JP2012178679A - Sound processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To maintain the low frequency range intensities of signals after sound source separation.SOLUTION: An observation signal x1(t) and an observation signal x2(t) resulting from picking-up of a mixture of a sound S1 from a sound source PS1 and a sound S2 from a sound source PS2 by a sound pickup PM1 and a sound pickup PM2 are supplied to a sound processing device. A direction identifying unit 62 identifies the arriving direction θe1 of the sound S1 and the arriving direction θe2 of the sound S2. An orientation processor 64, by executing on the observation signal x1(t) and the observation signal x2(t) dead angle control type beam formation that forms a dead angle of sound picking-up in the arriving direction θei (i=1, 2), generates an orientation component Zi[n, u]. A coefficient value generator 66 generates as a processing coefficient value αi[n, u] the ratio of the amplitude of the orientation component Zi[n, u] to the amplitude sum of an orientation component Z1[n, u] and an orientation component Z2[n, u].

Description

  The present invention relates to a technique for emphasizing (separating or extracting) sound from a specific sound source among a plurality of mixed sound generated by different sound sources.

  Sound source separation technology has been proposed that separates sound from each sound source by performing sound source separation on multiple observation signals obtained by collecting multiple sounds such as voice and noise with multiple sound collection devices. Yes. The separation matrix (inverse mixing matrix) applied to sound source separation is calculated for each frequency by learning processing (iterative update) using, for example, frequency-domain independent component analysis (FDICA). .

  In Patent Literature 1 and Non-Patent Literature 1, a separation matrix is generated by learning processing using a plurality of observation signals for frequencies selected from a plurality of frequencies under a predetermined condition, and the separation matrix after learning processing is used. A technique for supplementing the separation matrix of non-selected frequencies is disclosed. For example, blind angle control type beam forming (NBF (Null Beam Former)) is used to generate a separation matrix of non-selected frequencies. That is, a separation matrix of non-selected frequencies is generated such that a dead angle of sound collection is formed in the sound arrival direction estimated from the separation matrix after the learning process.

JP 2010-117653 A

Osako et al., “Fast method of blind source separation using frequency band interpolation by blind spot control type beamformer”, Proceedings of the Acoustical Society of Japan, Acoustical Society of Japan, March 2007, p.549-p.550

  However, the above technique has a problem that the intensity at the low frequency side of the signal after the sound source separation is lowered due to blind spot control type beam formation. In view of the above circumstances, an object of the present invention is to maintain the low-frequency side strength of a signal after sound source separation.

  Means employed by the present invention to solve the above problems will be described. In order to facilitate the understanding of the present invention, in the following description, the correspondence between the elements of the present invention and the elements of the embodiments described later will be indicated in parentheses, but the scope of the present invention will be exemplified in the embodiments. It is not intended to be limited.

  The sound processing apparatus of the present invention is configured to output mixed sound of sounds (for example, sound S1 and sound S2) coming from a plurality of sound sources (for example, sound source PS1 and sound source PS2) to a plurality of sound collection devices (for example, sound collection device PM1 and sound collection device). PM2) is an acoustic processing device for processing a plurality of observation signals (for example, observation signal x1 (t) and observation signal x2 (t)) collected by sound, and for each of the plurality of sound sources, the direction of arrival of sound (for example, the arrival direction) direction specifying means (for example, direction specifying unit 62) for specifying θe1 and direction of arrival θe2), and for each of a plurality of arrival directions specified by the direction specifying means, a blind spot control type beam that forms a dead angle of sound collection in the arrival direction. Directional processing means (for example, directional processing unit 64) for generating directional signals (for example, directional component Z1 [n, u] and directional component Z2 [n, u]) by executing formation on a plurality of observation signals; The compound generated by the means Coefficient value generation means (for example, coefficient value generation) that generates a processing coefficient value (for example, processing coefficient value αi [n, u]) corresponding to the ratio of the amplitude of one directional signal to the sum of the directional signal amplitudes Unit 66) and first signal processing means (for example, signal processing unit 68) for applying a processing coefficient value of the frequency to each frequency component of the observation signal.

  In the above embodiment, since the processing coefficient value corresponding to the ratio of the amplitude of one directional signal to the sum of the amplitudes of the plurality of directional signals is applied to the observation signal, for example, the directional signal generated by the directional processing means is separated from the sound source. As compared with the configuration determined as the acoustic signal, it is possible to maintain the low-side intensity of the signal after the sound source separation.

  In a preferred embodiment of the present invention, a plurality of frequencies (for example, K frequencies F [1] to F [K]) are set as a first frequency (for example, M first frequencies FA [1] to FA [M]). Frequency selecting means (for example, frequency selecting unit 24) for selecting the second frequency (for example, N second frequencies FB [1] to FB [N]), and separating from the components of the first frequencies in the plurality of observation signals Separation matrix generation means (for example, a separation matrix generation unit 54) for generating a matrix (for example, a separation matrix W [m]) for each first frequency, and separation of the first frequency into each first frequency component in a plurality of observation signals Second signal processing means (for example, signal processing unit 52) for generating a first separated signal (for example, separated component YA1 [m, u] or separated component YA2 [m, u]) by applying a matrix; The processing means generates a directional signal for each direction of arrival from each second frequency component in the plurality of observation signals, and generates a coefficient value The generating means generates a processing coefficient value for each second frequency, and the first signal processing means performs the second separation by applying the processing coefficient value of the second frequency to each second frequency component in the plurality of observation signals. A component (eg, separated component YB1 [n, u] or separated component YB2 [n, u]) is generated. In the above aspect, the first separated signal is generated by processing the observation signal using the separation matrix for the first frequency among the plurality of frequencies, and the observation using the processing coefficient value for the second frequency among the plurality of frequencies. A second separated signal is generated by processing the signal. Therefore, there is an advantage that the calculation amount and necessary storage capacity of the sound processing device can be reduced as compared with the configuration in which the separation matrix is generated for all frequencies while maintaining the low-frequency side strength in the second separated signal.

