JP5272141B2 - Voice processing apparatus and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a featured value capable of highly accurately recognizing a speaker. <P>SOLUTION: A feature extraction section 32 calculates an autocorrelation sequence AV of frequency components of a voice signal V exceeding a predetermined frequency fc as a featured value FV. The autocorrelation sequence AV is a sequence (time sequence) of an autocorrelation value a(m) of a voice signal V. A storage section 24 stores an autocorrelation sequence AREF for reference. A speaker recognition section 34 performs speaker recognition by comparing the autocorrelation sequence AV of the voice signal V with the autocorrelation sequence AREF for reference. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a technique for processing an audio signal.

  Conventionally, speaker recognition techniques using feature amounts extracted from speech signals have been proposed. For example, in Patent Document 1, each section is classified for each speaker by using an average power spectrum or MFCC (mel-frequency cepstral coefficient) extracted from each of a plurality of sections obtained by dividing a voice signal on a time axis. Classification techniques have been proposed.

JP 2009-20458 A

  However, the configuration using the average power spectrum and MFCC may not be able to realize sufficiently accurate speaker recognition. Therefore, there has been a strong demand and expectation for a new feature amount that can realize speaker recognition with high accuracy. In view of the above circumstances, an object of the present invention is to propose a feature amount capable of realizing highly accurate speaker recognition.

  In order to solve the above problems, the speech processing apparatus according to the first aspect of the present invention executes feature recognition using feature extraction means for calculating an autocorrelation sequence of speech signals and speaker recognition using the autocorrelation sequence as a feature quantity. Speaker recognition means. In the above embodiment, since the autocorrelation sequence of the speech signal is used as a feature amount for speaker recognition, it is possible to realize speaker recognition with high accuracy.

  There is a tendency that fluctuations in the autocorrelation sequence corresponding to the content of the utterance are suppressed as the higher frequency component of the audio signal. Therefore, from the viewpoint of realizing highly accurate speaker recognition that is not affected by the content of the utterance, a configuration that calculates an autocorrelation sequence for a component that exceeds a predetermined frequency in the speech signal is particularly suitable.

  By the way, for the determination of the similarity (presence / absence of correlation) of a plurality of audio signals, for example, a correlation coefficient (cross-correlation) indicating the correlation between the feature amounts of each audio signal is used. However, the difference in intensity in the distribution (spectrum, time waveform, etc.) represented by the numerical sequence that defines the feature quantity is not necessarily reflected in the correlation coefficient. For example, the correlation coefficient of a plurality of sine waves having the same frequency but different amplitudes has the same numerical value as the correlation coefficient of a plurality of sine waves having the same frequency and amplitude (that is, the waveform perfectly matches) ( Maximum value). Therefore, the speaker may not be distinguished with high accuracy.

  In order to solve the above-described problem, a speech processing apparatus according to the second aspect of the present invention includes feature extraction means for extracting feature values represented by a plurality of numerical sequences from a speech signal, and a plurality of numerical sequences. Storage means for storing the reference feature quantity to be represented, and component addition means for adding auxiliary components including different numerical values to corresponding positions in the feature quantity extracted by the feature extraction means and the reference feature quantity, respectively. And an index calculation means for calculating the similarity index value indicating the similarity of each feature quantity after the addition of the auxiliary component, and a recognition processing means for executing speaker recognition using the similarity index value. In the above aspect, since the feature value of the audio signal and the feature value for reference are compared after the addition of the common auxiliary component, each numerical value series of the feature value of the audio signal is represented by each value of the reference feature value. Even when a relationship that is a constant multiple of the sequence is established, the difference between the feature value of the audio signal and the reference feature value can be manifested by the similarity index value. Therefore, highly accurate speaker recognition is possible.

  In the technique of Patent Document 1, since the similarity of the feature amount is determined for a plurality of combinations for selecting two sections from a plurality of sections into which the audio signal is divided, each section of the audio signal is determined in real time. It is difficult to classify by speaker. Therefore, the speech processing apparatus according to the third aspect of the present invention includes speech classification means for dividing a speech signal into a plurality of sections, feature extraction means for extracting feature quantities for each section of the speech signal, and reference features. Storage means for storing the amount, and speaker recognition means for classifying each of the plurality of sections into a set for each speaker by using the feature amount of each section. About the calculated feature quantity of one section, the similarity index value indicating similarity with the reference feature quantity stored in the storage means, and the feature quantity corresponding to one or more sections classified into the existing set Index calculation means for calculating the similarity index value indicating similarity, and when the feature quantity of one section is similar to the reference feature quantity, classify the one section into a new set, and set the existing set A recognition processing means for classifying one section into an existing set when the feature amount is similar to No. In the above aspect, each section is classified into a set for each speaker by comparing the feature quantity of one section into which the speech signal is divided with the reference feature quantity and the feature quantity of the existing set. All sections of the audio signal are not necessary for classification of each section. Therefore, there is an advantage that each section of the audio signal can be classified in real time (that is, the section can be classified into any set whenever each section of the audio signal is supplied).

  In the specific example of the speech processing apparatus according to the third aspect, the recognition processing means determines whether the feature quantity of one section is similar to the feature quantity of the existing set even when the feature quantity of one section is similar to the reference feature quantity. When the index value is a numerical value on the similar side with respect to the predetermined threshold, one section is classified into the existing set. According to the above aspect, there exists an advantage that possibility that the several area | region where a speaker is in common will be classified into a separate set is reduced. In addition, the recognition processing means may determine whether the similarity index value with the feature value of the existing set is not equal to the predetermined threshold value even if the feature value of one section is similar to the feature value of the existing set. If it is a numerical value on the similar side, one section is classified into a new set. According to the above aspect, there is an advantage that the possibility that a plurality of sections with different speakers are classified into a common set is reduced.

