JP2009175474A - Sound processing device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a sound processing device and a program capable of highly accurately determining a speech sound and a non-speech sound. <P>SOLUTION: A modulation spectrum specifying section 32 specifies a modulation spectrum MS of an input sound VIN for each of multiple units TU. An index calculation section 34 calculates an index value D1 according to an intensity degree L1 in which a modulation frequency is 10 Hz or smaller in the modulation spectrum MS. a storage device 24 stores a sound model M which is generated from a sound of a vowel. An index calculation section 54 calculates an index value D2 for indicating the similarity of the input sound VIN and the sound model M for each unit. A determination section 42 determines whether the input sound VIN of each unit TU is the speech sound or the non-speech sound based on the index value D1 and the index value D2 of the unit TU. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a technique for discriminating between human voices (hereinafter referred to as “speech”) and sounds other than speech (hereinafter referred to as “non-speech”).

Conventionally, a technique for distinguishing sound such as recorded sound by a sound collecting device (hereinafter referred to as “input sound”) into a voice section and a non-voice section has been proposed. For example, Patent Document 1 discloses a technique for determining the presence or absence of sound based on the intensity of a component belonging to a predetermined frequency band in an input sound.
JP 2000-132177 A

  However, noise characteristics (frequency) vary, and noise may occur in a frequency band used for determining the presence or absence of speech. Therefore, it is difficult to determine the presence or absence of speech with sufficiently high accuracy under the technique of Patent Document 1. In view of the above circumstances, an object of the present invention is to determine voice / non-voice with high accuracy.

  In order to solve the above problems, a sound processing apparatus according to the first aspect of the present invention includes a modulation spectrum specifying unit that specifies a modulation spectrum of an input sound for each of a plurality of unit sections, and a modulation among the modulation spectra. First index calculation means (for example, index calculation unit 34 in FIG. 10) that calculates a first index value corresponding to the intensity of a component whose frequency falls within a predetermined range, and a memory that stores an acoustic model generated from the vowel sound Means, second index calculation means (for example, the index calculation unit 54 in FIG. 10) for calculating, for each unit section, a second index value indicating similarity between the input sound and the acoustic model, Determination means for determining whether the input sound is speech or non-speech based on the first index value and the second index value of the unit section. In the above aspect, whether the input sound in each unit section is speech or non-speech based on the intensity of the component whose modulation frequency falls within a predetermined range in the modulation spectrum and the similarity of the input sound with respect to the acoustic model of the vowel. Therefore, it is possible to identify speech / non-speech with high accuracy compared to the technique of Patent Document 1 using the frequency spectrum of the input sound.

  The sound processing apparatus according to a preferred aspect of the present invention includes third index calculation means (for example, the index calculation unit 62 in FIG. 10) that calculates the weighted sum of the first index value and the second index value as the third index value. The determination means determines whether the input sound of each unit section is speech or non-speech based on the third index value of the unit section. In the above aspect, by appropriately selecting the weighted sum of the first index value and the second index value, which of the first index value and the second index value is to be prioritized over voice / non-speech determination. It is possible to set.

  In the configuration including the third index calculation means, further provided is a weight value setting means for variably setting the weight value applied by the third index calculation means to the calculation of the third index value according to the SN ratio of the input sound. Also good. For example, on the assumption that the first index value is more susceptible to the noise of the input sound as compared to the second index value, the weight value setting means has a lower second index value as the SN ratio of the input sound is lower. The weight value is increased relative to the weight value of the first index value (that is, the second index value is given priority). According to the above aspect, it is possible to determine the voice / non-voice of the input sound with high accuracy regardless of the noise of the input sound.

  The sound processing apparatus according to the second aspect of the present invention includes a modulation spectrum specifying unit that specifies a modulation spectrum of an input sound for each of a plurality of unit sections, and an intensity of a component whose modulation frequency is within a predetermined range of the modulation spectrum. 1st index calculation means (for example, the index calculation unit 34 in FIG. 2) that calculates a first index value according to the above, and determination to determine whether the input sound of each unit section is voice or non-speech based on the first index value Means. In the above aspect, since the input sound in each unit section is determined to be speech or non-speech based on the intensity of the component whose modulation frequency falls within a predetermined range in the modulation spectrum, the frequency spectrum of the input sound is used. Compared with the technique of Patent Document 1, it is possible to identify voice / non-voice with high accuracy.

  The range used for calculating the first index value in the modulation spectrum is such that when the input sound is one of speech and non-speech, the intensity of the modulation spectrum within the range is high, and the input sound is speech and non-speech. In the other case, it is set experimentally or statistically so that the intensity of the modulation spectrum outside the range becomes high. Now, when the input sound is speech, the intensity within the range in which the modulation frequency falls below a predetermined boundary value (for example, 10 Hz) increases in the modulation spectrum, and when the input sound is non-speech, the modulation frequency of the modulation spectrum. Pay attention to the tendency that the intensity within the range exceeding the boundary value increases. In the case where the first index value is defined so that the intensity of the component of the modulation spectrum whose modulation frequency is lower than the boundary value is higher, the determination means, for example, the input sound when the first index value exceeds a threshold value. When the first index value falls below the threshold, the input sound is determined as non-speech. In addition, when the first index value is defined so that the intensity of the component of the modulation spectrum whose modulation frequency is lower than the boundary value is higher, the determination unit inputs, for example, when the first index value is lower than the threshold value. The sound is determined as sound, and the input sound is determined as non-speech when the first index value exceeds the threshold value. Further, when the first index value is defined so that the intensity of the component whose modulation frequency exceeds the boundary value in the modulation spectrum increases, the determination unit inputs, for example, when the first index value exceeds a threshold value. The sound is determined as non-speech, and the input sound is determined as sound when the first index value falls below the threshold. In addition, when the first index value is defined so that the intensity of the component of the modulation spectrum whose modulation frequency exceeds the boundary value is higher, the determination unit determines the input sound when the first index value exceeds the threshold value. When the first index value falls below the threshold, the input sound is determined as non-speech. All the modes exemplified above are included in the concept of the process of “determining whether the input sound is voice or non-voice based on the first index value” in the first or second aspect.

