JP2009020458A - Voice processing apparatus and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately classify a plurality of sections obtained by segmenting a voice signal, for each speaker. <P>SOLUTION: A feature extracting section 41 extracts a feature amount for each of N sections B in which the speech signal S is segmented on a time axis. An index calculating section 43 calculates a similarity index value for indicating similarity of the feature amount in two sections B for all combinations for selecting two sections B out of N sections B. A voice classification section 45 classifies N sections B into a plurality of sets on the basis of the similarity index value of each section so that each of N sections B, and a section B where the feature amount is most similar with that of the section B may belong to the same cluster. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a technique for classifying (clustering) a plurality of sections obtained by dividing an audio signal on a time axis for each speaker.

If audio signals recorded in an environment (for example, a conference) where a plurality of speakers speak at any time can be classified and classified for each speaker, it is convenient to use it for the preparation of conference minutes, for example. In Patent Document 1, an acoustic feature amount is extracted for each of a plurality of sections obtained by dividing an audio signal on a time axis, and a plurality of sections in which the degree of matching (similarity) of feature amounts exceeds a threshold value are the same speaker. A technique for classifying a voice signal as a voice signal is disclosed.
JP 2005-321530 A

  However, in the technique of Patent Document 1, since the threshold value for comparing the matching degree is a fixed value, depending on the utterance conditions (for example, utterance length, S / N ratio, etc.), the speech signals of each section are correctly classified. There are cases where it is not possible. Although it is conceivable to control the threshold value variably according to the conditions at the time of utterance, it is difficult to set an optimum threshold value according to various conditions at the time of utterance. In view of the above circumstances, an object of the present invention is to solve the problem of accurately classifying a plurality of sections into which speech signals are classified for each speaker.

  In order to solve the above-described problems, a speech processing apparatus according to the present invention includes a feature extraction unit that extracts a feature amount for each of a plurality of sections obtained by dividing an audio signal on a time axis, and 2 out of the plurality of sections. For a plurality of combinations for selecting individual sections, an index calculating means for calculating similarity index values indicating similarity of feature quantities in two sections, and each of the plurality of sections and the feature quantities are most similar to the sections Voice classification means for classifying a plurality of sections into a plurality of sets based on similarity index values of the sections so that the sections belong to the same set. According to the above configuration, since each of the plurality of sections and the section having the most similar feature amount to the section are classified into the same set, it is theoretically unnecessary to compare the similarity index value with a predetermined threshold value. It is. Therefore, it is possible to accurately classify a plurality of sections by reducing the influence of the conditions at the time of utterance (for example, sound volume and noise volume) on the accuracy of classification.

  In a preferred aspect of the present invention, the index calculation means calculates the cross-correlation value of the feature values of the two sections as the similarity index value. According to this aspect, there is an advantage that the processing load for calculating the similarity index value is reduced. However, the similarity index value is not limited to the cross-correlation value of the feature amount. For example, in an aspect in which the feature extraction unit extracts a time series of feature vectors of an audio signal for each of a plurality of sections as a feature amount, the distribution of feature vectors in one of the two sections is a plurality of probability distributions. In the configuration in which the index calculation means calculates the average value of the likelihood that each feature vector of the other section appears from the mixed model modeled in step 1 as the similarity index value, or the feature vector of one section of the two sections A configuration is adopted in which the index calculation means calculates the average value of the vector quantization distortion between the codebook obtained by vector quantization of the time series and each feature vector in the other section as the similarity index value.

  In a preferred aspect of the present invention, the speech classification means includes a section in which the similarity index value for each of all the other sections is dissimilar among the plurality of sections (for example, the rank of similarity falls below a predetermined value). The section having the lowest section or similarity in each of the other sections) is classified into a set separate from all other sections. According to this aspect, even if, for example, a single speaker utters only in one section, the utterance section can be classified into a set different from other sections.