  In a further preferred aspect, the direction specifying means estimates the sound arrival direction for each of the plurality of sound sources from the separation matrix generated by the separation matrix generation means for each first frequency. In the above aspect, since the direction of arrival of each sound is estimated from the separation matrix generated by the separation matrix generation unit for the first frequency, the direction specifying unit separates each sound from the generation of the separation matrix by the separation matrix generation unit. Compared with the configuration in which the direction of arrival is estimated, there is an advantage that the calculation amount and necessary storage capacity of the sound processing device are reduced.

  The acoustic processing device according to a preferred aspect of the present invention is an index calculation means (for example, index calculation) that calculates a significant index value indicating the significance of learning processing for generating a separation matrix from components of each frequency in a plurality of observation signals for each frequency. And a frequency sorting means sorts a plurality of frequencies into a first frequency and a second frequency according to a significant index value of each frequency. In the above aspect, since the plurality of frequencies are selected according to the significance index value indicating the significance of the learning process (for example, independent component analysis), the plurality of frequencies are selected regardless of the significance of the learning process. In comparison, it is possible to generate a separation matrix capable of highly accurate sound source separation.

  The acoustic processing apparatus of each aspect described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to voice processing, and a general-purpose arithmetic processing apparatus such as a CPU (Central Processing Unit). This is also realized through collaboration with programs. The program according to the present invention converts a mixed sound of sounds (for example, sound S1 and sound S2) coming from a plurality of sound sources (for example, sound source PS1 and sound source PS2) to a plurality of sound collection devices (for example, sound collection device PM1 and sound collection device PM2). ) Is a program for processing a plurality of observation signals (for example, observation signal x1 (t) and observation signal x2 (t)), and the sound arrival direction (for example, arrival direction θe1) for each of the plurality of sound sources. And a direction-of-arrival θe2) direction specifying process (for example, the direction specifying unit 62) and a blind spot control type beam forming that forms a dead angle of sound collection in the direction of arrival for each of a plurality of directions of arrival specified by the direction specifying process. Is executed for a plurality of observation signals to generate a directional signal (for example, directional component Z1 [n, u] and directional component Z2 [n, u]) and a directional process to generate Multiple Coefficient value generation processing (for example, coefficient value generation unit) that generates a processing coefficient value (for example, processing coefficient value αi [n, u]) corresponding to the ratio of the amplitude of one directional signal to the sum of the directional signal amplitude values 66) and the first signal processing (for example, the signal processing unit 68) for causing the processing coefficient value of the frequency to act on each frequency component of the observation signal. According to the above program, the same operation and effect as the sound processing apparatus according to 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 sound processing apparatus according to a first embodiment. It is a block diagram of a 1st sound source separation part and a 2nd sound source separation part. It is a block diagram of an orientation processing unit. It is an amplitude spectrum before and after sound source separation. It is a graph which shows the relationship between the number of the 1st frequency used as the object of a learning process, and the precision (segmental SNR) of sound source separation. It is a graph which shows the relationship between the number of the 1st frequency used as the object of a learning process, and the precision (SIR variation | change_quantity) of sound source separation. It is a block diagram of the sound processing apparatus of 2nd Embodiment.

<A: First Embodiment>
FIG. 1 is a block diagram of a sound processing apparatus 100A according to the first embodiment. The sound collection device PM1 and the sound collection device PM2 that are arranged at a distance from each other are connected to the sound processing apparatus 100A. The sound collection device PM1 and the sound collection device PM2 are non-directional or directional microphones, for example. The sound source PS1 and the sound source PS2 exist at different positions around the sound collection device PM1 and the sound collection device PM2. The sound source PS1 is located in the direction θ1 with respect to the observation point (for example, the midpoint between the sound collection devices PM1 and PM2), and the sound source PS2 is located in the direction θ2 with respect to the observation point.

  The mixed sound of the sound S1 generated by the sound source PS1 and the sound S2 generated by the sound source PS2 reaches the sound collection device PM1 and the sound collection device PM2. The sound collection device PM1 generates an observation signal x1 (t), and the sound collection device PM2 generates an observation signal x2 (t). Each of the observation signal x1 (t) and the observation signal x2 (t) is an acoustic signal representing a time waveform of a mixed sound of the sound S1 and the sound S2 (t: time).