  In the sound processing device according to the first aspect or the second aspect, for example, the first index calculation means includes the intensity of a component whose modulation frequency belongs to a predetermined range in the modulation spectrum and a range including the predetermined range ( That is, the first index value is calculated based on the relative ratio with the intensity of the component belonging to a range that is wider than the predetermined range including the predetermined range. In the above aspect, in addition to the intensity of the component within the predetermined range in the modulation spectrum, the intensity of the component in the range including the range (for example, the entire range of the modulation frequency) is applied to the calculation of the first index value. The Therefore, for example, even when the intensity over a wide range of the modulation spectrum is affected by the noise of the input sound, compared with the configuration in which the first index value is calculated based only on the intensity within the predetermined range, It is possible to discriminate voice with high accuracy.

  The sound processing apparatus according to the first aspect or the second aspect includes, for example, an intensity specifying unit that specifies the maximum value of the intensity of the modulation spectrum, and the determination unit uses the first index value and the maximum value of the intensity. Based on this, it is determined whether the input sound is voice or non-voice. For example, assuming that the maximum value of the intensity of the modulation spectrum of non-speech is lower than the maximum value of the intensity of the modulation spectrum of sound, the determination means determines that the unit increases as the maximum value of the modulation spectrum intensity increases. The voice / non-speech is determined such that the input sound in the section is likely to be determined as speech (the possibility that the input sound is determined as non-speech increases as the maximum intensity value decreases). More specifically, the determination unit determines that the input sound is non-speech when the maximum value of the intensity of the modulation spectrum is lower than the threshold even if it can be determined as sound from the first index value. In the above aspect, since the maximum value of the intensity of the modulation spectrum is also used for voice / non-voice determination in addition to the first index value, the range of the modulation frequency having high intensity in the non-voice modulation spectrum and the voice Even in the case where the modulation frequency range of the modulation spectrum approximates high, it is possible to distinguish between speech and non-speech with high accuracy.

  In the sound processing apparatus according to the first aspect or the second aspect, for example, the modulation spectrum specifying unit separates the unit interval from the component extraction unit that specifies the logarithmic spectrum of the input sound or the time locus of the specific component in the cepstrum. Frequency analysis means for Fourier transforming the time trajectory for each of the plurality of divided sections, and averaging means for specifying the modulation spectrum of the unit section by averaging the results of the Fourier transform for each of the plurality of divided sections of the unit section It comprises. In the above aspect, the Fourier transform of the logarithmic spectrum or the cepstrum time trajectory is executed for each of the plurality of divided sections into which the unit sections are divided. Compared with the execution, the number of points of the Fourier transform is reduced. Therefore, there is an advantage that the processing load by the modulation spectrum specifying means and the storage capacity necessary for the processing are reduced.

  The sound processing apparatus according to the third aspect of the present invention includes a storage unit that stores an acoustic model generated from a vowel sound, and a second index value indicating similarity between the input sound and the acoustic model for each unit section. Second index calculation means for calculating (for example, the index calculation unit 54 in FIG. 9) and determination means for determining whether the input sound of each unit section is speech or non-speech based on the second index value of the unit section To do. In the above aspect, since it is determined whether the input sound in each unit section is speech or non-speech based on the similarity between the acoustic model of the vowel speech and the input sound, a patent that uses the frequency spectrum of the input sound Compared with the technique of Document 1, it is possible to identify speech / non-speech with high accuracy.

  In the first aspect and the third aspect, assuming that the similarity between the speech and the acoustic model is higher than the similarity between the non-speech and the acoustic model, the determination unit has the second index value If the input sound is on the similar side with respect to the threshold, the input sound is determined to be speech, and if the second index value is on the dissimilar side with respect to the threshold, the input sound is determined to be non-speech. For example, in the aspect in which the second index value is defined so as to increase as the input sound and the acoustic model are similar, the determination unit determines that the input sound is speech when the second index value exceeds a threshold value. In the aspect in which the second index value is defined so that the input sound and the acoustic model decrease as the input model becomes similar, the determination unit determines that the input sound is speech when the second index value falls below a threshold value.

  In the specific example of the sound processing apparatus according to the first aspect and the third aspect, the storage unit stores one acoustic model generated from the sounds of a plurality of types of vowels. In the above aspect, since one acoustic model generated in an integrated manner from a plurality of types of vowel sounds is used, the storage means is compared with a configuration in which a separate acoustic model is prepared for each vowel type. There is an advantage that the capacity required for the system is reduced.

  The sound processing apparatus according to each specific example of the first to third aspects includes a voiced index calculation means (for example, a voiced index calculation unit that calculates a voiced index value according to a ratio of voiced sound sections among a plurality of sections into which unit sections are divided. 10 is provided, and the determination means determines whether the input sound is voice or non-voice based on the voiced index value. For example, on the assumption that the temporal ratio of voiced sound is higher than that of non-speech in the voice, the determination means can determine the input sound of the unit section as voice as the ratio of voiced sound is higher The voice / non-speech is determined so as to increase the likelihood (the possibility of determining non-speech increases as the proportion of voiced sound decreases). More specifically, the determination means is a ratio of the voiced sound section even when the index value calculated by the index calculation means (at least one of the first index value to the third index value) can be determined as speech. If there are many, the input sound is determined as non-speech. In the above aspect, since the voiced index value is also used for voice / non-voice determination in addition to the index value calculated from the modulation spectrum and the acoustic model, the non-voice modulation is performed in the first mode or the third mode. The vowel sound of non-speech and speech in the second mode or the third mode when the range of the modulation frequency with high intensity in the spectrum approximates the range of the modulation frequency with high intensity in the modulation spectrum of speech Even when the similarity to the model is approximate, speech and non-speech can be distinguished with high accuracy.

  The sound processing apparatus according to each specific example of the first to third aspects includes threshold setting means for variably setting a threshold according to the SN ratio of the input sound, and the determination means is calculated from the input sound. Whether the input sound is speech or non-speech depending on the magnitude of the index value (any one of the first index value, the second index value, the third index value, the voiced index value, and the maximum modulation spectrum intensity) and the threshold value judge. In the above aspect, since the threshold value to be compared with the index value is variably controlled according to the SN ratio of the input sound, the accuracy of voice / non-voice is maintained at a high level regardless of the level of the SN ratio. It is possible.