  A speech processing apparatus according to a preferred aspect of the present invention includes speech classification means for partitioning a speech signal into a plurality of sections with each valley in the envelope of the speech signal waveform as a boundary. According to this aspect, since the audio signal is divided into a plurality of sections with each valley portion of the envelope of the audio signal as a boundary, for example, when a plurality of speakers speak in sequence with almost no interval, In addition, it is possible to divide utterances by each speaker into separate sections. Therefore, the accuracy of classification by the voice classification unit can be increased.

  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 feature extraction process (for example, step S2 in FIG. 3) for extracting a feature amount for each of a plurality of sections obtained by dividing an audio signal on a time axis, and two sections among the plurality of sections. Index calculation processing (for example, step S5 in FIG. 3) for calculating similarity index values indicating similarity of feature quantities in two sections, a plurality of sections, and features in the sections A voice classification process (for example, step S8 in FIG. 3) for classifying a plurality of sections into a plurality of sets based on the similarity index value of each section so that the most similar sections belong to the same set. Let it run. 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. A speech processing method according to an aspect of the present invention includes a feature extraction procedure for extracting a feature amount for each of a plurality of sections obtained by dividing an audio signal on a time axis, and selecting two sections from the plurality of sections. For a plurality of combinations, an index calculation procedure for calculating a similarity index value indicating similarity of feature quantities in two sections, and each of the sections and the section having the most similar feature quantities in the section are the same set And a speech classification procedure for classifying a plurality of sections into a plurality of sets based on the similarity index value of each section. According to the above method, the same operation and effect as the sound processing apparatus according to the present invention are exhibited.

<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 voice represented by the voice signal S of this embodiment is voice collected using a sound collection device in a conference where a plurality of participants speak at any time. FIG. 2 illustrates a waveform on the time axis of the audio signal S.

  The control device 10 in FIG. 1 creates a meeting minutes from the audio signal S by executing a program stored in the storage device 20. The minutes are conference records in which the contents of each participant's remarks are arranged in time series for each participant. As shown in FIG. 1, the control device 10 functions as a voice classification unit 12, a classification processing 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. Further, the control device 10 may be distributed and mounted on a plurality of integrated circuits.

  As shown in FIG. 2, the voice classification unit 12 classifies the voice signal S into a plurality of sections B along the time axis. One section B is a period in which it is estimated that there is a high possibility that one speaker speaks continuously. A series of human utterances (especially in a conference) generally has 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. In consideration of the above tendency, the audio classification unit 12 of the present embodiment, as shown in FIG. 2, uses the audio signal with each of a plurality of valleys D appearing in an envelope E of the waveform of the audio signal S as a boundary. S is divided into a plurality of sections B.

  According to the above configuration, for example, when the last part of one speaker's utterance overlaps with the first part of another speaker's utterance, or a plurality of speakers uttered sequentially without intervals. Even in this case, it is possible to divide the utterance by each speaker into separate sections B. In the following, it is assumed that the audio signal S is divided into N (N is a natural number) sections B. A unique identifier (number) is assigned to each of the N sections B.

  The classification processing unit 14 in FIG. 1 is means for classifying the N sections B into which the audio classification unit 12 has classified the audio signal S for each speaker (conference participant). That is, the section B that is likely to be uttered by the same speaker in the audio signal S is classified into a common set (cluster). The classification processing unit 14 stores the classification result in the storage device 20. That is, the classification processing unit 14 associates each identification code of a plurality of speaker, the start and end times of each section B classified into the cluster of the speaker, and the audio signal S of each section B. And stored in the storage device 20.

  The voice recognition unit 16 in FIG. 1 specifies the content of a utterance for each speaker as a character based on each section B of the voice signal S classified into each cluster. For the process of recognizing characters from the speech signal S of each section B, a known speech recognition technique is arbitrarily employed. 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 section B classified into one cluster. , Generating an acoustic model that uniquely reflects the features of the speaker corresponding to the cluster, and secondly, an acoustic model after speaker adaptation and a feature amount extracted from the speech signal S of each section B in the cluster By comparing, the character spoken by the speaker is specified.

