JP2009020460A - Voice processing device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To select a section in which a target sound exists and another section in which voice except for the target sound exists out of voice signals. <P>SOLUTION: A voice division part 12 divides the voice signals S into a pronouncing section PA and a non-pronouncing section PB on a time axis. A storing device 20 stores a general acoustic model of the target sound. A selection processing part 13 judges the existence of correlation of the acoustic model and the feature amount of the voice signal S in each pronouncing section PA, and selects the pronouncing section PA having correlation with the acoustic model into an effective section PA1 and selects the pronouncing section PA having no correlation with the acoustic model into a rejection section PA2. A voice classifying part 14 classifies the pronouncing section PA selected into the effective section PA1 out of a plurality of the pronouncing sections PA demarcated by the voice division part 12 in every speaker based on the characteristic amount of the voice signal S in the pronouncing section PA. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a technique for dividing an audio signal into a plurality of sections on a time axis.

Various techniques for dividing an audio signal into a plurality of sections along the time axis have been proposed. For example, Patent Literature 1 and Patent Literature 2 disclose a technique for classifying an audio signal into a sounding interval and a non-sounding interval according to a result of comparison between the SN ratio of the audio signal and a predetermined threshold value.
JP 59-99497 A International Publication No. 2007/017993 Pamphlet

  However, in the technique of selecting the sounding section and the non-sounding section according to the S / N ratio as in Patent Document 1 and Patent Document 2, noise at the time of recording a sound signal (operation sound of an air conditioner or door opening / closing sound) In some cases, a section in which sound is present is selected as a pronunciation section. Then, for example, when voices other than the original target voice such as human uttered sounds are mixed in the sound generation section, there is a problem that the accuracy of the processing of the sound signal for the sound generation section (for example, classification of each section) is lowered. . In view of the above circumstances, an object of the present invention is to solve the problem of distinguishing between a section where a target sound exists and a section where a target sound does not exist in an audio signal.

  In order to solve the above-described problems, a speech processing apparatus according to the present invention includes a storage unit that stores an acoustic model of a target sound, a speech classification unit that segments a speech signal into a plurality of sections on a time axis, and an acoustic model. And a correlation determination means for determining whether or not there is a correlation between the feature quantity of the audio signal in each section, and a section determined to correlate with the acoustic model among a plurality of sections is selected as an effective section and is not correlated with the acoustic model Section selection means for selecting the determined section as a rejection section. According to the above configuration, each section can be classified into an effective section and a rejection section depending on whether or not there is a correlation between the acoustic model of the target sound and the feature amount of the audio signal in each section. Therefore, for example, by selectively using only the effective section, it is possible to improve the accuracy of voice processing (for example, classification for each speaker or voice recognition) for the voice signal in each section.

  In a preferred aspect of the present invention, the voice classifying unit classifies the voice signal into a sounding period and a non-sounding period, and the correlation determining unit is an index value of the correlation between the acoustic model and the feature value of the sound signal in each period. Is compared with a first threshold value to determine whether or not there is a correlation, and threshold value setting means is provided for setting the first threshold value according to the characteristics of the audio signal in the non-sounding interval. According to the above configuration, the first threshold value is variably set according to the characteristics of the audio signal in the non-sounding section. Can increase the sex.

  In a preferred aspect of the present invention, there is provided voiced determination means for determining for each section whether or not the ratio of the number of frames of voiced sound in the section to the total number of frames in each of the plurality of sections exceeds a second threshold value. The section selection means selects the section determined by the correlation determination means to correlate with the acoustic model and determined by the voiced determination means that the ratio of the number of frames of voiced sound exceeds the second threshold value as an effective section. . According to the above configuration, since a section in which the ratio of the number of frames of voiced sound exceeds the second threshold is selected as an effective section, it is possible to select a section of noise similar to the target sound as a rejection section. .

  In a further preferred aspect, the correlation determining means compares only the feature amount of the frame of the voiced sound in each of the plurality of sections with the acoustic model. For example, since the difference between the target sound such as a voice produced by a human and noise is particularly significant with respect to the characteristics of the voiced sound, only the feature amount of the frame of the voiced sound is compared with the acoustic model. There is an advantage that the accuracy of the determination by is improved.