  In the sound processing device according to each specific example of the first to third aspects, when the determination unit determines non-speech for three or more consecutive unit sections, the sound processing apparatus is in the middle of the three or more unit sections. Sound processing means for muting only the input sound of the unit interval. In the above aspect, since the unit section determined to be non-speech is muted, the listener can clearly perceive only the sound of the input sound. Further, a unit section in the middle of three or more unit sections determined to be non-speech (that is, at least one unit section excluding the first unit section and the last unit section among three or more unit sections) Since only the sound is muted, the possibility that the beginning (the last unit section of three or more) and the end (the first unit section of three or more) of the sound are muted by the sound processing means processing is reduced. Is done.

  The sound processing apparatus according to all of the above aspects is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to processing of input sound, or a general purpose such as a CPU (Central Processing Unit). This is also realized by cooperation between the arithmetic processing unit and the program. The program according to the first aspect includes a modulation spectrum specifying process for specifying a modulation spectrum of an input sound for each of a plurality of unit sections, and a first corresponding to the intensity of a component whose modulation frequency belongs to a predetermined range in the modulation spectrum. A first index calculation process for calculating an index value, a second index calculation process for calculating a second index value indicating the similarity between an acoustic model generated from a vowel sound and an input sound for each unit section, and each unit The computer is caused to execute a determination process for determining whether the input sound of the section is voice or non-speech based on the first index value and the second index value of the unit section. The program according to the second aspect includes a modulation spectrum specifying process for specifying a modulation spectrum of an input sound for each of a plurality of unit sections, and a first corresponding to the intensity of a component whose modulation frequency belongs to a predetermined range in the modulation spectrum. The computer executes a first index calculation process for calculating the index value and a determination process for determining whether the input sound of each unit section is a voice or a non-speech based on the first index value. The program according to the third aspect includes a second index calculation process for calculating, for each unit section, a second index value indicating similarity between an acoustic model generated from a vowel sound and an input sound, and input for each unit section And causing the computer to execute determination processing for determining whether the sound is voice or non-voice based on the second index value of the unit section. According to the program of this invention, the effect | action and effect similar to the sound processing apparatus which concern on each above aspect are show | played. 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.

<A: First Embodiment>
FIG. 1 is a block diagram of a remote conference system according to the first embodiment of the present invention. The remote conference system 100 is a system in which a plurality of users U (conference participants) exchange voices with each other in geographically separated spaces R1 and R2. In each space R (R1, R2), a sound collecting device 12, a sound processing device 14, a sound processing device 16, and a sound emitting device 18 are installed.

  The sound collection device 12 is a device (microphone) that generates an acoustic signal SIN representing the waveform of the input sound VIN existing in the space R. Each sound processing device 14 in the space R1 and the space R2 generates an output signal SOUT from the acoustic signal SIN and transmits the output signal SOUT to the other sound processing device 16 in the space R1 and the space R2. The sound processing device 16 amplifies the output signal SOUT and outputs it to the sound emitting device 18. The sound emitting device 18 is a device (speaker) that emits sound waves according to the amplified output signal SOUT supplied from the sound processing device 16. With the above configuration, the utterance sound of each user U in the space R1 is output from the sound emitting device 18 in the space R2, and the utterance sound of each user U in the space R2 is output from the sound emitting device 18 in the space R1. Is output.

  FIG. 2 is a block diagram showing a configuration of the sound processing device 14 installed in each of the space R1 and the space R2. As shown in FIG. 2, the sound processing device 14 includes a control device 22 and a storage device 24. The control device 22 is an arithmetic processing device that functions as each element in FIG. 2 by executing a program. 2 are also realized by an electronic circuit such as a DSP. The storage device 24 stores a program executed by the control device 22 and various data used by the control device 22. A known storage medium such as a semiconductor storage device or a magnetic storage device is arbitrarily used as the storage device 24.

  The control device 22 determines that the input sound VIN is voice and sound for each of a plurality of sections (hereinafter referred to as “unit sections”) obtained by dividing the acoustic signal SIN (input sound VIN) supplied from the sound collection device 12 along the time axis. The function of determining which of the non-sounds is applicable and the function of generating the output signal SOUT by executing the process corresponding to the result of the determination of the sound / non-speech for the acoustic signal SIN. The voice is an utterance sound uttered by a human. Non-voice is sound other than voice (for example, environmental sounds (noise) such as door opening / closing sound of the space R, operation sound of air conditioning equipment, ringtone of mobile phone).

  The modulation spectrum specifying unit 32 in FIG. 2 specifies the modulation spectrum MS of the acoustic signal SIN (input sound VIN). The modulation spectrum MS is a result of executing a Fourier transform on a temporal variation (hereinafter referred to as “time locus”) of a component belonging to a specific frequency band in the logarithmic spectrum (frequency spectrum) of the acoustic signal SIN.

  FIG. 3 is a block diagram illustrating a functional configuration of the modulation spectrum specifying unit 32, and FIG. 4 is a conceptual diagram for explaining processing by the modulation spectrum specifying unit 32. As shown in FIG. 3, the modulation spectrum specifying unit 32 includes a frequency analysis unit 322, a component extraction unit 324, and a frequency analysis unit 326. The frequency analysis unit 322 performs frequency analysis including Fourier transform (for example, FFT (Fast Fourier Transform)) on the acoustic signal SIN, thereby converting the acoustic signal SIN to time as shown in part (A) of FIG. A logarithmic spectrum S0 is calculated for each of a plurality of frames segmented along the axis. Therefore, a spectrogram SP in which the logarithmic spectrum S0 is arranged for each frame along the time axis is generated. Each successive frame may overlap partially or may be set so as not to overlap each other.

  The component extraction unit 324 in FIG. 3 extracts the time trajectory ST of the intensity (energy) of the component belonging to the specific frequency band ω in the spectrogram SP as shown in part (A) and part (B) of FIG. . More specifically, the component extraction unit 324 calculates the intensity of the component belonging to the frequency band ω in the logarithmic spectrum S0 of each frame and arranges the intensity of the logarithmic spectrum S0 in time series for a plurality of frames. Generate ST. In the frequency band ω, the frequency characteristic (modulation spectrum MS) of the time trajectory ST when the input sound VIN is speech and the frequency characteristic of the time trajectory ST when the input sound VIN is non-speech are significantly different. Pre-selected experimentally or statistically. For example, the frequency band ω is selected in a range from 10 Hz (more preferably 50 Hz) to 800 Hz. A configuration is also employed in which the component extraction unit 324 extracts the time series of the intensity of one frequency component in each logarithmic spectrum S0 as the time locus ST.