  In addition, when “similarity index value is a numerical value on the similar side with respect to a predetermined threshold value”, the similarity index value defined so as to increase as each feature quantity is similar exceeds the predetermined threshold value And the case where the similarity index value defined so as to decrease as the feature amounts become similar is less than a predetermined threshold value. Similarly, when “similarity index value is a numerical value on a dissimilar side with respect to a predetermined threshold value”, the similarity index value defined so as to increase as each feature amount is similar to the predetermined threshold value. The case where it falls below and the case where the similarity index value defined so that each feature-value decreases like it exceed a predetermined threshold are included.

  Two or more aspects selected from the first aspect, the second aspect, and the third aspect are arbitrarily merged. For example, the feature quantities in the second and third aspects are essentially arbitrary, but the autocorrelation sequence of the first aspect may be adopted as the feature quantities in the second and third aspects. Is possible. Moreover, the structure which added the component addition means in the 2nd aspect to the 1st aspect or the 3rd aspect is also employ | adopted.

  “Speaker recognition” in this application refers to speaker authentication (speaker verification) for determining whether or not a voice speaker of a voice signal corresponds to a regular registrant, and a voice speaker of a voice signal. It is a concept that includes speaker identification to be identified. For speaker identification, the process of determining which of the plurality of registrants corresponds to the voice speaker of the voice signal, and which speaker is in each segment of the voice recorded in the situation where there are multiple speakers And a process of determining whether the voice corresponds to the voice (further, a process of classifying each section for each speaker). The present invention (the first to third aspects) can be applied to any process included in the concept of speaker recognition defined as described above. However, this is not intended to exclude the possibility of applying the present invention to processing other than speaker recognition.

  The sound processing apparatus according to each of the above aspects is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to processing of a sound signal, or a general-purpose operation such as a CPU (Central Processing Unit). This is also realized by cooperation between the processing device and the program. For example, the program according to the first aspect of the present invention causes a computer to execute a feature extraction process for calculating an autocorrelation sequence of speech signals and a speaker recognition process using the autocorrelation sequence as a feature quantity.

  The program according to the second aspect includes a feature extraction process for extracting feature amounts represented by a plurality of numerical value sequences from a speech signal, a feature amount extracted by the feature extraction processing, and a reference value represented by a plurality of numerical value sequences. Component addition processing for adding auxiliary components including different numerical values to corresponding positions in each feature amount, and index calculation processing for calculating similarity index values indicating similarity of each feature amount after the addition of auxiliary components And a speaker recognition process using the similarity index value. Further, the program according to the third aspect uses an audio classification process for dividing an audio signal into a plurality of sections, a feature extraction process for extracting a feature quantity for each section of the audio signal, and a feature quantity of each section. A program for causing a computer to execute speaker recognition processing for classifying each of a plurality of sections into a set for each speaker, and the speaker recognition processing refers to the feature amount of one section calculated by the feature extraction processing. An index calculation process for calculating an similarity index value indicating similarity with a feature quantity for use and an similarity index value indicating similarity with a feature quantity corresponding to one or more sections classified into an existing set; When the feature value of one section is similar to the reference feature value, the one section is classified into a new set, and when the feature value is similar to the feature value of the existing set, the one section is set to the existing set. And recognition processing to classify.

  According to the program according to each aspect described above, the same operations and effects as the sound processing apparatus according to each aspect of the present invention are exhibited. The program of the present invention is provided to a user in a form stored in a computer-readable recording medium and installed in the computer, or provided from a server device in a form of distribution via a communication network and installed in the computer. Is done.

1 is a block diagram of a speech processing apparatus according to a first embodiment of the present invention. It is a graph which shows the autocorrelation sequence of an audio | voice signal (low-frequency component) for every speaker. It is a graph which shows the autocorrelation sequence of an audio | voice signal (high-frequency side component) for every speaker. It is a block diagram of a feature extraction unit. It is a conceptual diagram for demonstrating the relationship between two types of waveforms and a correlation coefficient. It is a conceptual diagram for demonstrating the correlation coefficient at the time of adding an auxiliary component to two types of waveforms. It is a block diagram of the speech processing unit concerning a 2nd embodiment. It is a block diagram of the speech processing unit concerning a 3rd embodiment. It is a conceptual diagram for demonstrating the division | segmentation of an audio | voice signal. It is a flowchart of operation | movement of a speaker recognition part.

<A: First Embodiment>
FIG. 1 is a block diagram of a speech processing apparatus 100A according to the first embodiment of the present invention. As shown in FIG. 1, a signal supply device 12 and an output device 14 are connected to the sound processing device 100A. The signal supply device 12 supplies an audio signal V representing a waveform on the time axis of the audio to the audio processing device 100A. For example, a sound collection device that collects ambient sounds and generates an audio signal V, a playback device that acquires the audio signal V from various recording media, and a communication device that receives the audio signal V from a communication network are Used as the supply device 12.

  The voice processing device 100A is a device (speaker recognition device) that performs speaker recognition using the voice signal V. Specifically, the speaker recognition performed by the speech processing apparatus 100A is speaker identification for determining which of a plurality of registrants the speaker of the speech of the speech signal V corresponds to. The output device 14 outputs the result of speaker recognition (speaker identification) by the sound processing device 100A as an image or sound. For example, a display device, a printing device, or a sound emitting device (speaker or headphones) is used as the output device 14.

  As shown in FIG. 1, the sound processing device 100 </ b> A is a computer system that includes an arithmetic processing device 22 and a storage device 24. The storage device 24 stores a program 26 executed by the arithmetic processing device 22 and data used by the arithmetic processing device 22. A known recording medium such as a semiconductor recording medium or a magnetic recording medium is arbitrarily employed as the storage device 24.