  The acoustic processing device 100A is a signal processing device that generates a separated signal y1 (t) and a separated signal y2 (t) by sound source separation for the observed signal x1 (t) and the observed signal x2 (t). The separated signal y1 (t) is an acoustic signal that emphasizes the sound S1 (suppresses the sound S2), and the separated signal y2 (t) is an acoustic signal that emphasizes the sound S2 (suppresses the sound S1). That is, the sound S1 and the sound S2 are separated (sound source separation).

  The separated signal y1 (t) and the separated signal y2 (t) are reproduced as sound by being supplied to a sound emitting device (not shown) such as a speaker or a headphone. A configuration that generates only one of the separated signal y1 (t) and the separated signal y2 (t) (for example, a configuration that discards the separated signal y2 (t) as noise) is also employed. Also, an A / D converter that converts the observation signal x1 (t) and the observation signal x2 (t) from analog to digital, and a D that converts the separation signal y1 (t) and separation signal y2 (t) from digital to analog. The illustration of the / A converter is omitted for convenience.

  As shown in FIG. 1, the sound processing device 100 </ b> A is realized by a computer system including an arithmetic processing device 12 and a storage device 14. The storage device 14 stores programs executed by the arithmetic processing device 12 and various types of information used by the arithmetic processing device 12. A known recording medium such as a magnetic recording medium or a semiconductor recording medium or a combination of a plurality of types of recording media is arbitrarily employed as the storage device 14. A configuration in which the observation signal x1 (t) and the observation signal x2 (t) are recorded in advance and stored in the storage device 14 (therefore, the sound collection device PM1 and the sound collection device PM2 are omitted) is also preferable.

  The arithmetic processing unit 12 executes a program stored in the storage device 14 to thereby execute a plurality of elements (frequency analysis unit 22, frequency selection unit 24, index calculation unit 26, first sound source separation unit 31, second sound source separation unit). 32, a frequency integration unit 42, and a waveform synthesis unit 44). A configuration in which each function of the arithmetic processing unit 12 is distributed over a plurality of integrated circuits, or a configuration in which a dedicated electronic circuit (DSP) realizes each function may be employed.

The frequency analysis unit 22 divides the observation signal x1 (t) into K frequency components X1 [k, u] (X1) obtained by dividing the observation signal x1 (t) for each frequency F [k] (k = 1 to K) (for each band) on the frequency axis. [1, u] to X1 [K, u]) and K frequency components X2 [k, u] (X2 [1, u] to X) obtained by dividing the observation signal x2 (t) for each frequency F [k] X2 [K, u]) are sequentially generated for each unit period (frame). Symbol k is a variable indicating each frequency on the frequency axis, and symbol u is a variable indicating each time point on the time axis (for example, a unit period number). For generating each frequency component X1 [k, u] and each frequency component X2 [k, u], for example, known frequency analysis such as short-time Fourier transform is arbitrarily employed. Further, K band pass filters (filter banks) having different pass bands can be used as the frequency analysis unit 22. Observation vector Xv [k, u] (Xv [k, u] = [X1 [k, u], X2 [k, u] having frequency component X1 [k, u] and frequency component X2 [k, u] as elements) u]] ^T ) are sequentially stored in the storage device 14. Note that the symbol T means transposition of a matrix.

  The frequency selection unit 24 selects the K frequencies F [1] to F [K] from the M first frequencies FA [1] to FA [M] and the N second frequencies FB [1] to FB [N]. ] For each unit period (M and N are natural numbers, K = M + N). Of the K frequency components X1 [1, u] to X1 [K, u] generated by the frequency analysis unit 22, M frequency components XA1 [of the first frequencies FA [m] (m = 1 to M) are included. 1, u] to XA1 [M, u] are supplied to the first sound source separation unit 31, and N frequency components XB1 [1, u] to the second frequencies FB [n] (n = 1 to N) are provided. XB1 [N, u] is supplied to the second sound source separation unit 32. Similarly, M frequency components XA2 [1, u] to XA2 [M, u] of each first frequency FA [m] among K frequency components X2 [1, u] to X2 [K, u]. Is supplied to the first sound source separation unit 31, and N frequency components XB2 [1, u] to XB2 [N, u] of each second frequency FB [n] are supplied to the second sound source separation unit 32.

The index calculation unit 26 sets the significant index value σ [k] (σ [1] to σ [K]), which serves as a reference for selection of the frequency F [k] by the frequency selection unit 24, to K frequencies F [1] to Calculate for each of F [K]. The calculation of the significant index value σ [k] is executed at a predetermined cycle (for example, every predetermined unit period). The significant index value σ [k] is a numerical value that is a measure of the significance of learning processing (independent component analysis) that generates a separation matrix from the observation vector Xv [k, u] of the frequency F [k]. The index calculation unit 26 according to the first embodiment uses a time-series covariance matrix Rxx [k] (Rxx [k] = E [) of the observation vector Xv [k, u] of the frequency F [k] over a predetermined number of unit periods. Xv [k, u] Xv [k, u] ^H ]) is calculated as a significant index value σ [k] of the frequency F [k]. The symbol H means conjugate transpose of the matrix, and the symbol E [] means an average value (expected value) or an added value over a predetermined number of unit periods. The calculation of the determinant of the covariance matrix Rxx [k] is described in detail in Patent Document 1.