  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 specified by the speech recognition unit 16 regarding the content of the utterance are arranged in time series. indicate.

Next, the classification processing unit 14 will be described in detail. As shown in FIG. 1, the classification processing unit 14 includes a feature extraction unit 41, an index calculation unit 43, and a voice classification unit 45. The feature extraction unit 41 extracts the average power spectrum (hereinafter referred to as “average power spectrum”) of the audio signal S for each of the N sections B as an acoustic feature quantity. The index calculation unit 43 calculates the cross-correlation value Cor of the average power spectrum in the two sections B for all combinations ( _N C ₂ types) that select the two sections B from the _N sections B. . The cross-correlation value Cor is a numerical value (similarity index value) serving as an index of similarity between two types of average power spectrum shapes. The voice classification unit 45 classifies N sections B into a plurality of clusters based on the cross-correlation value Cor of each section B (clustering) so that the sections B whose average power spectra are similar to each other belong to the same cluster. To do.

  FIG. 3 is a flowchart showing a specific operation of the classification processing unit 14. When the voice classifying unit 12 classifies the voice signal S into N sections B in response to an instruction to create the minutes, the process of FIG. 3 is started. The feature extraction unit 41 executes steps S1 to S3, the index calculation unit 43 executes steps S4 to S7, and the speech classification unit 45 executes steps S8 and S9.

  As shown in FIG. 3, the feature extraction unit 41 selects one section B from the N pieces and acquires the audio signal S in the section B from the storage device 20 (step S1). And the feature extraction part 41 extracts the average power spectrum of the area B selected by step S1 as a feature-value (step S2). That is, the feature extraction unit 41 calculates the power spectrum of each frame by performing frequency analysis including FFT (Fast Fourier Transform) processing on the audio signal S of each of the plurality of frames obtained by dividing the section B. The average power spectrum is calculated by averaging the power spectrum for all frames in B. The intensity at a specific frequency in the average power spectrum calculated in step S2 is an average value of the intensity at the frequency in the power spectrum of each frame in the section B. The processes of step S1 and step S2 are repeated for each of the N sections B (step S3). When the average power spectrum (N types) is calculated for all of the N sections B by the above process, the process proceeds to step S4.

In step S <b> 4, the index calculation unit 43 selects one section B (hereinafter referred to as “selected section B” in particular) from among the N sections B. The index calculation unit 43 then selects the average power spectrum of the selected section B and all ((N−1)) sections B other than the selected section B (hereinafter referred to as “contrast section B” in order to distinguish them from the selected section B). Is calculated as the similarity index value (step S5). The cross-correlation value Cor between the average power spectrum SPa in the selected section B and the average power spectrum SPb in one contrast section B is calculated by the following equation (1), for example.

  SPa (i) in equation (1) is the intensity of the average power spectrum SPa at a frequency specified by the variable i (F1≤i≤F2) among a plurality of frequencies (or frequency bands), and SPa_AVE is the frequency F1 Is the average value of the intensity of the average power spectrum SPa in the band from to F2. Similarly, SPb (i) is the intensity of the average power spectrum SPb at the frequency corresponding to the variable i, and SPb_AVE is the average value of the intensity of the average power spectrum SPb in the band from the frequency F1 to the frequency F2. When the average power spectrum SPa and the average power spectrum SPb completely match, the cross-correlation value Cor becomes the maximum value “1”, and the cross-correlation value Cor decreases as the difference between the two increases. Note that the frequency F1 and the frequency F2 are set statistically or experimentally so as to be the lower limit value (F1) and the upper limit value (F2) of the frequency band in which the difference for each speaker tends to be remarkable in the average power spectrum.

  Next, the index calculation unit 43 sorts the identifiers of the (N−1) comparison sections B in the order in which the cross-correlation value Cor with the selected section B is large (that is, the order of high similarity) (step S6). ). For example, it is assumed that the cross-correlation value Cor of the section B of the identifier “13” is the maximum value and the cross-correlation value Cor of the section B of the identifier “16” is the minimum value with respect to the selected section B of the identifier “1”. For example, as shown in FIG. 4, (N-1) identifiers are arranged so that the identifier “13” is the highest and the identifier “16” is the lowest in the selected section B of the identifier “1”. .