  The arithmetic processing unit 22 executes a program 26 stored in the storage device 24, thereby realizing a plurality of functions (feature extraction unit 32, speaker recognition unit 34) for executing speaker recognition of the audio signal V. To do. It should be noted that a configuration in which each element of the arithmetic processing unit 22 is distributedly mounted on a plurality of devices (integrated circuits) or a configuration in which an electronic circuit (DSP) dedicated to processing the audio signal V realizes each element is also employed. The

  The feature extraction unit 32 in FIG. 1 extracts the feature amount FV of the audio signal V. Specifically, the feature extraction unit 32 generates the autocorrelation sequence AV of the audio signal V as the feature amount FV. The autocorrelation sequence AV corresponds to a series (time series) of autocorrelation values a (m) of the audio signal V. The symbol m is an integer indicating the movement amount (time difference) on the own time axis with respect to the audio signal V.

  FIGS. 2 and 3 are graphs illustrating the autocorrelation sequence AV (autocorrelation function) of the voice signal V collected on a trial basis for each of the three speakers S (SA to SC). FIG. 2 shows the autocorrelation sequence AV of the low frequency component (1.5 kHz or less) of the audio signal V, and FIG. 3 shows the self component of the high frequency component (2 kHz to 7.8 kHz) of the audio signal V. A correlation sequence AV is shown. 2 and 3, the horizontal axis represents the movement amount of the audio signal V on the time axis (time difference from itself), and the vertical axis represents the autocorrelation value a (m). In the graph of each speaker S, an autocorrelation sequence AV of a plurality of speech signals V having different utterance contents is written together.

  As can be understood from FIGS. 2 and 3, the autocorrelation sequence AV has characteristics that are specific to the speaker and independent of the content of the utterance. Therefore, the autocorrelation sequence AV can be used as a feature amount for speaker recognition of the voice signal V. However, the low-frequency component illustrated in FIG. 2 is more likely to vary the autocorrelation value a (m) due to the content of the utterance than the high-frequency component illustrated in FIG. For high-accuracy speaker recognition, features that are independent of the content of the utterance (for example, features that reflect the resonance characteristics of the vocal tract of the speaker) are required. The autocorrelation sequence AV having a component exceeding a predetermined frequency in the audio signal V is particularly suitable.

  FIG. 4 is a specific block diagram of the feature extraction unit 32 of FIG. 4 is a filter (high-pass filter) that suppresses a component of the audio signal V that is below a predetermined frequency fc. The fluctuation of the autocorrelation value a (m) due to the content of the utterance is suppressed to such an extent that the desired accuracy is realized by the speaker recognition for the processed speech signal V by the low frequency suppressing unit 51. Thus, the frequency fc is selected experimentally or statistically. Specifically, a numerical value within the range of 1.5 kHz or more and 2.0 kHz or less is suitable as the frequency fc.

The time-frequency conversion unit 52 generates a frequency spectrum Q for each of a plurality of frames obtained by dividing the audio signal V (for example, a component of 2.0 kHz to 7.8 kHz) processed by the low frequency suppressing unit 51 on the time axis. For the generation of the frequency spectrum Q, a known technique such as fast Fourier transform is arbitrarily adopted. The power calculator 53 calculates the square of the absolute value of the frequency spectrum (amplitude spectrum) Q as the power spectrum | Q | ² . The averaging unit 54 generates an average power spectrum (average frequency characteristic) P by averaging (or adding) the power spectrum | Q | ² calculated by the power calculating unit 53 for a plurality of frames. The number and position of the frames of the power spectrum | Q | ² used for calculating the average power spectrum P are arbitrary.

The frequency-time conversion unit 55 performs an inverse Fourier transform on the average power spectrum P generated by the average unit 54. From the Wiener-Khintchine theorem, the numerical sequence in the time domain obtained by performing the inverse Fourier transform on the average power spectrum P corresponds to the autocorrelation sequence AV. Specifically, the frequency-time conversion unit 55 calculates each autocorrelation value a (m) by the calculation (inverse Fourier transform) of the following formula (1). Note that the symbol k in Equation (1) is an integer that specifies one of a plurality of frequencies (frequency bins) discretely set on the frequency axis, and the symbol p (k) in Equation (1) is: It means the intensity (power) at the frequency indicated by the symbol k in the average power spectrum P.

  As shown in FIG. 1, in the storage device 24, a reference feature quantity FREF (dictionary) of the same type as the feature quantity FV is stored in advance for each voice of a plurality of different registrants. Specifically, the autocorrelation sequence AREF generated for each registrant in the same manner as the autocorrelation sequence AV is stored in the storage device 24 as the feature amount FREF. Note that the feature extraction unit 32 can be used to generate the autocorrelation sequence AREF. However, a method of generating the autocorrelation sequence AREF for each registrant by a device separate from the speech processing device 100A and storing it in the storage device 24 is also employed. The number of autocorrelation values a (m) constituting the autocorrelation sequence AV is common to the number of autocorrelation values a (m) constituting the reference autocorrelation sequence AREF.

  The speaker recognition unit 34 compares the feature amount FV (autocorrelation sequence AV) extracted by the feature extraction unit 32 with each feature amount FREF (autocorrelation sequence AREF) stored in the storage device 24 to recognize the speaker. Execute. As shown in FIG. 1, the speaker recognition unit 34 is configured to include an index calculation unit 42 and a recognition processing unit 44. The index calculation unit 42 stores the similarity index value R indicating the similarity (correlation between the feature amount FV and the feature amount FREF) between the autocorrelation sequence AV of the audio signal V and the reference autocorrelation sequence AREF. Is calculated for each of the plurality of autocorrelation sequences AREF stored in. Specifically, the speaker recognition unit 34 calculates a correlation coefficient Cor defined by the following formula (2) as the similarity index value R.