  The frequency selecting unit 24 selects the K frequencies F [1] to F [K] from the M first frequencies FA [1] to FA according to each significant index value σ [k] calculated by the index calculating unit 26. [M] and N second frequencies FB [1] to FB [N] are selected. As disclosed in Patent Document 1, the smaller the determinant of the covariance matrix Rxx [k], the more significant the matrix processing to which the observation vector Xv [k, u] is applied (the accuracy of sound source separation before and after the learning processing). Tend to be low). Therefore, the frequency selection unit 24 has M significant sign values [sigma] [k] among the K frequencies F [1] to F [K] (for example, the top M numbers in descending order of the significant index value [sigma] [k]). Or the frequency F [k] of which the significant index value σ [k] exceeds a predetermined threshold) is selected as the first frequencies FA [1] to FA [M], and the significant index value σ [k] is small N The individual frequencies F [k] are selected as the second frequencies FB [1] to FB [N].

  The first sound source separation unit 31 in FIG. 1 performs each sound source separation by applying independent component analysis to each frequency component XA1 [m, u] and each frequency component XA2 [m, u]. M separation components YA1 [1, u] to YA1 [M, u] and M separation components YA2 [1, u] to YA2 [M, u] corresponding to the frequency FA [m] for each unit period. To generate. The separated component YA1 [m, u] is a frequency component in which the component of the sound S1 at the first frequency FA [m] is emphasized (the sound S2 is suppressed), and the separated component YA2 [m, u] is the second frequency FB. This is a frequency component in which the component of the sound S2 in [m] is emphasized (the sound S1 is suppressed).

  The second sound source separation unit 32 performs the signal processing different from that of the first sound source separation unit 31 on each frequency component XB1 [n, u] and each frequency component XB2 [n, u], so A unit period of N separated components YB1 [1, u] to YB1 [N, u] and N separated components YB2 [1, u] to YB2 [N, u] corresponding to two frequencies FB [m] Generate every time. The separated component YB1 [n, u] is a frequency component in which the component of the sound S1 at the second frequency FB [n] is emphasized (the sound S2 is suppressed), and the separated component YB2 [n, u] is the second frequency FB. This is a frequency component in which the component of the sound S2 in [n] is emphasized (the sound S1 is suppressed).

  The frequency integration unit 42 includes M separation components YA1 [1, u] to YA1 [M, u] generated by the first sound source separation unit 31 and N separation components YB1 [generated by the second sound source separation unit 32. 1, u] to YB1 [N, u] are arranged (integrated) in the order of frequencies to generate K separation components Y1 [1, u] to Y1 [K, u] for each unit period. Similarly, the frequency integration unit 42 arranges M separated components YA2 [1, u] to YA2 [M, u] and N separated components YB2 [1, u] to YB2 [N, u]. K separation components Y2 [1, u] to Y2 [K, u] are generated for each unit period.

  The waveform synthesizer 44 generates a time-domain separated signal y1 (t) from the K separated components Y1 [1, u] to Y1 [K, u] generated by the frequency integrating unit 42 for each unit period. Specifically, the waveform synthesizer 44 converts the sequence (frequency spectrum) of the K separated components Y1 [1, u] to Y1 [K, u] into the time domain by inverse Fourier transform, and before and after the unit period. Are separated from each other to generate a separation signal y1 (t). Similarly, the waveform synthesizer 44 generates a separation signal y2 (t) from the K separation components Y2 [1, u] to Y2 [K, u] generated by the frequency integration unit 42 for each unit period.

FIG. 2 is a block diagram of the first sound source separation unit 31 and the second sound source separation unit 32. As shown in FIG. 2, the first sound source separation unit 31 includes a signal processing unit 52 and a separation matrix generation unit 54. The signal processing unit 52 applies the separation matrix W [m] of the first frequency FA [m] to the frequency component XA1 [m, u] and the frequency component XA2 [m, u] of each first frequency FA [m]. By doing so, the separation component YA1 [m, u] and the separation component YA2 [m, u] are generated. Specifically, the signal processing unit 52 performs calculation (sound source separation) of the following formula (1) for each of the M first frequencies FA [1] to FA [M].

The separation matrix generation unit 54 in FIG. 2 converts the separation matrix W [m] (W [1] to W [M]) that the signal processing unit 52 applies to the sound source separation of Expression (1) to M first frequencies FA. Each of [1] to FA [M] is generated for each unit period. For the generation of the separation matrix W [m], a learning process (cumulative update of the separation matrix W [m]) to which independent component analysis is applied is employed. Although a known technique can be arbitrarily employed for the learning process of the separation matrix W [m], the immediately following separation matrix W _{p + 1} [m] is calculated from the p-th updated separation matrix W _p [m]. The following equation (2) is preferable. A predetermined initial matrix W ₀ [m] (for example, a unit matrix) is applied to the calculation of the first separation matrix W ₁ [m].

The symbol η in Equation (2) means a predetermined constant (step size), and the symbol off-diag () means an operator that replaces the diagonal component with zero. The symbol φ [m, n] is a predetermined nonlinear function (for example, a hyperbolic tangent function). The symbol Yv _p [m, u] in Equation (2) is an operation of Equation (1) that causes the separation matrix W _p [m] to act on the frequency component XA1 [m, u] and the frequency component XA2 [m, u]. This means the calculated vector (Yv _p [m, u] = [YA1 [m, u], YA2 [m, u]] ^T ). The separation matrix generation unit 54 determines the separation matrix W _{p + 1} [m] at the point of time when the calculation of Expression (2) is repeated a predetermined number of times as the separation matrix W [m]. The above is the configuration and operation of the first sound source separation unit 31.