  As shown in part (B) and part (C) of FIG. 4, the frequency analysis unit 326 in FIG. 3 performs a Fourier transform (for example, FFT) on the time trajectory ST, thereby converting the time trajectory ST into a time axis. The modulation spectrum MS is calculated for each of a plurality of unit intervals TU divided along the line. The unit section TU is a period of a predetermined time length (for example, about 1 second) composed of a plurality of frames. In the present embodiment, a configuration in which the unit sections TU do not overlap is illustrated for convenience, but a configuration in which the adjacent unit sections TU partially overlap is also employed.

  FIG. 5 shows a typical modulation spectrum MS of voice (human uttered sound), and FIG. 6 shows a modulation spectrum of non-voice (a crumbling sound when a net-like portion covering the tip of the sound collecting device 12 is scratched). Indicates MS. As understood from the comparison between FIG. 5 and FIG. 6, the range of the modulation frequency having a high intensity in the modulation spectrum MS tends to be different between voice and non-voice.

  In the modulation spectrum MS of a normal human speech sound (ie, speech), the intensity often becomes maximum at a modulation frequency of about 4 Hz corresponding to the frequency at which the syllable is switched during speech. Therefore, the intensity of the modulation spectrum MS (FIG. 5) is high in a low frequency range of 10 Hz or less, whereas the modulation frequency is 10 Hz in many non-voice modulation spectra MS (FIG. 6). There is a difference that the strength increases in the range exceeding. In consideration of the above differences, in the present embodiment, the modulation frequency of the modulation spectrum MS specified by the modulation spectrum specifying unit 32 depends on the intensity of the component belonging to a predetermined range (hereinafter referred to as “determination target range”) A. It is determined whether the input sound VIN is voice or non-voice. In this embodiment, a range of 10 Hz or less (more preferably, a range of 2 Hz to 8 Hz) is set as the determination target range A.

2 calculates an index value D1 corresponding to the intensity (energy) of the component belonging to the determination target range A for the modulation spectrum MS specified by the modulation spectrum specifying unit 32 for each unit section TU. More specifically, the index calculation unit 34 firstly calculates the intensity of the component whose modulation frequency belongs to the determination target range A in the modulation spectrum MS (for example, the added value or average of the intensity at each modulation frequency in the determination target range A). Value) L1 and the intensity of the modulation spectrum MS over the entire range of the modulation frequency (addition value or average value of the intensity at all modulation frequencies) L2. Second, the index calculator 34 calculates the index value D1 based on the following arithmetic expression (A) including the relative ratio (L1 / L2) between the intensity L1 and the intensity L2.
D1 = 1- (L1 / L2) (A)
As understood from the content of the arithmetic expression (A), the index value D1 increases as the intensity L1 of the component in the determination target range A of the modulation spectrum MS increases (that is, the possibility that the input sound VIN is a voice is higher). Is a small number. Therefore, the index value D1 is an index as to whether the input sound VIN is speech or non-speech. Further, since the determination target range A includes the frequency at which the syllable is switched during speech, the index value D1 is also grasped as an index as to whether or not the input sound VIN includes a rhythm peculiar to speech (speech rhythm). Is done.

  However, there is also a non-speech in which the intensity of the component in the determination target range A in the modulation spectrum MS is relatively high compared to other ranges. In the modulation spectrum MS of non-voice (phone push tone) shown in FIG. 7, an intensity peak occurs at a modulation frequency of about 5 Hz to 8 Hz included in the determination target range A. However, in the case of non-speech with characteristics as shown in FIG. 7, the maximum value P of the intensity of the modulation spectrum MS tends to be lower than that of speech. Considering the above tendency, in the present embodiment, it is determined whether the input sound VIN is voice or non-voice based on the index value D1 and the maximum value P of the intensity of the modulation spectrum MS. The intensity specifying unit 36 in FIG. 2 specifies the maximum intensity P of the modulation spectrum MS for each unit interval TU.

  The determination unit 42 determines whether the input sound VIN of each unit section TU is speech or non-speech based on the index value D1 calculated by the index calculation unit 34 and the maximum value P specified by the intensity specifying unit 36. Identification data d indicating a result (speech / non-speech distinction) is generated for each unit interval TU. FIG. 8 is a flowchart showing a specific operation of the determination unit 42. The process of FIG. 8 is executed every time the index value D1 and the maximum value P are specified for one unit section TU.

  The determination unit 42 determines whether or not the index value D1 exceeds the threshold value THd1 (step SA1). The threshold THd1 is selected experimentally or statistically so that the voice index value D1 is lower than the threshold THd1 and the non-voice index value D1 is higher than the threshold THd1. When the result of step SA1 is affirmative (for example, when the input sound VIN is non-speech with the characteristics of FIG. 6), the determination unit 42 determines that the input sound VIN of the unit interval TU that is the object of the current process is non-speech. Judgment is made (step SA2). That is, the determination unit 42 generates identification data d indicating non-voice.

  On the other hand, if the result of step SA1 is negative, the determination unit 42 determines whether or not the maximum value P of the intensity of the modulation spectrum MS is below the threshold value THp (step SA3). If the result of step SA3 is affirmative, the determination unit 42 shifts the process to step SA2 to generate identification data d indicating non-voice. That is, even when the input sound VIN can be determined to be a sound if only the index value D1 is considered, the maximum value P is below the threshold value THp (for example, when the input sound VIN is a non-speech sound having the characteristics shown in FIG. 7). The input sound VIN is determined as non-speech.

  When the result of step SA3 is negative (for example, when the input sound VIN is a sound having the characteristics shown in FIG. 5), the determination unit 42 determines that the input sound VIN of the unit section TU that is the target of the current process is a sound. (Step SA4). That is, the determination unit 42 generates identification data d indicating a voice. As described above, only the input sound VIN of the unit section TU in which both the intensity L1 and the maximum intensity value P in the determination target range A in the modulation spectrum MS are high is determined as the sound.