  In step S7, the index calculation unit 43 determines whether or not the processing from step S4 to step S6 has been completed for N sections B. If the result of step S7 is negative, the index calculation unit 43 selects a section B different from the current stage as a new selection section B (step S4), and (N-1) contrast sections B. The cross-correlation value Cor is calculated (step S5) and the identifiers of the respective comparison sections B are rearranged (step S6). Therefore, at the stage where the result of step S7 is affirmative (that is, the stage from step S4 to step S6 is completed for N sections B), as shown in FIG. 4, N (N = 16 in FIG. 4). For each section B, a table (hereinafter referred to as “similarity map”) M in which the identifiers of the other (N−1) sections B are arranged in descending order of the cross-correlation value Cor is completed.

  When the similarity map M is created, the speech classification unit 45 includes each of the N sections B and the section B having the maximum cross-correlation value Cor (similarity index value) for the section B belonging to the same cluster. In this way, the N sections B are classified into a plurality of clusters with reference to the similarity map M (step S8). That is, the speech classification unit 45 includes one section B and the section B of the identifier positioned at the highest position with respect to the section B in the similarity map M in the same cluster. For example, in the similarity map M illustrated in FIG. 4, the identifier “13” is positioned at the highest position for the section B with the identifier “1” (that is, the average power spectrum of the section B with the identifier “1” has an identifier The average power spectrum of the section B of “13” is most similar), the identifier “13” is positioned at the top for the section B of the identifier “9”, and the identifier for the section B of the identifier “13” “9” is at the top. Therefore, the voice classification unit 45 classifies the three sections B with the identifiers “1”, “9”, and “13” into the same cluster G1. Similarly, four sections B with identifiers “2”, “3”, “4” and “14” are classified into cluster G2, and two sections B with identifiers “5” and “10” are assigned to cluster G3. being classified.

  By the way, in step S8, since the section B having the maximum cross-correlation value Cor for each section B is classified into the same cluster, for example, even when one speaker speaks only in one section B, the section B B is classified into the same cluster as other section B (section B uttered by another speaker) having the maximum cross-correlation value Cor. Therefore, the speech classification unit 45 is a section B in which the similarity to (N−1) sections B other than itself among the N sections B is lower than a predetermined value (similarity with other sections is low). The section B) is excluded from the clusters classified in step S8 and is categorized as a single cluster (step S9).

  For example, in the case of FIG. 4, the identifier “12” is the highest for the section B with the identifier “16” (that is, the average power spectrum of the section B with the identifier “12” among the N sections B). Is similar to the average power spectrum of the section B having the identifier “16”), the section B having the identifier “16” is classified into the same cluster G4 as the identifier “12” in the step S8. However, the identifier “16” is set at the lowest level of the similarity map M with respect to the section B of all other identifiers “1” to “15”. That is, the average power spectrum of the section B with the identifier “16” has a low correlation with the average power spectrum of any other section B. Therefore, the voice classification unit 45 excludes the section B of the identifier “16” from the cluster G4 classified in step S8 and classifies it into an independent cluster G6. According to the above configuration, it is possible to appropriately classify the only section B (identifier “16”) uttered by a specific speaker without mixing with the clusters of other speakers.

  As described above, since each section B in which the cross-correlation value Cor of the average power spectrum is maximum is classified into the same cluster, it is not necessary to compare the cross-correlation value Cor with a predetermined threshold value. Therefore, it is possible to accurately classify the speech signal S of each section B for each speaker regardless of the conditions at the time of speaking, and create appropriate minutes in which each speech in the meeting is faithfully distinguished for each participant can do.