  As shown in FIG. 2, the second sound source separation unit 32 includes an arithmetic processing unit 60 and a signal processing unit 68. The arithmetic processing unit 60 processes the processing coefficient value α1 [n, u] (α1 [1, u] to α1 [N, u]) for each of the N second frequencies FB [1] to FB [N]. The coefficient value α2 [n, u] (α2 [1, u] to α2 [N, u]) is set. The calculation of the processing coefficient value α1 [n, u] and the processing coefficient value α2 [n, u] is executed every predetermined cycle (for example, every unit period).

  The processing coefficient value α1 [n, u] and the processing coefficient value α2 [n, u] of each second frequency FB [n] are the frequency components S1 [n, u of the second frequency FB [n] of the sound S1. ] | S1 [n, u] | and the amplitude | S2 [n, u] | of the frequency component S2 [n, u] of the second frequency FB [n] of the sound S2 (large or small) Accordingly, it is variably set within the range of 0 or more and 1 or less. Specifically, the processing coefficient value α1 [n, u] is set to a larger value as the amplitude | S1 [n, u] | of the sound S1 is larger than the amplitude | S2 [n, u] | The processing coefficient value α2 [n, u] is set to a larger numerical value as the amplitude | S2 [n, u] | is larger than the amplitude | S1 [n, u] |.

  The signal processing unit 68 in FIG. 2 applies the processing coefficient value α1 [of the second frequency FB [n] to the frequency component XB1 [n, u] and the frequency component XB2 [n, u] of each second frequency FB [n]. By applying n, u] and the processing coefficient value α2 [n, u], the separation component YB1 [n, u] and the separation component YB2 [n, u] are generated for each unit period. Specifically, the signal processing unit 68 performs the following calculations (3A) and (3B) for each of the N second frequencies FB [1] to FB [N].

  That is, by separating the frequency component XB1 [n, u] by the processing coefficient value α1 [n, u], a separated component YB1 [n, u] in which the frequency component S1 [n, u] of the sound S1 is emphasized is generated. A separated component YB2 [n, u] in which the frequency component S2 [n, u] of the sound S2 is emphasized is generated by multiplying the frequency component XB2 [n, u] by the processing coefficient value α2 [n, u]. Therefore, the processing coefficient value α1 [n, u] corresponds to a gain (spectral gain) for the frequency component XB1 [n, u], and the processing coefficient value α2 [n, u] is a gain for the frequency component XB2 [n, u]. It corresponds to.

  As shown in FIG. 2, the arithmetic processing unit 60 includes a direction specifying unit 62, a directing processing unit 64, and a coefficient value generating unit 66. The direction specifying unit 62 specifies the arrival direction of the sound S1 (the direction of the sound source PS1) θe1 and the arrival direction of the sound S2 (the direction of the sound source PS2) θe2. In the following description, the subscript “e” denotes an estimated value (estimate).

  The direction specifying unit 62 of the first embodiment arrives by using the separation matrix W [m] (W [1] to W [M]) generated by the separation matrix generation unit 54 for each first frequency FA [m]. The direction θe1 and the arrival direction θe2 are estimated. A known technique (for example, the method disclosed in Non-Patent Document 1) is arbitrarily employed for estimating the arrival direction θe1 and the arrival direction θe2. For example, the direction specifying unit 62 estimates the arrival direction θe1 [m] of the sound S1 and the arrival direction θe2 [m] of the sound S2 from the separation matrix W [m] of each first frequency FA [m], and M pieces. The representative values (for example, weighted sum, average value, and median) of the arrival directions θe1 [1] to θe1 [M] are determined as the arrival direction θe1 and are representative of the M arrival directions θe2 [1] to θe2 [M]. The value is determined as the direction of arrival θe2.

  The directivity processing unit 64 in FIG. 2 performs a process (hereinafter referred to as “dead angle control beam forming”) for forming a blind spot of sound collection (an area where the sensitivity of sound collection is low) in a predetermined direction as a frequency component XB1 [n, u]. And the directional component Z1 [n, u] (Z1 [1, u] to Z1 [N, u]) and the directional component Z2 [n, u] ( Z2 [1, u] to Z2 [N, u]) are generated for each unit period. Specifically, the directivity processing unit 64 performs blind angle control type beam forming (NBF) that forms a dead angle of sound collection in the arrival direction θe2 specified by the direction specifying unit 62, with frequency components XB1 [n, u] and frequency components XB2 [n, u] generates a directional component Z1 [n, u], and a blind spot control type beam forming that forms a dead angle of sound collection in the arrival direction θe1 is performed with a frequency component XB1 [n, u] and a frequency component. The directional component Z2 [n, u] is generated by executing it on XB2 [n, u]. Therefore, the directional component Z1 [n, u] suppresses the incoming sound (sound S2) from the arrival direction θe2, and the directional component Z2 [n, u] suppresses the incoming sound (sound S1) from the arrival direction θe1. .