  The speech processing apparatus according to a specific aspect of the present invention classifies, for each speaker, a plurality of sections selected by the section selection unit as effective sections among the plurality of sections based on the feature amount of the audio signal in the section. Voice classification means is provided. According to this aspect, since there is a high possibility that the effective section includes the target sound, a feature value that accurately reflects the characteristic of the target sound is extracted from the audio signal in the effective section. Therefore, according to this aspect in which only effective sections are targeted for classification, each section can be classified with high accuracy according to the characteristics of the audio signal.

  The audio processing apparatus according to the present invention is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to audio processing, and a general-purpose arithmetic processing apparatus such as a CPU (Central Processing Unit). It is also realized through collaboration with the program. The program according to the present invention includes a computer having storage means for storing an acoustic model of a target sound, voice classification processing for dividing a voice signal into a plurality of sections on a time axis, an acoustic model, and a voice signal in each section. A correlation determination process (for example, step S5 in FIG. 3) for determining whether or not there is a correlation with the feature amount, and a section determined to be correlated with the acoustic model among a plurality of sections is selected as an effective section to be processed, The computer is caused to perform section selection processing (for example, step S6 and step S10 in FIG. 3) for selecting a section determined to be uncorrelated with the acoustic model as a rejection section that is not subject to processing. Even with the above program, the same operations and effects as those of the speech processing apparatus according to the present invention are exhibited. The program of the present invention is provided to the user in a form stored in a computer-readable recording medium and installed in the computer, or is provided from the server device in the form of distribution via a communication network. Installed on.

  The present invention is also specified as a method of processing speech. An audio processing method according to one aspect of the present invention includes an audio classification procedure for dividing an audio signal into a plurality of sections on a time axis, an acoustic model stored in a storage device, and a feature amount of an audio signal in each section A correlation determination procedure for determining the presence or absence of correlation, and selecting a section determined to correlate with the acoustic model among a plurality of sections as an effective section to be processed, and processing a section determined not to be correlated with the acoustic model Section selection procedure for selecting a rejection section that is not subject to According to the above method, the same operation and effect as the sound processing apparatus according to the present invention are exhibited.

  In addition, the speech processing apparatus according to another aspect of the present invention includes speech classification means for segmenting an audio signal into a plurality of sections on a time axis, and voiced sound in the section with respect to the total number of frames in each of the plurality of sections. Voiced determination means for determining for each section whether or not the ratio of the number of frames exceeds a threshold; selecting a section where the ratio of the number of frames of the voiced sound exceeds a threshold as an effective section; Section selection means for selecting a section in which the ratio of the number of frames is below a threshold as a reject section. According to the above aspect, each section is sorted into an effective section and a rejection section according to the ratio of voiced sound frames in each section. Therefore, for example, by selectively using only the effective section, it is possible to improve the accuracy of voice processing (for example, classification for each speaker or voice recognition) for the voice signal in each section.

<A: First Embodiment>
FIG. 1 is a block diagram showing the configuration of the speech processing apparatus according to the first embodiment of the present invention. As shown in FIG. 1, the voice processing device 100 is a computer system that includes a control device 10 and a storage device 20. The control device 10 is an arithmetic processing device that executes a program. The storage device 20 stores a program executed by the control device 10 and various data used by the control device 10. A known storage medium such as a semiconductor storage device or a magnetic storage device is arbitrarily adopted as the storage device 20. An output device 30 is connected to the control device 10. The output device 30 according to the present embodiment is a display device that displays various images under the control of the control device 10.

  The storage device 20 stores an audio signal S representing a waveform on the time axis of audio. The audio represented by the audio signal S of this embodiment is audio recorded using a sound collection device in a conference where a plurality of participants speak at any time. A waveform on the time axis of the audio signal S is illustrated in part (A) of FIG. The control device 10 generates a meeting minutes from the audio signal S by executing a program stored in the storage device 20. The minutes are records of a meeting in which the contents (characters) of each of a plurality of participants are arranged in time series.

  Furthermore, the storage device 20 stores an acoustic model that represents the acoustic characteristics of the speech that is the object of processing by the speech processing device 100 (hereinafter referred to as “target sound”). In this embodiment, a voice uttered by a human is the target sound. That is, a process of extracting acoustic feature quantities (for example, MFCC (Mel Frequency Cepstral Coefficient)) from the uttered sound is executed for many speakers and utterances of various contents, and the extracted many feature quantities are statistically processed. As a result of the processing, an acoustic model is generated that shows the general (average) characteristics of the uttered sound by humans. The control device 10 may generate an acoustic model, or an acoustic model generated by an external device may be stored in the storage device 20.