相関係数Ｃorは、数値ｄ1(i)の系列と数値ｄ2(i)の系列との相関を示す数値である（ｉは整数）。数式(2)の記号ｄ1_aveは、複数の数値ｄ1(i)の平均値を意味し、数式(2)の記号ｄ2_aveは、複数の数値ｄ2(i)の平均値を意味する。数式(2)から理解されるように、数値ｄ1(i)の系列と数値ｄ2(i)の系列とが類似するほど（すなわち各系列が示す波形の相関が高いほど）、相関係数Ｃorは大きい数値となり、両者が完全に合致する場合に最大値「１」となる。

The correlation coefficient Cor is a numerical value indicating the correlation between the series of numerical values d1 (i) and the numerical value d2 (i) (i is an integer). The symbol d1_ave in Equation (2) means the average value of the plurality of numerical values d1 (i), and the symbol d2_ave in Equation (2) means the average value of the plurality of numerical values d2 (i). As understood from the equation (2), the more similar the series of the numerical value d1 (i) and the series of the numerical value d2 (i) (that is, the higher the correlation of the waveform shown by each series), the correlation coefficient Cor becomes It becomes a large numerical value, and when the two match completely, the maximum value is “1”.

  The index calculation unit 42 in FIG. 1 substitutes each autocorrelation value a (m) of the autocorrelation sequence AV for each numerical value d1 (i) of the formula (2), and each autocorrelation value a (m) of the autocorrelation sequence AREF. ) Is substituted for each numerical value d2 (i) in the equation (2), and the correlation coefficient Cor is calculated as the similarity index value R. Accordingly, the similarity index value R becomes a larger numerical value as the autocorrelation sequence AV and the autocorrelation sequence AREF are more similar.

  The recognition processing unit 44 in FIG. 1 executes speaker identification using the similarity index value R calculated by the index calculation unit 42 for each autocorrelation sequence AREF (for each registrant). Specifically, the recognition processing unit 44 identifies a registrant having a maximum similarity index value R between the autocorrelation sequence AV and the autocorrelation sequence AREF among a plurality of registrants. That is, the voice utterer (unknown) of the voice signal V is identified from a plurality of registrants. The result of identification by the recognition processing unit 44 is output from the output device 14.

  In the above embodiment, since the autocorrelation sequence AV of the audio signal V is used as the feature amount FV for speaker recognition, it is possible to realize speaker recognition with high accuracy. In addition, since the autocorrelation sequence AV of the component of the voice signal V that exceeds the frequency fc (that is, the component in which the autocorrelation value a (m) varies less due to the content of the utterance) is applied to speaker recognition, A special effect is realized that speaker recognition with high accuracy is possible regardless of the content of the.

<B: Second Embodiment>
Next, a second embodiment of the present invention will be described. In the second embodiment, the target for calculating the similarity index value R is different from that of the first embodiment. In addition, about the element in which an effect | action and a function 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.

  5 and 6 are conceptual diagrams for explaining the principle of the method for calculating the similarity index value R in the second embodiment. FIG. 5 shows two types of waveforms (WA, WB) having the same frequency but different amplitudes. Correlation coefficient Cor calculated by equation (2) from a numerical sequence representing the waveform WA (sequence of numerical values d1 (i)) and a numerical sequence representing the waveform WB (sequence of numerical values d2 (i)) is the waveform WA Although the amplitude is different from that of the waveform WB, the maximum value “1” means that the two match. Therefore, the waveform WA and the waveform WB cannot be distinguished from the correlation coefficient Cor.

  On the other hand, FIG. 6 shows a case where a common component (hereinafter referred to as “auxiliary component”) Wp is added to the corresponding position in each of the waveform WA and the waveform WB. In the example of FIG. 6, the auxiliary component Wp is added immediately after each of the waveform WA and the waveform WB. The auxiliary component Wp is a component whose intensity varies (non-DC component). Specifically, a waveform component (for example, a sine wave component) whose intensity varies periodically is suitably employed as the auxiliary component Wp. FIG. 6 illustrates the case where the amplitude of the auxiliary component Wp exceeds the amplitude of the waveform WA and is lower than the amplitude of the waveform WB, and the frequency of the auxiliary component Wp exceeds the frequencies of the waveform WA and the waveform WB. The amplitude and frequency of the auxiliary component Wp are arbitrarily changed.

  A numerical sequence (a series of numerical values d1 (i)) representing the waveform WAp with the auxiliary component Wp added to the waveform WA, and a numerical sequence (a series of numerical values d2 (i)) representing the waveform WBp with the auxiliary component Wp added to the waveform WB. Therefore, the correlation coefficient Cor calculated by the equation (2) is a numerical value (for example, 0.9) that is less than the maximum value “1”. That is, by adding the common auxiliary component Wp, the difference between the waveform WA and the waveform WB (difference in amplitude) can be manifested by the correlation coefficient Cor. On the other hand, if the waveform WA and the waveform WB completely match, the waveform naturally matches even after the auxiliary component Wp is added, so the correlation coefficient Cor becomes the maximum value. In the second embodiment, the difference between the feature amount FV (autocorrelation sequence AV) of the audio signal V and the reference feature amount FREF (autocorrelation sequence AREF) is made obvious by using the above principle.