  In addition, not only the section B having the maximum cross-correlation value Cor with each selected section B is specified, but also a similarity map M indicating the rank of similarity with respect to each section B is created. Therefore, the similarity map M is such that, for example, the section B (for example, the section B with the identifier “16” in FIG. 4) having a lower rank order than the other sections B is excluded from the cluster in step S9 in FIG. It is possible to improve the accuracy of classification of each section B by referring to.

  Note that the method of calculating the cross-correlation value Cor in the first embodiment is changed as appropriate. For example, an addition value (or weighted sum) of cross-correlation values in each of a plurality of frequency bands obtained by dividing the average power spectrum on the frequency axis may be calculated as the cross-correlation value Cor. That is, the index calculation unit 43 calculates the cross-correlation value Cor_a for a specific band in the average power spectrum (SPa, SPb) and calculates the cross-correlation value Cor_b for another band. The weighted sum is calculated as a cross-correlation value Cor (Cor = α · Cor_a + β · Cor_b: α and β are constants). According to the above configuration, it is possible to reflect the characteristics of the band in which the difference for each speaker in the average power spectrum is particularly remarkable in the cross-correlation value Cor in a detailed and effective manner.

<B: Second Embodiment>
Next, a second embodiment of the present invention will be described. In the first embodiment, the cross correlation value Cor of the average power spectrum is exemplified as the similarity index value of each section B. On the other hand, in this embodiment, a result (average likelihood) obtained by collating the mixed model expressing the audio signal S of each section B with the feature amount of each other section B is employed as the similarity index value. In addition, about the element which an effect | action and function are the same as that of 1st Embodiment in each following form, the same code | symbol as FIG. 1 is attached | subjected and each detailed description is abbreviate | omitted suitably.

  In step S2 of FIG. 3, the feature extraction unit 41 performs frequency analysis on the audio signal S in the section B selected in step S1, and a MFCC (Mel Frequency Cepstral Coefficient) vector (hereinafter referred to as MFCC) in each frame in the section B. A time series of x) (referred to as “feature vector”) is extracted as a feature quantity. Further, in step S2, the feature extraction unit 41 generates a mixed model λ that models the distribution of the plurality of feature vectors x in the section B as a weighted sum of M normal distributions (M is a natural number). For the generation of the mixed model λ, a known technique such as an EM (Expectation-Maximization) algorithm is arbitrarily employed. By repeating the above process N times, the time series of the feature vector x and the mixed model λ are specified for each of the N sections B defined by the speech classification unit 12.

The mixed model λ is expressed by, for example, the following formula (2).
λ = {pi, μi, Σi} (i = 1 to M) (2)
In the equation (2), pi is a weight value (weight value) of the i-th normal distribution. The sum of the weights p1 to pM is 1. In the equation (2), 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 (2), this specification means that the symbol means a vector, for example, by clearly expressing the expression `` average vector ''. The vector symbol (the arrow pointing right on the character) is omitted.

  In step S5 of FIG. 3, the index calculation unit 43 converts the mixed model λ of the selected section B selected in step S4 and the time series of the feature vectors x extracted from the (N−1) comparison sections B. Based on this, the average likelihood L is calculated as the similarity index value for each of the (N−1) comparison sections B. As will be described in detail below, the average likelihood L is a probability (likelihood) that a feature vector x in another comparison section B appears from the mixed model λ in the selection section B, and a plurality of features in the comparison section B. It is a numerical value averaged for the vector x.

Assuming that one feature vector x is a D-dimensional vector, the likelihood that the feature vector x appears from the mixed model λ is calculated by the following equation (3).

The index calculation unit 43 substitutes the K feature vectors x (x1 to xK) extracted by the feature extraction unit 41 for one contrast section B other than the selected section B into the equation (4), thereby calculating the average likelihood L ( The average value of the probability that the feature vectors x1 to xK in the comparison section B appear from the mixed model λ in the selection section B) is calculated.