The acoustic model of the present embodiment is, for example, a mixed model λ that models the distribution of feature values (MFCC vectors) extracted from a large number of various utterances as samples as a weighted sum of M probability distributions ( M is a natural number of 2 or more). For the generation of the mixed model λ, a known technique such as an EM (Expectation-Maximization) algorithm is arbitrarily employed. The mixed model λ of this embodiment is a Gaussian mixed model expressed by the following equation (1) as a weighted sum of M normal distributions.
λ = {pi, μi, Σi} (i = 1 to M) (1)
In the equation (1), pi is a weight value (weight value) of the i-th normal distribution. The sum of the weights p1 to pM is 1. In Expression (1), μi is an average vector of the i-th normal distribution, and Σi is a covariance matrix of the i-th normal distribution. It should be noted that even if a symbol actually means a vector, such as μi in equation (1), this specification means that the symbol means a vector, for example, by clearly expressing it as an `` average vector ''. The vector symbol (the arrow pointing right on the character) is omitted.

  As illustrated in FIG. 1, the control device 10 functions as a voice classification unit 12, a selection processing unit 13, a voice classification unit 14, and a voice recognition unit 16. Each function of the control device 10 of FIG. 1 is also realized by an electronic circuit such as a DSP dedicated to voice processing. In addition, the control device 10 may be implemented as a plurality of integrated circuits.

  As shown in part (D) of FIG. 2, the voice classification unit 12 classifies the voice signal S stored in the storage device 20 into a plurality of sound generation sections PA and a plurality of non-sound generation sections PB along the time axis. To do. The sound generation section PA is a section where speech (target sound or noise) exists, and the non-sound generation section PB is a section where there is no sound or a volume is sufficiently small.

  The voice classification unit 12 executes a first process and a second process. As shown in part (B) of FIG. 2, the first process is a process of detecting a section in which the S / N ratio and the volume (amplitude) of the audio signal S exceed the threshold as the sound generation section PA. Sections other than the sound generation section PA are non-sound generation sections PB.

  When utterances by a plurality of speakers are continuous without being spaced apart or partially overlapped, it is difficult to classify the audio signal S for each speaker by only the first process. Therefore, the sound classification unit 12 generates a sound with each of a plurality of valleys D appearing in the envelope (envelope) E of the waveform of the sound signal S as a boundary, as shown in part (C) and part (D) of FIG. A second process for dividing the section PA is executed. In a series of utterances by humans, generally, there is a tendency that the volume gradually increases from the start point of the utterance and gradually decreases from an intermediate point to the end point of the utterance. Therefore, according to the configuration in which the sound generation section PA is divided with the valley portion D as a boundary, even if a plurality of utterances are continuous or overlapped, the utterances by each speaker are divided into separate sound generation sections PA. In the following, the total number of sounding sections PA after the classification by the voice classifying unit 12 is J (J is an integer of 2 or more). In addition to the above examples, a known technique is arbitrarily adopted for detection of the sounding section PA and the non-sounding section PB.

  By the way, when the section where the SN ratio and the sound volume exceed the threshold is detected as the sound generation section PA as described above, the section where the voice other than the target sound (for example, the ringing tone of the telephone) exists in the voice signal S is the sound generation section PA. There is a possibility of being detected. Therefore, as shown in part (E) of FIG. 2, the selection processing unit 13 selects a plurality of sound generation sections PA defined by the speech classification unit 12 as sections (hereinafter referred to as “effective sections”) where the target sound is highly likely to exist. And PA1 and a section (hereinafter referred to as “rejection section”) PA2 where there is a low possibility that the target sound exists. That is, the section where the target sound does not exist among the plurality of sound generation sections PA is removed as the reject section PA2. The specific operation of the sorting processing unit 13 will be described later.

  The voice classification unit 14 in FIG. 1 classifies the voice signal S of each effective section PA1 selected by the selection processing unit 13 among the plurality of sound generation sections PA for each speaker. The non-sound generation section PB defined by the voice classification section 12 and the rejection section PA2 selected by the selection processing section 13 are excluded from the classification targets. A known clustering technique is arbitrarily employed for classification of each effective section PA1.