  FIG. 7 is a block diagram of an audio processing device 100B according to the second embodiment. As shown in FIG. 7, the sound processing device 100B has a configuration in which a component addition unit 36 is added to the sound processing device 100A of the first embodiment. The component addition unit 36 adds a common auxiliary component Wp to the autocorrelation number sequence AV generated by the feature extraction unit 32 and each of the plurality of autocorrelation number sequences AREF stored in the storage device 24. The auxiliary component Wp is set as a series of a plurality of numerical values w including different numerical values (for example, a time series of sine wave intensity).

  The index calculation unit 42 in FIG. 7 stores a plurality of similarity index values R of the autocorrelation sequence AV added with the auxiliary component Wp and the reference autocorrelation sequence AREF added with the auxiliary component Wp stored in the storage device 24. Are calculated for each autocorrelation sequence AREF. Specifically, the index calculation unit 42 substitutes each autocorrelation value a (m) of the autocorrelation sequence AV and each numerical value w of the auxiliary component Wp into each numerical value d1 (i) of the equation (2), and By substituting each autocorrelation value a (m) of the autocorrelation sequence AREF and each value w of the auxiliary component Wp into each value d2 (i) of Equation (2), the autocorrelation sequence AV and the autocorrelation sequence AREF The similarity index value R (correlation coefficient Cor) is calculated. The method of speaker recognition by the recognition processing unit 44 is the same as in the first embodiment.

  In the second embodiment, speaker recognition is performed by comparing the autocorrelation sequence AV with the common auxiliary component Wp added thereto and the autocorrelation sequence AREF. Therefore, compared with the first embodiment without adding the auxiliary component Wp, The difference between the autocorrelation sequence AV and the autocorrelation sequence AREF can be made obvious. Therefore, there is an advantage that speaker recognition with higher accuracy than that of the first embodiment is realized.

<C: Third Embodiment>
Next, a third embodiment of the present invention will be described. Speaker identification, which is one aspect of speaker recognition, is a process for determining which of a plurality of registrants a voice speaker and each section of speech recorded in a situation where a plurality of speakers exist. Is roughly divided into processing for determining which speaker's voice corresponds to. In the first embodiment and the second embodiment, the former speaker identification is illustrated, but in the third embodiment, the latter speaker identification is illustrated. In the following, the third embodiment will be described based on the configuration of the first embodiment, but it is naturally possible to add the component addition unit 36 in the second embodiment to the third embodiment.

  FIG. 8 is a block diagram of a speech processing apparatus 100C according to the third embodiment. As shown in FIG. 8, the speech processing apparatus 100C is configured by adding a speech classification unit 38 to the speech processing apparatus 100A of the first embodiment and replacing the recognition processing unit 44 of the first embodiment with a recognition processing unit 46. It is. An audio signal V recorded in a situation where there are a plurality of speakers (for example, a conference where a plurality of participants exist) is supplied from the signal supply device 12 to the audio processing device 100C.

  The audio classification unit 38 divides the audio signal V into a plurality of sections (blocks) B on the time axis. Each section B is a period in which it is estimated that there is a high possibility that one speaker is continuously generated. Each section B is given a unique identifier Ib. The speech processing apparatus 100C determines which speaker's speech each section B of the speech signal V corresponds to.

  In a normal utterance (especially in a conference), 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, as shown in FIG. 9, the audio classification unit 38 divides the audio signal V into a plurality of sections B with each of a plurality of valleys D appearing in an envelope E of the waveform of the audio signal V as a boundary. . According to the above configuration, for example, when the last part of the utterance by one speaker overlaps with the first part of the utterance by another speaker, or when a plurality of speakers uttered sequentially without any interval. Even in this case, it is possible to divide the utterance by each speaker into separate sections B. However, the method of dividing the audio signal V into a plurality of sections B is arbitrary in the present invention.

  The feature extraction unit 32 calculates a feature amount FV (autocorrelation sequence AV) for each of the plurality of sections B of the audio signal V. On the other hand, the storage device 24 stores, as a dictionary, a feature quantity FREF (autocorrelation sequence AREF) of the same kind as the feature quantity FV for each of a plurality of representative voices (samples) of voice qualities. That is, in the first embodiment and the second embodiment, the feature amounts FREF of a plurality of registrants who are candidates for the speaker of the audio signal V are generated in advance and stored in the storage device 24. In the third embodiment, however, The feature value FREF of the speaker of the audio signal V is not stored in the storage device 24.

  The speaker recognition unit 34 in FIG. 8 classifies each of the plurality of sections B into a set (cluster) CLj (j is a natural number) for each speaker using the feature amount FV of each section B of the audio signal V. The classification of the section B is executed in real time whenever the feature extraction unit 32 calculates the feature amount FV of each section B. As shown in FIG. 8, a storage area Mj (M1, M2,...) For each set CLj is set in the storage device 24 according to the classification result by the speaker recognition unit 34. In the storage area Mj of the set CLj, the feature quantity corresponding to the identifier Ib of each section B classified into the set CLj, the audio signal V in the section B, and the feature quantity FV of each section B classified into the set CLj FC is stored. The feature value FC of the set CLj is, for example, an average value of the feature values FV of one or more sections B classified into the set CLj.

  FIG. 10 is a flowchart of the operation of the speaker recognition unit 34. As shown in FIG. 10, when the first section B of the audio signal V is acquired (step S1), the recognition processing unit 46 classifies the section B into a new set CL1 (step S2). That is, the recognition processing unit 46 stores the identifier Ib, the voice signal V, and the feature amount FV (feature amount FC) of the section B acquired in step S1 in the storage area M1 of the storage device 24.

  When the next section B (hereinafter referred to as “target section B”) is acquired (step S3), the index calculating unit 42 calculates the similarity index value R between the feature amount FV of the target section B and the reference feature amount FREF. Calculation is performed for each of the plurality of feature values FREF stored in the storage device 24 (step S4). Further, the index calculation unit 42 is a similarity index value between the feature amount FV of the target section B and the feature set FC of the existing set (that is, the set in which the recognition processing unit 46 classifies one or more sections B in the past) CLj. R is calculated (step S5).