  The sound processing unit 44 in FIG. 2 generates an output signal SOUT by executing processing corresponding to the identification data d of each unit section TU on the acoustic signal SIN of the unit section TU. For example, the sound processing unit 44 outputs the acoustic signal SIN as the output signal SOUT for the unit interval TU in which the identification data d indicates speech, while setting the volume to zero for the unit interval TU in which the identification data d indicates non-speech. The output signal SOUT is output (that is, the acoustic signal SIN is not output). Therefore, in each of the space R1 and the space R2, the non-speech of the input sound VIN in the other space R is removed, and only the sound that the user originally needs to listen to passes through the sound processing device 16. And emitted from the sound emitting device 18.

  As described above, in this embodiment, since the voice / non-speech is determined based on the intensity L1 (presence / absence of utterance rhythm) of the component within the determination target range A in the modulation spectrum MS, the input sound VIN It is possible to identify speech / non-speech with high accuracy compared to the technique of Patent Document 1 using the frequency spectrum of In addition to the intensity L1 of the component in the determination target range A, the maximum value P of the intensity of the modulation spectrum MS is also used for the determination. Therefore, the intensity L1 of the component in the determination target range A is compared with other ranges. High non-speech can also be determined as non-speech.

  When the volume of non-speech is high, the modulation spectrum MS has a high intensity over the entire band of the modulation frequency. Therefore, in the configuration in which the voice / non-speech of the input sound VIN is identified based only on the intensity L1 within the determination target range A of the modulation spectrum MS, there is a high possibility that the non-speech having a large volume is erroneously determined as a voice. In this embodiment, since voice / non-speech is determined based on the relative ratio between the intensity L1 in the determination target range A and the intensity L2 over the entire range of the modulation frequency, even if the volume of non-speech is high. There is an advantage that voice / non-voice can be accurately determined.

<B: Second Embodiment>
Next, a second embodiment of the present invention will be described. In the following embodiments, elements having the same functions and functions as those of the first embodiment are denoted by the same reference numerals as above, and detailed descriptions thereof are appropriately omitted.

  FIG. 9 is a block diagram of the sound processing device 14. One acoustic model M is stored in the storage device 24 of this embodiment. The acoustic model M is a statistical model that models the average acoustic characteristics of a plurality of types of vowels produced by many speakers. The acoustic model M of the present embodiment models the distribution of speech feature values (for example, MFCC (Mel-Frequency Cepstrum Coefficient)) as a weighted sum of probability distributions. For example, a Gaussian mixture model (GMM (Gaussian Mixture Model)) that models a feature amount of speech as a weighted sum of a plurality of normal distributions is suitable as the acoustic model M.

  The acoustic model M is created, for example, when the control device 22 executes the following process. First, the control device 22 collects voices when a large number of speakers utter various sentences, classifies each voice into phonemes, and a plurality of types of vowels (a, i, u, e, o). Only the waveform corresponding to is extracted. Secondly, the control device 22 extracts an acoustic feature amount (feature vector) for each of a plurality of frames obtained by dividing the waveform of each portion corresponding to a vowel along the time axis. The time length of each frame is, for example, 20 milliseconds, and the time difference between successive frames is about 10 milliseconds. Thirdly, the control device 22 generates the acoustic model M by integrally processing the feature quantities extracted from a large number of sounds with respect to a plurality of types of vowels. For the generation of the acoustic model M, a known technique such as an EM (Expectation-Maximization) algorithm is arbitrarily employed. Note that since the vowel feature quantity is affected by the immediately preceding phoneme (consonant), the acoustic model M generated by the above procedure is not a statistical model purely modeling only the vowel characteristics. That is, the acoustic model M can be said to be a statistical model (or a statistical model of voiced voices) created with a plurality of vowels as a center.

  As shown in FIG. 9, the sound processing device 14 includes a feature extraction unit 52 and an index calculation unit 54 instead of the modulation spectrum specification unit 32, the index calculation unit 34, and the intensity specification unit 36 of FIG. 2. The feature extraction unit 52 extracts a feature quantity (for example, MFCC) X of the same type as the feature quantity used for generating the acoustic model M for each frame of the acoustic signal SIN. A known technique is arbitrarily employed for extracting the feature amount X by the feature extraction unit 52.

演算式(B)から理解されるように、音響モデルＭと単位区間ＴU内の入力音ＶINとで特徴量が類似するほど指標値Ｄ2は小さくなる。非音声と比較すると音声は母音の割合が多い（したがって音響モデルＭとの音色の類似の程度が高い）という傾向がある。したがって、入力音ＶINが音声である場合に算定される指標値Ｄ2は、入力音ＶINが非音声である場合に算定される指標値Ｄ2と比較して小さい数値となる。すなわち、指標値Ｄ2は、入力音ＶINが音声であるか非音声であるかの指標となる。したがって、音響モデルＭは、音声（人間の発話音）の統計モデルとしても把握される。 The index calculation unit 54 calculates an index value D2 corresponding to the similarity between the input sound VIN and the acoustic model M represented by the acoustic signal SIN for each unit section TU of the acoustic signal SIN. More specifically, the index value D2 represents the likelihood (probability) p (X | M) that the feature quantity X extracted from the acoustic signal SIN of each frame is generated from the acoustic model M within the unit interval TU ( (n) frames are averaged values. That is, the index value D2 is calculated by the following arithmetic expression (B).

As can be understood from the arithmetic expression (B), the index value D2 decreases as the feature amount is similar between the acoustic model M and the input sound VIN in the unit interval TU. Compared with non-speech, speech tends to have a higher proportion of vowels (thus, the degree of timbre similarity with the acoustic model M is higher). Therefore, the index value D2 calculated when the input sound VIN is speech is a smaller numerical value than the index value D2 calculated when the input sound VIN is non-speech. That is, the index value D2 is an index as to whether the input sound VIN is voice or non-voice. Therefore, the acoustic model M is also grasped as a statistical model of speech (human speech sound).

  The determination unit 42 in FIG. 9 determines whether the input sound VIN of each unit section TU is speech or non-speech based on the index value D2 calculated by the index calculation unit 54, and uses the identification data d indicating the determination result as the unit section. Generated for each TU. The index value D2 is a numerical value indicating the timbre similarity between the input sound VIN and the acoustic model M. That is, in the first embodiment, it is determined whether or not the rhythm of the input sound VIN (intensity L1 within the determination target range A) is sound, whereas in the present embodiment, whether or not the timbre of the input sound VIN is sound. Determine whether.