  For example, the audio classification unit 14 performs frequency analysis including FFT (Fast Fourier Transform) processing on the audio signal S in each effective interval PA1, thereby acoustic characteristics of the audio signal S in the effective interval PA1. A quantity (for example, MFCC) is extracted, and a plurality of effective sections PA1 are classified into clusters so that each effective section PA1 having a similar feature amount belongs to a common cluster. Therefore, the effective section PA1 that is highly likely to be uttered by the same speaker in the audio signal S is classified into a common cluster. Then, the voice classification unit 14 identifies each of the plurality of speaker's identification codes, the start and end times of each effective section PA1 classified into the speaker's cluster, and the sound signal S in each effective section PA1. Are stored in the storage device 20. In the configuration in which the user designates the number of participants in the conference as a known number, a configuration in which a plurality of effective sections PA1 are classified into a number of clusters corresponding to the number of people is preferably employed.

  The voice recognition unit 16 specifies the content of the utterance for each speaker as a character from the voice signal S of each effective section PA1 classified into each cluster. A known speech recognition technique is arbitrarily employed for the process of recognizing characters from the speech signal S in each effective section PA1. For example, the speech recognition unit 16 firstly updates (speaker adaptation) the initial acoustic model according to the acoustic feature amount of the speech signal S of each effective section PA1 classified into one cluster. Then, an acoustic model that inherently reflects the voice characteristics of the speaker corresponding to the cluster is generated, and secondly, it is extracted from the acoustic model after speaker adaptation and the voice signal S of each effective section PA1 in the cluster. The character of the utterance is identified by comparing with the feature amount.

  The control device 10 outputs the processing result by the voice recognition unit 16 to the output device 30. The output device 30 displays an image of the minutes in which the time of the utterance, the identification code of the utterer (for example, the name of the utterer), and the characters identified by the voice recognition unit 16 regarding the content of the utterance are arranged in time series. To do.

  Next, a specific example of processing by the sorting processing unit 13 will be described with reference to FIG. The process in FIG. 3 is started when the process by the voice classifying unit 12 is completed. As shown in the figure, the selection processing unit 13 extracts the feature amount of the audio signal S for each of the J sound generation sections PA (step S1). More specifically, the selection processing unit 13 performs frequency analysis for each of a plurality of frames F (see part (A) in FIG. 2) that divide each sounding section PA, so that each sounding section PA in each sounding section PA is analyzed. A time series of MFCC vectors (hereinafter referred to as “feature vectors”) in the frame F is extracted as a feature amount. But the feature-value extracted in step S1 is not limited to MFCC.

  Next, the selection processing unit 13 selects the unselected and earliest (oldest) sounding section PA from among the J sounding sections PA (step S2). Next, the selection processing unit 13 sets a threshold value TH1 according to the sound signal S of the latest non-sounding section PB before the start of the sounding section PA selected in step S2 (hereinafter, particularly referred to as “selected section PA_S”). (Step S3). Since the non-sound generation section PB is basically a section in which only noise (environmental sound) exists, the process of step S3 corresponds to the process of setting the threshold value TH1 according to the noise when the audio signal S is recorded. Specifically, the selection processing unit 13 calculates the average intensity (hereinafter referred to as “noise level”) of the audio signal S in the non-sound generation section PB immediately before the selection section PA_S, and the threshold increases as the noise level increases. The threshold value TH1 is variably controlled so that TH1 becomes small.

  Next, the selection processing unit 13 calculates a correlation index value indicating the degree of correlation between the mixed model λ stored in the storage device 20 and the feature vector x of the speech signal S in the selected section PA_S (step S4). More specifically, the selection processing unit 13 is a numerical value obtained by averaging the probability (likelihood) that each feature vector x in the selected section PA_S appears from the mixed model λ with respect to all the feature vectors x in the selected section PA_S (hereinafter referred to as the following). L (referred to as “average likelihood”) is calculated as a correlation index value.

When one feature vector x is a D-dimensional vector, the likelihood p (x | λ) at which the feature vector x appears from the mixed model λ is calculated by the following equation (2).

In step S4, the selection processing unit 13 calculates the average likelihood L by substituting the K feature vectors x (x1 to xK) in the selected section PA_S into the equation (3). As understood from the equation (3), the average likelihood L increases as the sound feature represented by the acoustic model is similar to the feature of the sound signal S in the selected section PA_S.