  FIG. 3 is a block diagram of the directional processing unit 64. In FIG. 3, the sound S1 (frequency component S1 [n, u]) radiated by the sound source PS1 and the sound S2 (frequency component S2 [n, u]) radiated by the sound source PS2 are collected by the sound collecting device PM1 and the sound collecting device. A model of a propagation path to reach each of PM2 is shown for convenience.

  The symbol Ai [n] (i = 1, 2) in FIG. 3 means the propagation loss (gain gained in the propagation path) of the frequency component Si [n, u] of the acoustic Si. Note that the propagation delay of the frequency component Si [n, u] is omitted in FIG. 3 in consideration of the fact that it is reflected in the propagation loss Ai [n]. The symbol τi1 in FIG. 3 means a delay (time difference) from when the frequency component Si [n, u] reaches the sound collection device PM2 until it reaches the sound collection device PM1, and the symbol τi2 represents the frequency component Si [ n, u] means a delay from reaching the sound collection device PM1 to reaching the sound collection device PM2.

As understood from FIG. 3, the frequency component XB1 [n, u] and the frequency component XB2 [n, u] after sound collection by the sound collection device PMj are expressed by the following equations (4A) and (4B). Is done. In the equations (4A) and (4B), the symbol ω [n] means an angular frequency corresponding to the second frequency FB [n], and the symbol j means an imaginary unit.

  As shown in FIG. 3, the directional processing unit 64 includes a first processing unit 72 that generates a directional component Z1 [n, u] and a second processing unit 74 that generates a directional component Z2 [n, u]. . The first processing unit 72 includes a delay unit 721 that adds a delay τe22 to the frequency component XB1 [n, u], a delay unit 723 that adds a delay τe21 to the frequency component XB2 [n, u], a delay unit 721, and a delay And a calculation unit 725 that generates a difference between the outputs of the unit 723 as a directional component Z1 [n, u]. Similarly, the second processing unit 74 includes a delay unit 741 that applies a delay τe11 to the frequency component XB2 [n, u], a delay unit 743 that applies a delay τe12 to the frequency component XB1 [n, u], and a delay unit. 741 and a calculation unit 745 that generates a difference between the outputs of the delay unit 743 as a directional component Z2 [n, u]. The delay τeij is an estimated value of the delay τij given in the propagation path. The delay τe21 and the delay τe22 are set so that a sound collection dead angle is formed in the arrival direction θe2, and the delay τe11 and the delay τe12 are set so that a sound collection dead angle is formed in the arrival direction θe1.

As understood from FIG. 3, the directional component Z1 [n, u] and the directional component Z2 [n, u] are expressed by the following formulas (5A) and (5B).

By substituting Equation (4A) and Equation (4B) into Equation (5A) for transformation, the following Equation (6A) is derived. Similarly, the following formula (6B) is derived by substituting the formula (4A) and the formula (4B) into the formula (5B) for transformation.

Assuming that the accuracy of the estimation of the arrival direction θe1 and the arrival direction θe2 by the direction specifying unit 62 is sufficiently high (θe1≈θ1, θe2≈θ2), the delay τeij applied to the directivity processing unit 64 is actually propagated. It is possible to approximate by the delay τij in the path (τeij≈τij). Therefore, the second term and the fourth term on the right side of the formula (6A) are canceled and the following formula (7A) is derived, and the first and third terms on the right side of the formula (6B) are offset. The following formula (7B) is derived.

Now, the sum of the amplitude | Z1 [n, u] | of the directional component Z1 [n, u] and the amplitude | Z2 [n, u] | of the directional component Z2 [n, u] (hereinafter referred to as “amplitude sum”) The ratio of the amplitude | Z1 [n, u] | of the directional component Z1 [n, u] with respect to) is expressed as the following equation (8A) in consideration of the equations (7A) and (7B). Similarly, the ratio of the amplitude | Z2 [n, u] | of the directivity component Z2 [n, u] to the amplitude sum is expressed as in the following formula (8B).

  Of the equations (7A) and (7B), the delay term related to the delay (phase) τij (the parentheses in the latter half) is common to the directional component Z1 [n, u] and the directional component Z2 [n, u]. Therefore, the delay term is eliminated in Equation (8A) and Equation (8B).

The coefficient value generation unit 66 in FIG. 2 has a ratio of the amplitude | Z1 [n, u] | of the directional component Z1 [n, u] to the sum of amplitudes as shown in the following equations (9A) and (9B) ( Equation (8A)) is calculated for each second frequency FB [n] as the processing coefficient value α1 [n, u], and the amplitude | Z2 [n, u] | of the directional component Z2 [n, u] with respect to the sum of amplitudes is calculated. The ratio (formula (8B)) is calculated for each second frequency FB [n] as the processing coefficient value α2 [n, u].