  As understood from the above description, the average likelihood L increases as the feature vector x is similar between the audio signal S in the selected section B and the audio signal S in the comparison section B. Therefore, as in the first embodiment, the index calculation unit 43 assigns each identifier of the (N−1) comparison sections B in order of decreasing average likelihood L (that is, similarity) in step S6 of FIG. Sort in ascending order). By calculating the average likelihood L (step S5) and sorting the identifiers (step S6) N times, a similarity map M similar to that in the first embodiment is completed. The operation (step S8 and step S9) of the voice classification unit 45 is the same as that in the first embodiment. In this embodiment, the same effect as that of the first embodiment is obtained.

<C: Third Embodiment>
A third embodiment of the present invention will be described. In this embodiment, the result (VQ (Vector Quantization) distortion) obtained by comparing the codebook when the speech signal S of each section B is vector quantized with the feature quantity of each other section B is adopted as the similarity index value. To do.

In step S2 in FIG. 3, the feature extraction unit 41 extracts the time series of the feature vector x (for example, MFCC) as the feature amount for the audio signal S in the section B selected in step S1 by the same method as in the second embodiment. , to create a codebook C ^a from the time series of a plurality of feature vector x in the section B. For vector quantization of the feature vector x, a known technique such as a k-means method or an LBG algorithm is arbitrarily employed. By step S2 it is repeated over N times, and the time series and the codebook C ^A feature vector x for each of the N phases B voice classification unit 12 is defined is specified.

In step S5 in FIG. 3, the index calculator 43, based on the codebook C ^A selection section B selected in step S4, the feature vector x (N-1) pieces of the contrast period B, (N -1) VQ distortion D is calculated as an similarity index value for each of the comparison sections B. The VQ distortion D is calculated by, for example, the following formula (5).

| C ^A | in Equation (5) is the size of the code book C ^{A in} the selected section B, 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 contrast period n _B number extracted from B (the number of frames in contrast period _{B) (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 of the selected section B and the feature vector x of the contrast section B, and n _B feature vectors x 1 to is a value obtained by averaging over xn _B. Therefore, the VQ distortion D decreases as the feature vector x is similar between the audio signal S in the selected section B and the audio signal S in the comparison section B.

  In step S6 of FIG. 3, the index calculation unit 43 sorts the identifiers of the (N-1) comparison sections B in the order in which the VQ distortion D is small (that is, the order of high similarity). By calculating the VQ distortion D (step S5) and sorting the identifiers (step S6) N times, a similarity map M similar to that of the first embodiment is completed. The operation of the voice classification unit 45 is the same as that of the first embodiment. In this embodiment, the same effect as that of the first embodiment is obtained.

<D: 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 forms, all of the N sections B into which the speech signal S is divided are classified, but the N sections N are classified into sound generation sections and non-sound generation sections (only noise in the environment where the sound is recorded is recorded). Existing sections), and only the pronunciation sections may be classified. For example, the voice classifying unit 12 excludes, from among the N sections B, the section B whose peak value is lower than the threshold as a non-sounding section from the classification target.

(2) Modification 2
Although the use of the threshold value for classification of the section B is not necessary according to each of the above forms, the configuration using the threshold value for the classification of the section B is not intended to exclude from the scope of the present invention. For example, when the section B that is most similar to one section B (the section B in which the identifier is at the highest position in the similarity map M) is selected in step S8 in FIG. Only when the index value exceeds the threshold value (when the similarity is high), both sections B are classified into the same cluster, and when the similarity index value falls below the threshold value, they are not classified into the same cluster. Also in this modification, when compared with the conventional configuration in which each section B is classified based only on the result of comparing the similarity index value and the threshold, the conditions (for example, the magnitude of noise) at the time of recording the audio signal S are different. The influence on the accuracy of the classification of the section B is reduced.

(3) Modification 3
The feature amount extracted by the feature extraction unit 41 is not limited to the above examples. For example, in the first embodiment, the feature extraction unit 41 may extract the average of the MFCC extracted from each frame in the section B in the section B as a feature amount instead of the average power spectrum. In the second embodiment or the third embodiment, the average value, maximum value, or fundamental frequency of the intensity of the audio signal S in the section B may be calculated as the feature amount.