  Next, the selection processing unit 13 determines whether or not the average likelihood L of the selected section PA_S is lower than the threshold value TH1 (step S5). Since the mixed model λ comprehensively reflects various human utterances, it is unlikely that the speech in the selected section PA_S whose average likelihood L is lower than the threshold TH1 is a human utterance. Therefore, when the result of step S5 is affirmative (L <TH1), the sorting processor 13 sorts the selected section PA_S at the current stage into the reject section PA2 (step S6). As described above, the process of step S5 is a process of determining whether or not there is a possibility that the sound in the selected section PA_S is a human uttered sound according to the presence or absence of the correlation between the sound signal S and the mixed model λ. It is.

  Since the feature vector x extracted in step S1 is affected by noise in the audio signal S, if the threshold value TH1 is a fixed value, the higher the noise level of the audio signal S, the more actual the target sound. There is a high possibility that the result of step S5 will be affirmative in spite of the sound generation period PA that includes it. In this embodiment, the threshold TH1 is set to a smaller numerical value as the noise level of the audio signal S is higher (that is, the rate at which the result of step S5 becomes affirmative is lower), so that the selection section PA_S including the target sound is the rejection section PA2. It is possible to reduce the possibility of erroneous determination.

  By the way, even if the target sound is not actually included in the selection section PA_S, if the selection section PA_S includes noise similar to a human voice, the result of step S5 is negative (ie, It is not determined to be the rejection section PA2.) Therefore, if the result of step S5 is negative, the sorting processor 13 sorts the selected section PA_S into the valid section PA1 or the reject section PA2 by a method that does not use the mixed model λ (step S7 to step S9).

  When a human utters naturally (for example, unless only intentionally utters unvoiced sound continuously), there is a tendency that voiced sound exists for a length of time exceeding a predetermined ratio in a section where utterance continues. Therefore, in the present embodiment, the selection section PA_S is selected as the effective section PA1 (sounding section PA rich in voiced sound) and the rejection section PA2 (sounding section PA rich in unvoiced sound) according to the proportion of the voiced sound section in the selection section PA_S. ) And sorting.

  In step S7, the selection processing unit 13 determines, for each of the plurality of frames F in the selected section PA_S, whether the voice indicated by the voice signal S is a voiced sound or an unvoiced sound. A known technique is arbitrarily employed for voiced / unvoiced determination. For example, the selection processing unit 13 calculates the maximum value (hereinafter referred to as “autocorrelation value”) of the autocorrelation function that is an index of the periodicity of the audio signal S for each frame F, and the autocorrelation value exceeds a predetermined value. The frame F (that is, the frame F having a high periodicity of the audio signal S) is determined as a voiced sound, and the frame F having an autocorrelation value lower than a predetermined value is determined as an unvoiced sound. A configuration in which only a frame F in which a clear pitch (basic frequency) is detected from the audio signal S is determined as a voiced sound is also preferably employed.

  Next, the selection processing unit 13 calculates a ratio R of the number of frames F determined to be voiced in step S7 out of the total number of frames F in the selected section PA_S (step S8), and the ratio R is a predetermined threshold value. It is determined whether or not TH2 is exceeded (step S9). When the determination in step S9 is negative (that is, when the ratio of the unvoiced sound frame F is high in the selection section PA_S), the selection processing unit 13 selects the selection section PA_S in the current stage as the rejection section PA2 (step S6). On the other hand, when the determination in step S9 is affirmative (that is, when the ratio of the frame F of voiced sound is high in the selection section PA_S), the selection processing unit 13 selects the selection section PA_S into the effective section PA1 (step S10).

  When step S6 or step S10 is executed, the selection processing unit 13 determines whether or not all sound generation sections PA of the audio signal S have been selected (step S11). If the result of step S11 is negative, the selection processing unit 13 selects the sound generation section PA immediately after the current selection section PA_S as the new selection section PA_S in step S2, and then executes the processes after step S3. To do. When the selection of all the sound generation sections PA is completed (step S11: YES), the selection processing unit 13 ends the process of FIG.

  As described above, in this embodiment, the sound generation section PA is distinguished into the effective section PA1 and the rejection section PA2 according to the presence or absence of the target sound, so that the rejection section PA2 that does not include the target sound is designated as the speech classification unit 14 or By excluding it from the target of processing by the speech recognition unit 16, it is possible to effectively reduce the influence of noise. For example, by reducing the influence of noise, it is possible to extract a feature value that faithfully reflects the characteristics of the uttered sound of each speaker, so that the classification of each sounding section PA (effective section PA1) by the speech classification unit 14 The accuracy of speech processing using feature quantities such as speaker adaptation and speech recognition by the speech recognition unit 16 is improved. Therefore, an accurate minutes can be created from the audio signal S.