  In step S5 immediately after the start of the process of FIG. 10, the first section B is only at the stage where it is classified into the set CL1, so the recognition processing unit 46 classifies only the feature quantity FC of the set CL1 with the feature quantity FV. A rejection index value R is calculated. On the other hand, at the stage where the process of FIG. 10 has progressed and a plurality of sets CLj has been generated, the recognition processing unit 46 generates a plurality of similarity index values R between the feature quantity FC of the set CLj and the feature quantity FV of the target section B. Each of the sets CLj is calculated in step S5. Note that the method of calculating the similarity index value R in step S4 and step S5 is the same as in the first embodiment. Also, the order of steps S4 and S5 can be reversed.

  The recognition processing unit 46 determines that the maximum value among the plurality of similarity index values R calculated in step S4 and step S5 is the similarity index value R (step S4) of the reference feature quantity FREF and the existing set. It is determined which of the CL feature quantity FV is the similarity index value R (step S5) (step S6).

  When the similarity index value R with the reference feature quantity FREF is the maximum value (S6: YES), the recognition processing unit 46 classifies the feature quantity FV of the target section B and the feature quantity FC of the existing set CLj. It is determined whether or not the maximum value Rmax1 in the rejection index value R (step S5) exceeds a predetermined threshold value RTH1 (that is, whether both are sufficiently similar) (step S7). When the result of step S7 is negative (that is, when the feature value FV of the target section B is not similar to the feature value FC of the existing set CLj), the recognition processing unit 46 changes the target section B to the new set CLj. Classify (step S10). That is, the recognition processing unit 46 stores the identifier Ib, the audio signal V, and the feature amount FV (feature amount FC) of the target section B in the storage area Mj of the set CLj.

  On the other hand, when the result of step S7 is affirmative (that is, when the feature value FV of the target section B is similar to the reference feature value FREF but sufficiently similar to the feature value FC of the existing set CLj). The recognition processing unit 46 classifies the target section B into the set CLj having the maximum similarity index value R calculated in step S5 among the existing set CLj (step S8). Specifically, the recognition processing unit 46 adds the identifier Ib and the audio signal V of the target section B to the storage area Mj of the existing set CLj, and uses the feature value FC of the set CLj as the feature of the target section B. The number is updated to a numerical value corresponding to the amount FV (for example, an average value of the feature amount FV of each section B and the feature amount FV of the target section B previously classified into the set CLj). As can be understood from the above description, the similarity index value between the two is only when the speaker in the target section B and the speaker in the section B in the existing set CLj can be determined with sufficiently high accuracy. The threshold value RTH1 applied in step S7 is set experimentally or statistically so that R exceeds the threshold value RTH1.

  When it is determined in step S6 that the similarity index value R with the feature value FC of the existing set CLj is the maximum value (S6: NO), the recognition processing unit 46 and the feature value FV of the target section B and the existing set It is determined whether or not the maximum value Rmax2 in the similarity index value R (step S5) with the feature quantity FC of CLj is below a predetermined threshold value RTH2 (that is, whether or not they are sufficiently different). Step S9). When the result of step S9 is negative (that is, when the feature value FV of the target section B is similar to the feature value FC of the set CLj), the recognition processing unit 46 calculates in step S5 of the existing set CLj. The target section B is classified into the set CLj having the maximum similarity index value R (step S8). The process in step S8 is as described above.

  On the other hand, if the result of step S9 is affirmative (that is, if the feature value FV of the target section B is not similar to either the reference feature value FREF or the feature value FC of the existing set CLj), the recognition processing unit 46 Classifies the target section B into a new set CLj (step S10). The process of step S10 is as described above.

  When the process of step S8 or step S10 is executed, the speaker recognition unit 34 determines whether or not to end the speaker recognition (step S11). When an instruction to end speaker recognition is given by the user or when classification of all sections B of the audio signal V is completed, the recognition processing unit 46 ends speaker recognition (S11: YES). On the other hand, when the speaker recognition is not terminated, the speaker recognition unit 34 moves the process to step S3, and executes the processes after step S4 with the section B to be acquired next as the new target section B. Therefore, among the plurality of sections B of the audio signal V, each section B having a similar feature amount FV (that is, a section B that can be determined that the speaker is common) is classified into a common set CLj. As can be understood from the above description, the similarity index value between the two is only when the speaker in the target section B and the speaker in the section B in the existing set CLj can be determined with sufficiently high accuracy. The threshold value RTH2 applied in step S9 is set experimentally or statistically so that R is below the threshold value RTH2.

  The classification result by the recognition processing unit 46 is output from the output device 14. For example, the minutes of the meeting are output as an image from the output device 14. In the minutes, for each section B of the voice signal V, the identifier of the set CLj (speaker identifier) into which the section B is classified, and the character string specified by the speech recognition of the voice signal V in the section B (That is, the content of a statement) are arranged in time series.

  In the above embodiment, each section B is determined for each speaker by comparing the feature quantity FV of one section B into which the audio signal V is divided with the reference feature quantity FREF and the feature quantity FC of the existing set CLj. Since it is classified into the set CLj, all the sections B of the audio signal V are not necessary for classification of one section B. Therefore, there is an advantage that each section B of the audio signal V can be classified in real time. In addition, without storing the reference feature quantity FREF in the storage device 24, the similarity index value R between the feature quantity FC of the existing set CLj and the feature quantity FV of the section B is compared with a threshold value, thereby comparing each section. A configuration (hereinafter referred to as “proportional”) in which B is classified for each speaker is also assumed. However, there is a problem that it is difficult to appropriately select a threshold value that is a criterion for determination of similarity under a comparative ratio. On the other hand, in the third embodiment, the classification of each section B is executed according to the size of the similarity index value R for the feature quantity FC of the existing set CLj and the similarity index value R for the reference feature quantity FREF. Therefore, there is an advantage that the setting of the threshold value in the proportionality does not become a problem.