(4) Modification 4
In each of the above embodiments, step S9 may be executed prior to step S8 in FIG. That is, the speech classification unit 45 classifies the section B having a similarity lower than a predetermined value among (N−1) sections B other than itself among the N sections B into a single cluster ( In step S9), the classification in step S8 is executed for a section B other than the section B (that is, a section B whose similarity with any other section B exceeds a predetermined rank).

(5) Modification 5
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 classification processing unit 14 can be output. For example, according to the structure which outputs the audio | voice signal S in the area B containing the time which the user specified among the some area B which the audio | voice classification | category part 12 classified as a sound wave from a sound emission apparatus (for example, speaker), a user However, it is possible to effectively support the task of creating the minutes of the meeting while selectively listening to the statements of each speaker and confirming them appropriately. Moreover, although the audio | voice classification part 12 illustrated the structure which divides the audio | voice signal S into the some area B in the above form, it stores in the memory | storage device 20 in the state by which the audio | voice signal S was divided into the some area B in advance. May be. As described above, the voice classification unit 12 and the voice recognition unit 16 are not essential elements for the voice processing apparatus 100.

(6) Modification 6
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.

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 which shows the content of operation | movement of an audio | voice classification | category part. It is a flowchart which shows the content of operation | movement of an audio | voice classification | category part. It is a conceptual diagram which shows the content of a similarity map.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Voice processing apparatus, 10 ... Control apparatus, 12 ... Voice classification part, 14 ... Classification processing part, 41 ... Feature extraction part, 43 ... Index calculation part, 45 ... Voice classification part, 16 ... ... voice recognition unit, 20 ... storage device, 30 ... output device, S ... voice signal, B ... section.

Claims (7)

Feature extraction means for extracting a feature amount for each of a plurality of sections obtained by dividing an audio signal on a time axis;
An index calculating means for calculating an similarity index value indicating similarity of the feature quantity in the two sections for a plurality of combinations for selecting two sections from the plurality of sections;
Voice classification for classifying the plurality of sections into a plurality of sets based on similarity index values of the sections so that each of the plurality of sections and a section having the most similar feature amount to the section belong to the same set And a voice processing apparatus. The speech processing apparatus according to claim 1, wherein the index calculation unit calculates a cross-correlation value between feature quantities of the two sections as the similarity index value. The feature extraction means extracts, as the feature amount, a time series of feature vectors of the voice signal for each of the plurality of sections.
The index calculating means calculates an average value of the likelihood that each feature vector of the other section appears from a mixed model obtained by modeling the distribution of feature vectors of one section of the two sections with a plurality of probability distributions, The speech processing apparatus according to claim 1, wherein the speech processing apparatus calculates the similarity index value. The feature extraction means extracts the feature amount from a time series of feature vectors of the audio signal for each of the plurality of sections,
The index calculating means calculates an average value of vector quantization distortion of a codebook obtained by vector quantization of a feature vector time series of one of the two sections and each feature vector of the other section. The speech processing device according to claim 1, wherein the speech processing device is calculated as an index value. The voice classification means includes
The sections in which the similarity index values for each of all the other sections are dissimilar among the plurality of sections are classified into a set different from all the other sections. Any of the voice processing devices. The audio processing device according to any one of claims 1 to 5, further comprising: audio classification means for dividing the audio signal into the plurality of sections with each valley portion in the envelope of the waveform of the audio signal as a boundary. On the computer,
A feature extraction process for extracting a feature amount for each of a plurality of sections obtained by dividing an audio signal on a time axis;
An index calculation process for calculating an similarity index value indicating similarity of the feature quantity in the two sections, for a plurality of combinations for selecting two sections from the plurality of sections;
Voice classification for classifying the plurality of sections into a plurality of sets based on similarity index values of the sections such that each of the plurality of sections and a section having the most similar feature amount to the section belong to the same set. A program that executes processing and.

Images