  By the way, in the above form, the structure which selects sounding area PA into effective area PA1 and rejection area PA2 according to whether the audio | voice signal S correlates with the acoustic model (mixed model (lambda)) of a human vocal sound is illustrated. did. On the other hand, for example, a configuration (hereinafter referred to as “proportional”) using an acoustic model (hereinafter referred to as “noise model”) of noise that may occur during recording of the audio signal S is also assumed. In contrast, when the correlation between the speech signal S and the noise model is high, the sounding section PA is selected as the rejection section PA2, and when the correlation between the speech signal S and the noise model is low, the sounding section PA becomes the effective section PA1. Selected.

  However, the characteristics of noise are extremely diverse compared to the characteristics of human vocal sounds. Therefore, even if a noise model is created assuming specific noise, there is a high possibility that noise that cannot be covered by the noise model is included in the speech signal S. That is, in the comparative configuration, there is a problem that the sound generation section PA that does not include the target sound cannot be sufficiently removed. On the other hand, in this embodiment, since an acoustic model of a human uttered sound is used, even if the speech signal S includes various noises, the pronunciation period PA that does not include the target sound is effectively removed. There is an advantage that you can.

<B: Second Embodiment>
Next, a second embodiment of the present invention will be described. In this embodiment, VQ (Vector Quantization) distortion is adopted as a correlation index value between the acoustic model and the audio signal S instead of the average likelihood L in the first embodiment. In addition, about the element in which a function and an effect | action are equivalent to 1st Embodiment in each following form, the same code | symbol as the above is attached | subjected and each detailed description is abbreviate | omitted suitably.

Acoustic model stored in advance in the storage device 20, a number and variety of a number of feature amounts extracted from the utterance (MFCC) codebook generated from the vector of the sample (codebook) is C ^A. A known technique such as a k-means method or an LBG algorithm is arbitrarily employed for generating the code book.

In step S4 of FIG. 3, distinguishing processing unit 13, the codebook C ^A stored in the storage device 20, a plurality of feature vector x is extracted in step S1 from the sound signal S of the selection section PA_S (eg MFCC) Based on this, the VQ distortion D is calculated. The VQ distortion D is calculated by, for example, the following formula (4).

In formula (4), | C ^A | is the size of the code book C ^A , and C ^A (i) is the i-th code vector (centroid vector) in the code book C ^A. Further, xj denotes the j-th among the feature vectors x1 to xn _B of n _B-number that is extracted from the selected interval PA_S (number of frames F in a selected interval PA_S) (j = 1~n _B) . d (X, Y) is the Euclidean distance between the vector X and the vector Y. That is, the VQ distortion D is the minimum value (min) of | C ^A | centroid vectors in the codebook C ^A that is an acoustic model and the feature vector x of the selected section PA_S, and n _B feature vectors x 1 to x. is a value obtained by averaging over xn _B.

  As understood from the above description, the VQ distortion D becomes smaller as the voice in the selected section PA_S resembles a human voice. Therefore, in step S4 in FIG. 3, the selection processing unit 13 variably controls the threshold value TH1 so that the threshold value TH1 increases as the noise level in the non-sound generation period PB immediately before the selection period PA_S increases. In step S5 in FIG. 3, the selection processing unit 13 determines whether or not the VQ distortion D exceeds the threshold value TH1, and if it exceeds the threshold value TH1, the selection section PA_S is selected as the rejection section PA2 (step) S5: YES), if below the threshold TH1, the process proceeds to step S7 (step S5: NO). Other operations are the same as those in the first embodiment. In this embodiment, the same effect as that of the first embodiment is obtained.

<C: Modification>
Various modifications can be made to each of the above embodiments. An example of a specific modification is as follows. Two or more aspects may be arbitrarily selected from the following examples and combined.