  More specifically, the determination unit 42 determines whether or not the index value D2 of each unit section TU exceeds a predetermined threshold value THd2. The threshold value THd2 is selected experimentally or statistically so that the voice index value D2 is lower than the threshold value THd2 and the non-voice index value D2 is higher than the threshold value THd2. When the result of the determination is affirmative (D2> THd2), the determination unit 42 determines the input sound VIN of the unit section TU as a non-speech and generates identification data d. On the other hand, if the result of the determination is negative (D2 <THd2), the determination unit 42 determines the input sound VIN of the unit section TU as a sound and generates identification data d. The operation of the sound processing unit 44 according to the identification data d is the same as in the first embodiment.

  As described above, in this embodiment, since voice / non-speech is determined according to the similarity with the acoustic model M that models the voice of a vowel, Patent Literature that uses the frequency spectrum of the input sound VIN Compared with the first technique, it is possible to identify speech / non-speech with high accuracy. In addition, since one acoustic model M in which a plurality of types of vowels are modeled in an integrated manner is stored in the storage device 24, it is stored in comparison with a configuration in which individual acoustic models are prepared for each of the plurality of types of vowels. There is an advantage that the capacity required for the device 24 is reduced.

<C: Third Embodiment>
FIG. 10 is a block diagram of a sound processing apparatus 14 according to the third embodiment of the present invention. As in the first embodiment, the modulation spectrum specifying unit 32 and the index calculating unit 34 in FIG. 10 calculate the index value D1 for each unit interval TU of the input sound VIN, and the intensity specifying unit 36 is the maximum intensity of the modulation spectrum MS. The value P is specified. Further, the feature extraction unit 52 and the index calculation unit 54 calculate the index value D2 for each unit interval TU of the input sound VIN, as in the second embodiment.

The index calculation unit 62 calculates the weighted sum of the index value D1 calculated by the index calculation unit 34 and the index value D2 calculated by the index calculation unit 54 as the index value D3. The index value D3 is calculated by, for example, the following arithmetic expression (C).
D3 = D1 + α ・ D2 (C)
As understood from the calculation formula (C), the higher the possibility that the input sound VIN is a voice (that is, the higher the intensity L1 in the determination target range A in the modulation spectrum MS or the acoustic model M). The index value D3 is a small numerical value as the feature amount is similar to the input sound VIN in the unit interval TU. The weight value α is a positive number (α> 0) set by the weight value setting unit 66 of FIG. The index value D3 calculated by the index calculation unit 62 is used for voice / non-voice determination in the determination unit 42.

  The SN ratio specifying unit 64 in FIG. 10 calculates the SN ratio R of the sound signal SIN (input sound VIN) for each unit interval TU. The weight value setting unit 66 varies the weight value α applied by the index calculating unit 62 to calculate the index value D3 of each unit section TU based on the SN ratio R calculated by the SN ratio specifying unit 64 for the unit section TU. Set to.

  Here, the index value D1 calculated from the modulation spectrum MS tends to be easily influenced by noise of the input sound VIN as compared with the index value D2 calculated from the acoustic model M. Therefore, the weight value setting unit 66 variably controls the weight value α so that the weight value α increases as the SN ratio R is lower (the noise is higher). According to the above configuration, as the SN ratio R is lower, the influence of the index value D2 on the index value D3 is relatively increased (the influence of the index value D1 that is easily affected by noise is reduced). Even when noise is superimposed on the voice / non-voice, it is possible to determine voice / non-voice with high accuracy.

  10 determines whether the input sound VIN is a voiced sound or an unvoiced sound for each of a plurality of frames. A known technique is arbitrarily employed for the determination by the voiced / unvoiced discrimination unit 72. For example, the voiced / unvoiced discriminating unit 72 detects the pitch (fundamental frequency) for each frame of the input sound VIN and determines a frame in which a significant pitch is detected as a voiced sound, but a clear pitch is not detected. The frame is determined as an unvoiced sound.

  The index calculation unit 74 calculates a voiced index value DV for each unit section TU of the acoustic signal SIN. The voiced index value DV is the ratio (DV = NV / n) of the number of frames NV that the voiced / unvoiced discriminating unit 72 has determined to be voiced out of n frames in the unit interval TU. Voice (human utterance) tends to have a higher proportion of voiced sound than non-voice. Therefore, the voiced index value DV calculated when the input sound VIN is speech is a larger numerical value than the voiced index value DV calculated when the input sound VIN is non-speech.

  The determination unit 42 in FIG. 10 determines each unit interval TU based on the index value D3 calculated by the index calculation unit 62, the maximum value P specified by the strength specifying unit 36, and the voiced index value DV calculated by the index calculation unit 74. It is determined whether the input sound VIN is speech or non-speech, and identification data d indicating the determination result is generated for each unit interval TU. FIG. 11 is a flowchart showing a specific operation of the determination unit 42. The process of FIG. 11 is executed every time the index value D3, the maximum value P, and the voiced index value DV are specified for one unit section TU.

  The determination unit 42 determines whether or not the index value D3 exceeds the threshold value THd3 (step SB1). The threshold THd3 is selected experimentally or statistically so that the voice index value D3 is lower than the threshold THd3 and the non-voice index value D3 is higher than the threshold THd3. If the result of step SB1 is affirmative, the determination unit 42 determines that the input sound VIN of the current unit section TU is non-speech and generates identification data d (step SB2).

  On the other hand, when the result of step SB1 is negative, the determination unit 42 determines whether or not the maximum value P is lower than the threshold value THp, similarly to step SA3 of FIG. 8 (step SB3). If the result of step SB3 is affirmative, the determination unit 42 generates identification data d indicating non-voice in step SB2. If the result of step SB3 is negative, the determination unit 42 determines whether or not the voiced index value DV is below the threshold value THdv (step SB4).

  When the result of step SB4 is affirmative (that is, when the ratio of voiced frames in the unit interval TU is small), the determination unit 42 generates identification data d indicating non-speech in step SB2. On the other hand, if the result of step SB4 is negative, the determination unit 42 determines that the input sound VIN of the current unit section TU is a voice and generates identification data d. The operation of the sound processing unit 44 according to the identification data d is the same as in the first embodiment.