  Furthermore, in the above embodiment, even if the feature quantity FV of the target section B is most similar to the reference feature quantity FREF, the feature quantity FV of the target section B and the feature quantity FC of the existing set CLj When the similarity index value R is so similar that it exceeds the threshold value RTH1, the target section B is classified into the set CLj. Therefore, there is an advantage that the possibility that a plurality of sections B uttered by a common speaker is classified into another set CLj is reduced. Even if the feature value FV of the target section B is most similar to the feature value FC of the existing set CLj, the similarity index value between the feature value FV of the target section B and the feature value FC of the existing set CLj If the two differ so that R falls below the threshold value RTH2, the target section B is not classified into the set CLj. Therefore, a plurality of sections B uttered by another speaker may be classified into the same set CLj. There is an advantage that it is reduced.

<D: Modification>
Each form illustrated above can be variously modified. Specific modes of deformation are exemplified below. Note that two or more aspects arbitrarily selected from the following examples may be appropriately combined.

(1) Modification 1
In the first embodiment, the autocorrelation sequence AV is calculated by inverse Fourier transform of the average power spectrum P. However, a configuration in which the autocorrelation sequence AV is calculated from the audio signal V by time domain calculation is also employed. However, according to the first embodiment in which the autocorrelation sequence AV is calculated by calculation in the frequency domain, there is an advantage that the amount of calculation by the feature extraction unit 32 is reduced. Further, the position of the low-frequency suppressing unit 51 is arbitrarily changed. For example, a configuration in which the low-frequency suppressing unit 51 is arranged after the time-frequency converting unit 52 to suppress the low-frequency component of the frequency spectrum Q is also employed. However, the low frequency suppression unit 51 is not essential in the present invention. That is, since a characteristic unique to the speaker appears in the autocorrelation sequence AV of the low frequency component of the audio signal V, a configuration in which the low frequency suppression unit 51 is omitted (that is, the entire band of the audio signal V is targeted). The autocorrelation sequence AV (feature value FV) that can be used for speaker recognition can be calculated.

(2) Modification 2
In each of the above embodiments, the autocorrelation sequence (AV, AREF) is exemplified as the feature quantity (FV, FREF, FC). However, the types of feature quantities (FV, FREF, FC) in the second and third embodiments are described. Is arbitrarily changed. For example, the average power spectrum P of speech (the average value of the square of the absolute value of the frequency spectrum Q) or the average (average cepstrum) over a plurality of frames of the cepstrum calculated from the frequency spectrum Q is used as the feature quantity (FV, FREF, FC ) Can be used.

  Further, the similarity index value R is not limited to the correlation coefficient Cor of the equation (2), and an appropriate similarity index value R corresponding to the type of the feature amount FV or the feature amount FREF is selected. For example, in the configuration using the average cepstrum as the feature quantity FV and the feature quantity FREF, the difference between the feature quantity FV and the feature quantity FREF (the difference between the feature quantity FV and the feature quantity FC in the third embodiment) is the similarity index. Calculated as the value R. In the configuration in which the average power spectrum P is used as the feature amount FV or the feature amount FREF, the correlation coefficient Cor in Expression (2) is used as the similarity index value. Also, depending on the type of feature quantity FV and feature quantity FREF, a configuration in which distance and likelihood are calculated as similarity index value R is also suitable.

  As described above, the definition of the similarity index value R is arbitrary. The relationship between the similarity index value R and the similarity between the feature quantity FV and the feature quantity FREF is determined according to the definition of the similarity index value R. That is, in each of the above embodiments, the similarity index value R is defined so that the similarity index value R becomes a larger numerical value as the feature quantity FV and the feature quantity FREF are similar, but the feature quantity FV and the feature quantity FREF are The similarity index value R is defined so that the similarity index value R becomes smaller as the similarity is similar (for example, a configuration in which the distance between the feature quantity FV and the feature quantity FREF is the similarity index value R) is also adopted. Is done.

(3) Modification 3
When the average power spectrum P is adopted as the feature amount (FV, FREF) in the second embodiment, the component adding unit 36 averages the average power spectrum PV (feature amount FV) of the voice signal V and the voice of each registrant. A common auxiliary component Wp is added to each of the power spectrum PREF (feature value FREF). The index calculation unit 42 substitutes the numerical value of the intensity (power) for each frequency in the average power spectrum PV and each numerical value of the auxiliary component Wp into each numerical value d1 (i) of Equation (2), and in the average power spectrum PREF. The correlation coefficient Cor when the numerical value of the intensity for each frequency and each numerical value of the auxiliary component Wp is substituted for each numerical value d2 (i) of the equation (2) is calculated as the similarity index value R.

  Further, the position where the auxiliary component Wp is added is appropriately changed. For example, the auxiliary component Wp may be added or inserted at the beginning or in the middle of the feature amount FV (autocorrelation sequence AV or average power spectrum PV) and feature amount FREF (autocorrelation sequence AREF or average power spectrum PREF). The effect similar to 2nd Embodiment is implement | achieved. That is, a configuration in which the auxiliary component Wp is added to the corresponding position (the same position in both) of the feature amount FV and the feature amount FREF is preferable in the present invention, but the auxiliary component Wp in the feature amount FV and the feature amount FREF is preferred. The specific position of is unquestioned. Furthermore, the waveform indicated by the auxiliary component Wp is also arbitrary. That is, regardless of the waveform indicated by the auxiliary component Wp, by adding the common auxiliary component Wp to the feature quantity FV and the feature quantity FREF, the same effect as in the second embodiment is realized.