(1) Modification 1
In each of the above embodiments, the feature amount (feature vector x) extracted from each frame F of the sound production section PA is compared with the acoustic model regardless of whether it is voiced sound or unvoiced sound. A configuration in which only the feature amount extracted from the frame F of the voiced sound is compared with the acoustic model is also employed. The acoustic model stored in the storage device 20 is generated based on the feature amount in the voiced sound section excluding the unvoiced sound section and the silent section from the sample voice. The selection processing unit 13 uses only the feature amount extracted from the frame F of the voiced sound among the plurality of frames F in the selected section PA_S, and uses the average likelihood L (in the second embodiment, in the second embodiment). VQ distortion D) is calculated, and whether or not there is a correlation between the acoustic model and the audio signal S in the selected section PA_S is determined in step S5. Since the difference between the noise and the target sound is particularly remarkable with respect to the characteristics of the voiced sound, according to the configuration in which only the frame F of the voiced sound in the sound generation section P is used for comparison with the acoustic model as in the above modification. The accuracy of the determination in step S5 can be improved.

(2) Modification 2
In the above embodiment, the threshold value TH1 is set based on the noise level in the non-sound generation section PB immediately before the selected section PA_S (step S3), but the criterion for setting the threshold value TH1 is appropriately changed. For example, a configuration in which the threshold value TH1 is set based on the noise level in the leading non-sounding section PB of the audio signal S and the threshold value TH1 is commonly applied in step S5 for selecting each sounding section PA is also adopted. However, according to the configuration of the first embodiment in which the noise level of the non-sound generation section PB immediately before the selection section PA_S is applied to the selection section PA_S, the noise level changes at a point in the middle of the audio signal S. Even so, since the threshold value TH1 is updated according to the noise level after the change, the possibility that the accuracy of selection in step S5 is reduced is reduced.

  In each of the above embodiments, the threshold value TH2 in step S9 is a fixed value. However, there is a configuration in which the threshold value TH2 is variably controlled by a method similar to the threshold value TH1 (the method exemplified in the first embodiment and this modification). Adopted. As the noise level of the audio signal S is higher, the error of the ratio R calculated in step S8 increases. Therefore, in the embodiment in which the threshold value TH2 is a fixed value, the selection section PA_S including the target sound is erroneously determined as the rejection section PA2. The possibility of being increased. Therefore, the selection processing unit 13 sets the threshold value TH2 so that the threshold value TH2 becomes smaller as the noise level becomes higher in the non-sound generation period PB (or the first non-sounding period PB of the audio signal S) immediately before the selection period PA_S. According to the above configuration, it is possible to reduce the possibility that the selection section PA_S including the target sound is erroneously determined as the rejection section PA2.

(3) Modification 3
In each of the above embodiments, one acoustic model is used. However, a plurality of acoustic models may be selectively used to sort the sound generation section PA into the effective section PA1 and the rejection section PA2. For example, a plurality of acoustic models generated from a plurality of types of sounds having different average pitches are created in advance and stored in the storage device 20. In step S4 of FIG. 3, the selection processing unit 13 detects the pitch (average pitch) of the audio signal S in the selected section PA_S, and uses the acoustic model corresponding to the pitch among the plurality of acoustic models to calculate the average likelihood. Degree L (VQ distortion D in the second embodiment) is calculated. According to the above configuration, even when the voice signal S includes voices of various pitches, such as when male voices and female voices are mixed, the pronunciation period PA is accurately set as the valid period PA1. And the rejection section PA2.

(4) Modification 4
The method by which the audio classification unit 12 classifies the audio signal S is not limited to the above examples. For example, the first process of dividing the audio signal S into the sounding section PA and the non-sounding section PB according to the S / N ratio and volume of the sound signal S, and the first process of dividing the sound signal S with the valley portion D of the envelope E as a boundary. Only one of the two processes may be executed. In addition, a configuration in which the audio signal S is divided into sections of a fixed or variable time length set regardless of the characteristics of the audio signal S is also adopted. In other words, the division between the sounding section PA and the non-sounding section PB is not essential in the embodiment of the present invention.

(5) Modification 5
In each of the above embodiments, the determination in step S5 using the correlation index value (average likelihood L and VQ distortion D) for the acoustic model and the determination in step S9 using the ratio R of the voiced frame F are executed. did. However, a configuration in which each sounding section PA is selected as an effective section PA1 and a rejection section PA2 based on only the determination result in step S5 (that is, a configuration in which steps S7 to S9 in FIG. 3 are omitted) is also employed. A configuration is also adopted in which each sounding segment PA is sorted into an effective segment PA1 and a rejection segment PA2 based only on the determination in step S9 (that is, a configuration in which steps S3 to S5 in FIG. 3 are omitted).