  As described above, in the present embodiment, voice / non-speech is determined from the viewpoints of both the rhythm (index value D1) and timbre (index value D2) of the input sound VIN. Compared with the second embodiment, it is possible to distinguish the input sound VIN into voice and non-voice with high accuracy. In addition to the index value D1 and the index value D2, the voiced index value DV is also applied to the determination of voice / non-speech. For example, even if the input sound VIN has a rhythm or tone similar to the voice, the voiced index value DV If is low, it can be determined as non-voice.

<D: Modification>
Various modifications are added to 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
The modulation spectrum specifying unit 32 is changed to the configuration of FIG. 12 includes an averaging unit 328 in addition to a frequency analysis unit 322, a component extraction unit 324, and a frequency analysis unit 326 similar to those in FIG. The time trajectory ST generated by the component extraction unit 324 is divided into m sections (hereinafter referred to as “divided sections”) obtained by further dividing the unit section TU (m is a natural number of 2 or more). The frequency analysis unit 326 calculates a modulation spectrum for each divided section by performing Fourier transform on the time trajectory ST of each divided section. The averaging unit 328 calculates the modulation spectrum MS of the unit section TU by averaging the m modulation spectra calculated for each divided section constituting the unit section TU. According to the configuration of FIG. 12, since the number of points of Fourier transform executed by the frequency analysis unit 326 is reduced compared to the first embodiment, the load (computation amount) of Fourier transform by the frequency analysis unit 326 and the Fourier transform are reduced. There is an advantage that the capacity of the necessary storage device 24 is reduced.

(2) Modification 2
A configuration in which the threshold value TH (THd1, THd2, THd3, THp, THdv) used for voice / non-voice judgment is variably controlled is also suitable. For example, as shown in FIG. 13, a threshold setting unit 68 is added to the sound processing apparatus 14 of the third embodiment. The threshold setting unit 68 variably controls the threshold TH in accordance with the SN ratio R calculated by the SN ratio specifying unit 64.

  Even if the input sound VIN is actually a voice, if the SN ratio R is low, the determination unit 42 is more likely to erroneously determine the input sound VIN as a non-voice. Therefore, the threshold value setting unit 68 controls each threshold value TH in such a direction that the input sound VIN is more likely to be determined to be a voice as the SN ratio R calculated by the SN ratio specifying unit 64 is lower. For example, as the SN ratio R is lower, the threshold value THd3 is increased and the threshold value THp and the threshold value THdv are decreased. According to the above configuration, it is possible to reduce the possibility that the input sound VIN including sound is erroneously determined as non-speech. A configuration in which the threshold value TH is variably controlled according to a numerical value other than the SN ratio R (for example, the volume of the input sound VIN) is also employed. Moreover, although the modification of 3rd Embodiment was illustrated in FIG. 13, the S / N ratio specific | specification part 64 and the threshold value setting part 68 were similarly added also about the sound processing apparatus 14 of 1st Embodiment or 2nd Embodiment. Configuration is adopted.

(3) Modification 3
In each of the above forms, when the proportion of the voice included in the unit section TU is small (for example, when the voice is included only in a short section of the unit section TU), the unit section TU is determined as non-speech. there is a possibility. Therefore, in the configuration in which the input sound VIN is muted uniformly for all the unit intervals TU determined to be non-speech, the unit interval TU slightly including the start and end portions (particularly the unvoiced consonant portion) of the sound is not included. In some cases, the sound is judged to be sound and muted. Therefore, it is preferable that the input sound VIN is silenced in each unit section TU in consideration of determination by the determination unit 42 for a plurality of unit sections TU.

  For example, when one unit section TU is determined to be non-speech, the sound processing unit 44 does not mute the unit section TU and, as shown in FIG. When the input sound VIN is determined to be non-speech with respect to the unit interval TU of (natural number), the sound processing unit 44 first (first) and last (kth) unit interval among the k unit intervals TU. The input sound VIN of each unit section TU excluding TU (that is, the unit section TU in the middle of the k sets) is muted. The input sound VIN is not muted for the first and kth unit intervals TU. For example, only the input sound VIN of the second unit section TU among the three (k = 3) unit sections TU determined as non-voice is muted. According to the above configuration, the unit section TU including speech only immediately after the start point (for example, the first unit section TU among the k unit sections TU in FIG. 14) or the unit section TU including speech immediately before the end point. Since mute is not executed for (for example, the k-th unit interval TU in FIG. 14), there is an advantage that voice loss is prevented.

(4) Modification 4
The definition of each index value D (D1, D2, D3) is changed as appropriate. Therefore, the relationship between the magnitude of each index value D (D1, D2, D3) and voice / non-voice is arbitrary. For example, in the first embodiment, the index value D1 is defined such that the smaller the index value D1 is, the higher the possibility that the input sound VIN is determined to be speech. For example, the relative ratio of the intensity L1 to the intensity L2 is defined as the index value. If it is defined as D1 (D1 = L1 / L2), the greater the index value D1, the higher the possibility of being determined to be speech. Further, the index value D3 is defined using one weight value α, but the index value D3 (D3 = β) is applied by applying the weight values (β, γ) set independently for the index value D1 and the index value D2. A configuration for calculating D1 + γ · D2) is also suitable. Also, the weight values (α, β, γ) used for calculating the index value D3 may be fixed values.

(5) Modification 5
In the first embodiment and the third embodiment, the modulation spectrum MS is specified by performing Fourier transform on the time trajectory ST of the component belonging to the frequency band ω in the logarithmic spectrum S0, but the acoustic signal SIN (input sound VIN) is specified. A configuration is also adopted in which the modulation spectrum MS is specified by performing Fourier transformation on the time trajectory of the cepstrum. More specifically, the frequency analyzing unit 322 of the modulation spectrum specifying unit 32 calculates a cepstrum for each frame of the acoustic signal SIN, and the component extracting unit 324 is a component having a quefrency within a specific range in the cepstrum of each frame. The frequency analysis unit 326 performs a Fourier transform for each unit section TU (or for each divided period as in Modification 1) with respect to the time path ST of the cepstrum, so that each unit section TU is extracted. The modulation spectrum MS is calculated.