  As can be understood from the above description, the component addition unit 36 in the second embodiment is a numerical sequence (autocorrelation sequence or average power spectrum) indicating each of the feature amount FV and the feature amount FREF extracted by the feature extraction unit 32. It is included as an element for adding a common auxiliary component Wp (typically a non-DC component) to a set of numerical values constituting the component.

(4) Modification 4
In each of the above embodiments, speaker identification has been exemplified. However, in the speech processing apparatus 100A of the first embodiment and the speech processing apparatus 100B of the second embodiment, the voice speaker of the voice signal V corresponds to a regular registrant. It is also used for speaker authentication (speaker verification) for determining whether or not to do so. For example, the feature value FREF (for example, autocorrelation sequence AREF) extracted from the voice of the regular registrant is stored in the storage device 24, and the index calculation unit 42 uses the feature value FV (for example, autocorrelation) extracted from the sound signal V. The similarity index value R between the numerical sequence AV) and the registrant's feature value FREF is calculated. The recognition processing unit 44 determines the legitimacy of the voicer of the voice signal V according to the similarity index value R. Specifically, when the similarity index value R exceeds a predetermined threshold value (when the feature value FV and the feature value FREF are similar), the recognition processing unit 44 authenticates the speaker and verifies the similarity. If the index value R is below the threshold, authentication is denied.

(5) Modification 5
In each of the above embodiments, one type of feature value FV is used. However, a configuration in which a combination of different types of feature values is used as the feature value FV (and also the feature value FREF) for speaker recognition is also suitable. For example, the feature extraction unit 32 extracts a combination of two or more feature amounts selected from the autocorrelation sequence AV, the average power spectrum P, and the average cepstrum as the feature amount FV. The index calculation unit 42 calculates the similarity index value with the reference feature quantity FREF for each feature quantity of the feature quantity FV, and calculates the weighted sum of the similarity index values of the feature quantities for the speaker recognition similarity index. Calculated as the value R. According to the above configuration, since various viewpoints (properties) of the voice are reflected in the determination of the similarity between the feature amount FV and the feature amount FREF, the accuracy is higher than in the case of using one type of feature amount. Advantageous speaker recognition. Further, since the weighted sum of the similarity index values of the feature quantities is used as the similarity index value R for speaker recognition, an operation of giving priority to a specific feature quantity over other feature quantities is possible.

100A, 100B, 100C ... speech processing device, 12 ... signal supply device, 14 ... output device, 22 ... calculation processing device, 24 ... storage device, 26 ... program, 32 ... feature extraction unit, 34 …… Speaker recognizing part 36 …… Component adding part 38 ??? Speech classification part 42 ??? Indicator calculating part 44 ...... Recognition processing part 46 ...... Recognition processing part 51 ...... Low frequency suppressing part 52... Time-frequency conversion unit, 53... Power calculation unit, 54... Average unit, 55.

Claims (7)

Feature extraction means for extracting a feature amount represented by a series of numerical values from an audio signal ;
Storage means for storing feature values for reference represented by a series of a plurality of numerical values;
Component addition means for adding a common auxiliary component including different numerical values to corresponding positions in each of the feature quantity extracted by the feature extraction means and the reference feature quantity;
An index calculating means for calculating an similarity index value indicating the similarity of each feature quantity after addition of the auxiliary component;
A speech processing apparatus comprising recognition processing means for performing speaker recognition using the similarity index value . The speech processing apparatus according to claim 1, wherein the feature extraction unit extracts a feature amount for a component that exceeds a predetermined frequency in the speech signal. Voice classification means for dividing the voice signal into a plurality of sections;
The feature extraction means extracts a feature amount for each section ,
The index calculation means calculates a similarity index value for each section,
The speech processing apparatus according to claim 1, wherein the recognition processing unit classifies each of the plurality of sections into a set for each speaker using the similarity index value of each section. The index calculation unit, the feature amounts of one section the feature extraction means is calculated, indicating the similarity between the feature quantity of a reference stored in said storage means is extracted from the audio samples representative voice Calculating the similarity index value and the similarity index value indicating similarity between the feature quantity corresponding to one or more sections classified into the existing set ;
If the recognition processing means, the feature quantity of the one section is the case similar to the feature quantity for the reference, which classifies the one segment to a new collection, similar to the feature amount of the existing set And classify the one section into the existing set
The speech processing apparatus according to claim 3 . Said recognition processing means, even if the characteristic quantity of the one section is similar to the feature quantity for the reference, similar against a threshold similarity index value is in a predetermined feature amount of the existing set If it is a numerical value on the side, classify the one section into the existing set
The speech processing apparatus according to claim 4 . The recognition processing means may determine whether the similarity index value with the feature value of the existing set is equal to a predetermined threshold value even when the feature value of the one section is similar to the feature value of the existing set. If the value is on the dissimilar side, classify the one section into a new set
The speech processing apparatus according to claim 4 or 5 . A feature extraction process for extracting a feature value represented by a series of numerical values from a speech signal ;
A common auxiliary component including different numerical values is added to the corresponding position in each of the characteristic amounts extracted in the characteristic extraction processing and the reference characteristic amounts represented by a plurality of numerical value sequences stored in the storage means. Component addition processing,
An index calculation process for calculating an similarity index value indicating the similarity of each feature quantity after the addition of the auxiliary component;
A program for causing a computer to execute speaker recognition processing using the similarity index value .

Images