(6) Modification 6
A printing device that prints the minutes created by the voice processing device 100 may be adopted as the output device 30. However, it is not necessary to output the result of processing by the speech processing apparatus 100 in the form of minutes (characters), and for example, the result of classification by the speech classification unit 14 can be output. For example, according to the configuration in which the sound signal S in the effective section PA1 including the time specified by the user among the plurality of effective sections PA1 classified by the voice classification unit 14 is output as sound waves from a sound emitting device (for example, a speaker). It is possible to effectively support the user's work of creating the minutes of the meeting while selectively listening to the statements of each speaker and confirming them appropriately. A configuration is also adopted in which the selection processing unit 13 outputs the result of selecting the sound generation section PA into the effective section PA1 and the rejection section PA2 from the speech processing apparatus 100 to an external device. In the external device, the same processing as the speech classification unit 14 in FIG. 1 or other appropriate processing is executed on the output from the speech processing device 100. For example, only the effective section PA1 selected by the selection processing unit 13 among the plurality of sound generation sections PA is selectively output to an external device, and predetermined processing (classification or speech recognition for each speaker) is performed for each effective section PA1. Is executed by an external device. As described above, the speech classification unit 14 and the speech recognition unit 16 are not essential elements for the speech processing apparatus 100.

(7) Modification 7
In each of the above embodiments, the audio signal S stored in advance in the storage device 20 is the target of processing, but is sequentially supplied via the audio signal S supplied from the sound collection device (microphone) and the communication network. The processing may be executed in real time for the audio signal S. In addition, the type of sound represented by the audio signal S is arbitrary in the present invention. For example, according to the configuration in which an acoustic model having a performance sound of a specific instrument as a target sound is stored in the storage device 20, a sound other than the target sound (for example, applause) from the sound section S recorded at the performance of the musical instrument. It is possible to exclude the section of sound) as the rejection section PA2.

It is a block diagram which shows the structure of the audio processing apparatus which concerns on 1st Embodiment of this invention. It is a conceptual diagram for demonstrating operation | movement of a speech processing unit. It is a flowchart which shows operation | movement of a selection process part.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Voice processing apparatus, 10 ... Control apparatus, 12 ... Voice classification part, 13 ... Sorting process part, 14 ... Voice classification part, 16 ... Voice recognition part, 20 ... Memory | storage device, 30 ... Output device, PA ... sounding section, PB ... non-sounding section, S ... speech signal.

Claims (5)

Storage means for storing an acoustic model of the target sound;
Audio classification means for dividing the audio signal into a plurality of sections on the time axis;
Correlation determining means for determining whether or not there is a correlation between the acoustic model and a feature amount of the audio signal in each section;
A speech processing apparatus comprising: a section selecting unit that selects a section determined to correlate with the acoustic model among the plurality of sections as an effective section, and selects a section determined not to be correlated with the acoustic model as a rejection section. . The voice classification means classifies the voice signal into a sounding section and a non-sounding section,
The correlation determination means determines the presence or absence of correlation by comparing the index value of the correlation between the acoustic model and the feature value of the audio signal in each section with a first threshold value,
The speech processing apparatus according to claim 1, further comprising a threshold setting unit configured to set the first threshold according to characteristics of the speech signal in the non-sounding section. Voiced determination means for determining for each section whether or not the ratio of the number of frames of voiced sound in the section to the total number of frames in each of the plurality of sections exceeds a second threshold;
The section selecting means determines the section determined by the correlation determining section to correlate with the acoustic model, and the voiced determining section determines that the ratio of the number of voiced frames exceeds a second threshold as an effective section. The voice processing device according to claim 1 or 2. The speech processing apparatus according to claim 1, wherein the correlation determination unit compares only a feature amount of a voiced sound frame in each of the plurality of sections with the acoustic model. In a computer having storage means for storing an acoustic model of the target sound,
Audio classification processing for dividing the audio signal into a plurality of sections on the time axis;
A correlation determination process for determining whether or not there is a correlation between the acoustic model and a feature amount of the audio signal in each section;
Of the plurality of sections, a section determined to correlate with the acoustic model is selected as an effective section to be processed, and a section determined not to be correlated with the acoustic model is selected as a rejection section that is not subject to the processing. A program that executes the section selection process.

Images