(6) Modification 6
Variables used for voice / non-voice determination are changed as appropriate. For example, the determination according to the maximum value P (step SA3 in FIG. 8 or step SB3 in FIG. 11) may be omitted in the first embodiment or the third embodiment, and the voiced index value in the third embodiment. The determination according to DV (step SB4 in FIG. 11) may be omitted. Moreover, the structure which added the voiced unvoiced determination part 72 and the parameter | index calculation part 74 to 1st Embodiment or 2nd Embodiment is also suitable.

(7) Modification 7
In each of the above embodiments, the identification data d and the output signal SOUT are generated by the sound processing device 14 in the space R that picks up the input sound VIN, but the position for generating the identification data d and the output signal SOUT are generated. The position is changed as appropriate. For example, in the configuration in which the sound processing device 14 outputs the acoustic signal SIN generated by the sound collection device 12 and the identification data d generated by the determination unit 42, the output signal SOUT is generated from the acoustic signal SIN and the identification data d. A sound processing unit 44 is arranged in the sound processing device 16 on the receiving side. Further, in the configuration in which the sound processing device 14 transmits the acoustic signal SIN generated by the sound collecting device 12, the same elements as those in FIG. 2 are installed in the sound processing device 16 on the receiving side. However, the remote conference system 100 is merely an example of the application of the present invention. Therefore, transmission / reception of the output signal SOUT and the acoustic signal SIN is not essential in the present invention.

(8) Modification 8
In each of the above embodiments, the sound processing unit 44 does not output the acoustic signal SIN of the unit section TU determined to be non-speech (sets the volume of the output signal SOUT to zero), but the sound processing unit 44 is exemplified. The content of the processing by is changed as appropriate. For example, a configuration in which the sound processing unit 44 outputs as the output signal SOUT a signal in which the volume of the acoustic signal SIN is reduced for the unit interval TU determined to be non-speech, or the unit interval TU determined to be speech and non-speech A configuration in which the sound processing unit 44 outputs a signal obtained by adding a separate acoustic effect to the acoustic signal SIN in the unit interval TU as the output signal SOUT is also preferable. Further, in a configuration in which speech recognition or speaker recognition (speaker identification or speaker authentication) is performed at the output destination of the output signal SOUT (sound processing device 16), the sound processing unit 44 is determined to be, for example, speech. For the unit section TU, the feature quantity used for speech recognition and speaker recognition is extracted from the acoustic signal SIN and output as the output signal SOUT, while the feature quantity is extracted for the unit section TU determined to be non-speech. Stop.

1 is a block diagram of a remote conference system according to a first embodiment of the present invention. It is a block diagram of the sound processing apparatus of FIG. It is a block diagram of the modulation spectrum specific | specification part of FIG. It is a conceptual diagram which shows the procedure of the process by the modulation spectrum specific | specification part of FIG. This is the modulation spectrum of speech. This is a non-voice modulation spectrum. This is a non-voice modulation spectrum. It is a flowchart which shows operation | movement of the determination part of FIG. It is a block diagram of the sound processing apparatus which concerns on 2nd Embodiment of this invention. It is a block diagram of the sound processing apparatus which concerns on 3rd Embodiment of this invention. It is a flowchart which shows operation | movement of the determination part of FIG. It is a block diagram of the modulation spectrum specific | specification part which concerns on a modification. It is a block diagram of the sound processing apparatus which concerns on a modification. It is a conceptual diagram which shows operation | movement of the sound processing part which concerns on a modification.

Explanation of symbols

100 …… Remote conference system, 12 …… Sound collecting device, 14 …… Sound processing device, 16 …… Sound processing device, 18 …… Sound emitting device, 22 …… Control device, 24 …… Storage device, 32 …… Modulation spectrum identification unit, 322... Frequency analysis unit, 324... Component extraction unit, 326... Frequency analysis unit, 328... Average unit, 34, 54, 62, 74. , 42... Determination unit, 44... Sound processing unit, 52... Feature extraction unit, 64... SN ratio specifying unit, 66... Weight setting unit, 68. Discriminating unit, VIN …… input sound, SIN …… acoustic signal, SOUT …… output signal, d …… identification data, MS …… modulation spectrum, M …… acoustic model, D1, D2, D3 …… index value, P …… Maximum value of intensity of modulation spectrum, R …… SNR, TU …… Unit interval.

Claims (7)

Modulation spectrum specifying means for specifying the modulation spectrum of the input sound for each of a plurality of unit sections;
First index calculating means for calculating a first index value according to the intensity of a component whose modulation frequency belongs to a predetermined range in the modulation spectrum;
Storage means for storing an acoustic model generated from vowel sounds;
Second index calculation means for calculating a second index value indicating similarity between the input sound and the acoustic model for each unit section;
A sound processing apparatus comprising: a determination unit that determines whether the input sound of each unit section is speech or non-speech based on the first index value and the second index value of the unit section. The sound processing apparatus according to claim 1, wherein the storage unit stores one acoustic model generated from a plurality of types of vowel sounds. Comprising a third index calculation means for calculating a weighted sum of the first index value and the second index value as a third index value;
The sound processing device according to claim 1, wherein the determination unit determines whether the input sound of each unit section is voice or non-speech based on the third index value of the unit section. The sound processing apparatus according to claim 3, further comprising weight value setting means for variably setting a weight value applied to the calculation of the third index value by the third index calculation means according to the SN ratio of the input sound. Comprising a voiced index calculating means for calculating a voiced index value according to a ratio of a voiced sound section among a plurality of sections into which the unit sections are divided;
5. The sound processing according to claim 1, wherein the determination unit determines whether the input sound is voice or non-voice based on the first index value, the second index value, and the voiced index value. apparatus. And a sound processing unit that silences only an input sound in a unit section in the middle of the three or more unit sections when the determination unit determines non-speech for three or more consecutive unit sections. The sound processing device according to any one of claims 1 to 5. A modulation spectrum specifying process for specifying the modulation spectrum of the input sound for each of the plurality of unit sections;
A first index calculation process for calculating a first index value according to the intensity of a component whose modulation frequency belongs to a predetermined range in the modulation spectrum;
A second index calculation process for calculating, for each unit section, a second index value indicating the similarity between the acoustic model generated from the vowel sound and the input sound;
A program for causing a computer to execute determination processing for determining whether the input sound of each unit section is speech or non-speech based on the first index value and the second index value of the unit section.

Images