminal, and the feature amount is converted into a word string using the acoustic model in the server-side decoder unit,
A terminal of a distributed speech recognition system, wherein the feature amount extracted under the acoustic analysis condition (I) is corrected using the method according to any one of claims 10 to 16 . An acoustic analysis condition normalization program that operates in an acoustic analysis condition normalization system for speech recognition,
Computer
Acoustic analysis means (I) for extracting feature values from the input signal to be recognized under the acoustic analysis condition (I), and acoustic analysis condition (I) in the acoustic analysis means (I) for the first learning input signal A speech model storage means for creating a speech model (I) configured by mixing a plurality of Gaussian distributions of the feature quantities extracted in step 1 and storing the created speech model (I);
Different from the acoustic analysis means (I) in the acoustic analysis means (O) for the second learning input signal, a plurality of feature quantities extracted under the same acoustic analysis conditions (O) as in the creation of the acoustic model used for speech recognition Correspondence table storage means for storing a correspondence table storing correspondence relationships between each Gaussian distribution included in the speech model (O) configured by mixing Gaussian distributions and each Gaussian distribution included in the speech model (I). ,
For each Gaussian distribution constituting the speech model (I), calculate a posterior probability for the feature quantity of the input signal to be recognized,
By correcting the average value of the corresponding Gaussian distribution using the correspondence table by weighted averaging using the posterior probability, the feature quantity of the input signal is corrected to match the acoustic analysis condition (O), Acoustic analysis condition normalization program that functions as acoustic analysis condition correction means for obtaining the corrected feature value The correspondence table includes a Gaussian distribution index constituting the speech model (I), and
The acoustic analysis condition normalization program according to claim 19 , wherein a pair of the index and an average value of a Gaussian distribution constituting the speech model (O) corresponding to the index is stored. Wherein when creating a speech model (I) and the voice model (O), the same training data by using, sets the corresponding relationship between the Gaussian distribution of the respective of claims 19 to 20 stored in the correspondence table The acoustic analysis condition normalization program according to any one of the above. At the time of creating the speech model (I) and the speech model (O), by creating each Gaussian distribution in units of phonemes, the correspondence relationship between the respective Gaussian distributions is set and stored in the correspondence table. The acoustic analysis condition normalization program according to any one of claims 19 to 20 . Characteristics of a microphone that acquires an input signal, a noise estimation method that estimates a noise component from an input signal, a noise suppression method that suppresses a noise component of an input signal to obtain an estimated speech, a feature amount calculation method, or one or more of these The acoustic analysis condition is a combination of
The acoustic analysis condition normalization program according to any one of claims 19 to 22 . The speech model storage means stores a plurality of speech models (i, ii,...) Created under a plurality of different acoustic analysis conditions (i, ii,...)
Making a computer function as a speech model selection means for selecting a speech model (I) using the input signal, the noise component, the estimated speech, the feature amount, or a combination of these,
The acoustic analysis condition correcting unit corrects the feature quantity of the input signal to be recognized using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalization program according to claim 23 . The speech model storage means stores a plurality of speech models (i, ii,...) Created under a plurality of different acoustic analysis conditions (i, ii,...)
Computer
SNR estimation means for estimating SNR from the input signal;
Function as speech model selection means for selecting a speech model (I) from the estimated SNR;
The acoustic analysis condition correcting unit corrects the feature quantity of the input signal to be recognized using the selected speech model (I) and the correspondence table.
The acoustic analysis condition normalization program according to any one of claims 19 to 23 .

Images