  As understood from the equations (8A) and (8B), the equations (9A) and (9B), the processing coefficient value α1 [n, u] and the processing coefficient value α2 [n, u] Corresponds to the internal ratio of the amplitudes of the sound S1 and the sound S2 for each unit period (contribution of each sound source PSi to each of the observation signal x1 (t) and the observation signal x2 (t)). That is, the amplitude ratio of the sound S1 at the observation point can be expressed by the processing coefficient value α1 [n, u], and the amplitude ratio of the sound S2 at the observation point can be expressed by the processing coefficient value α2 [n, u]. For example, processing is performed when the amplitude of the acoustic S1 at the observation point (A1 [n] | S1 [n, u] |) and the amplitude of the acoustic S2 (A2 [n] | S2 [n, u] |) are equal. The coefficient value α1 [n, u] and the processing coefficient value α2 [n, u] are 0.5, and the amplitude of the acoustic S1 (A1 [n] | S1 [n, u] |) is the amplitude of the acoustic S2 (A2 [A2 [n, u] | n] | S2 [n, u] |), the processing coefficient value α1 [n, u] exceeds the processing coefficient value α2 [n, u]. Therefore, the processing coefficient value α1 [n, u] in the equation (9A) and the processing coefficient value α2 [n, u] in the equation (9B) are variables representing the amplitude ratio between the sound S1 and the sound S2 at the observation point. It is reasonable.

  Since the processing coefficient value α1 [n, u] and the processing coefficient value α2 [n, u] are set as described above, Formula (3A) to which the processing coefficient value α1 [n, u] of Formula (9A) is applied. In the separation component YB1 [n, u] generated by the signal processing unit 68 in the calculation of (2), the frequency component S1 [n, u] of the sound S1 is emphasized, and the processing coefficient value α2 [n, u] of Equation (9B) is applied. In the separated component YB2 [n, u] generated by the calculation of the mathematical formula (3B), the frequency component S2 [n, u] of the sound S2 is emphasized. That is, for each of the N second frequencies FB [1] to FB [N], the sound S1 (frequency component S1 [n, u]) and the sound S2 (frequency component S1 [n, u]) are separated. .

  Part (A) in FIG. 4 is the amplitude spectrum of the acoustic Si radiated from the sound source PSi, and part (C) in FIG. 4 is the amplitude spectrum of the separated signal yi (t) generated in the configuration of the first embodiment. It is. Part (B) of FIG. 4 has a configuration in which the directional component Zi [n, u] generated by the blind spot control type beam forming by the directional processing unit 64 is a separated component YBi [n, u] (hereinafter referred to as “comparative”). ) Is an amplitude spectrum of the separated signal yi (t) generated.

^Each term (e ^{−jω [n] (τ11 + τ12)} , e ^{−jω [n] (τ12 + τ21)} ) of the expression (7A) and the expression (7B) indicating the directional component Zi [n, u] Since the smaller the angular frequency ω [n], the closer to 1, the smaller the angular frequency ω [n], the closer to zero the delay terms in the equations (7A) and (7B). Therefore, the directional component Zi [n, u] is suppressed toward the lower frequency side. That is, in the comparative configuration in which the separated signal yi (t) is generated with the directional component Zi [n, u] as the separated component YBi [n, u], as can be understood from the part (B) in FIG. There is a problem that the intensity (amplitude) of the low frequency side (particularly 0 Hz to 500 Hz) of the separated signal yi (t) is suppressed as compared with the original sound Si (part (A)).

  On the other hand, in the first embodiment, the processing coefficient value αi [n, u] calculated from the amplitudes of the directional component Z1 [n, u] and the directional component Z2 [n, u] is used as the frequency component XBi [n, u]. The separated component YBi [n, u] is generated by the action. As described above, in the processing coefficient value α1 [n, u] and the processing coefficient value α2 [n, u], the influence of the delay term in Expression (7A) and Expression (7B) is eliminated. As can be seen from C), the intensity of the low frequency side in the separated signal yi (t) can be maintained equal to that of the sound Si. That is, according to the first embodiment, there is an advantage that sound source separation with higher accuracy can be realized (each sound Si can be faithfully extracted) compared with the comparative example.

  Further, as understood from Equation (9A) and Equation (9B), the processing of the second sound source separation unit 32 (calculation of processing coefficient value αi [n, u] and calculation of Equation (3A) and Equation (3B)) ) Is less loaded than the process of the first sound source separation unit 31 (the process of generating the separation matrix W [m] by repeating the learning process). Therefore, according to the first embodiment in which the number M of the first frequencies FA [m] to be subjected to the learning process can be reduced, the processing load (power consumption) of the arithmetic processing unit 12 without reducing the sound source separation performance. There is an advantage that the storage capacity required for the storage device 14 can be reduced. The above effects are particularly advantageous when the sound processing device 100A is mounted on a portable information terminal (for example, a mobile phone) in which the performance, power supply capacity, and storage capacity of the arithmetic processing device 12 are restricted.

The number M of the first frequencies FA [m] to be subjected to learning processing and the accuracy of sound source separation will be described in detail below. In the following description, the observation signal xi (t) is expressed as in the following equations (10A) and (10B), and the separated signal yi (t) is expressed as in the following equations (11A) and (11B). Express as follows. The symbols xij (t) and yij (t) mean acoustic components that arrive at the sound collection device PMj from the sound source PSi.

  5 and 6 show the relationship between the number M (horizontal axis) of the first frequencies FA [m] for generating the separation matrix W [m] in the independent component analysis learning process and the evaluation index (vertical axis) of the sound source separation. It is a graph which shows. The symbol “FDICA” on the horizontal axis in FIGS. 5 and 6 represents the case where all of the K frequencies F [1] to F [K] (for example, K = 513) are selected as the first frequency FA [m] (that is, , A configuration in which the second sound source separation unit 32 is omitted. In FIGS. 5 and 6, for each of the first embodiment (solid line) and the proportionality (broken line), the result of processing the observation signal xi (t) recorded in the anechoic room and the reverberation time are shown. The result of processing the observation signal xi (t) recorded in the acoustic room of 500 milliseconds is also shown.

In FIG. 5, a segmental signal-to-noise ratio (SegSNR) after sound source separation is shown on the vertical axis as an evaluation index for sound source separation. The segmental SNR after the sound source separation is expressed by the following formula (12). The symbol xij (h, u) in Equation (12) represents the signal value (amplitude) at the time point h within the u-th unit period among the acoustic components xij (t) in Equation (10A) and Equation (10B). means. The symbol yi (h, u) in the equation (12) means a signal value (amplitude) at the time point h within the u-th unit period of the separated signal yi (t) after the sound source separation. As understood from the equation (12), it can be evaluated that the higher the segmental SNR after the sound source separation (that is, the closer the separated signal yi (t) is to the acoustic Si at the observation point), the higher the sound source separation accuracy.

  As can be seen from FIG. 5, in the comparative configuration, the accuracy of sound source separation (segmental SNR) decreases as the number M of the first frequencies FA [m] to be subjected to learning processing decreases. In the first embodiment, sufficiently accurate sound source separation is realized even when the number M of the first frequencies FA [m] is reduced. It can be understood from FIG. 5 that the first embodiment has higher accuracy of sound source separation even when compared with the case where all K pieces are selected as the first frequency FA [m] (FDICA).

On the other hand, in FIG. 6, a change amount ΔSIR of SIR (Signal-to-Interference Ratio) before and after sound source separation is shown on the vertical axis as an evaluation index of sound source separation. SIR _in before the sound source separation is expressed by the following formula (13A), and SIR _out after the sound source separation is expressed by the following formula (13B). The acoustic component x21 (t) and the acoustic component x12 (t) (Equation (10A) and Equation (10B)) of the equation (13A) and the acoustic component y21 (t) and the acoustic component y12 (t) (Equation (13B)) Expressions (11A) and (11B)) correspond to interference components (interfering sounds).

The change amount ΔSIR shown on the vertical axis in FIG. 6 corresponds to a difference value (ΔSIR = SIR _out −SIR _in ) between SIR _in before sound source separation and SIR _out after sound source separation. Therefore, it can be evaluated that the greater the change amount ΔSIR, the higher the accuracy of sound source separation. As can be seen from FIG. 6, in both the first embodiment and the comparative example, the accuracy (variation ΔSIR) of the sound source separation increases as the number M of the first frequencies FA [m] to be subjected to learning processing decreases. descend. The above tendency is particularly remarkable in an environment where reverberation occurs.

  As described above, in the first embodiment, the accuracy of the sound source separation evaluated from the viewpoint of the segmental SNR and the accuracy of the sound source separation evaluated from the viewpoint of the SIR (variation amount ΔSIR) are the first frequency FA [m ] Are mutually contradictory to the number M. Therefore, by selecting the number M of the first frequencies FA [m] so that the segmental SNR in FIG. 5 and the SIR variation ΔSIR in FIG. Sound source separation can be realized.

  For example, in the first embodiment, the smaller the number M of the first frequencies FA [m], the higher the segmental SNR. Therefore, from the viewpoint of improving the segmental SNR and from the viewpoint of reducing the processing load (power consumption) of the arithmetic processing device 12 and the capacity of the storage device 14, the number M of the first frequencies FA [m] is reduced. Is more advantageous. On the other hand, when the number M of the first frequencies FA [m] is extremely reduced, there is a possibility that the decrease in the change amount SIR of the SIR becomes obvious, but the number M of the first frequencies FA [m] is From FIG. 6, it can be seen from FIG. 6 that the decrease in the amount of change SIR ΔSIR does not become apparent if the frequency F [k] is in a range that exceeds approximately ¼ (M = 128) of the total number K (K = 513) of the frequency F [k]. . Further, when the number M of the separation matrix M [m] is extremely small, the estimation accuracy of the arrival direction θe1 and the arrival direction θe2 is lowered, but the number M is about 1/4 of the total number K of the frequencies F [k]. If so, the arrival direction θe1 and the arrival direction θe2 can be estimated with sufficient accuracy. Considering the above tendency, a configuration in which the number M of the first frequencies FA [m] is set to about 25% (for example, 20% to 30%) of the total number K of the frequencies F [k] is particularly suitable.

<B: Second Embodiment>
A second embodiment of the present invention will be described below. In the first embodiment, the sound source separation using the independent component analysis (first sound source separation unit 31) and the sound source separation using the blind spot control type beam forming (second sound source separation unit 32) are used in combination. In the form, sound source separation by independent component analysis is omitted. In addition, about the element in which an effect | action and a function are equivalent to 1st Embodiment in each structure illustrated below, the detailed description of each is abbreviate | omitted suitably using the code | symbol referred by the above description.

  FIG. 7 is a block diagram of the sound processing apparatus 100B of the second embodiment. As shown in FIG. 7, the sound processing apparatus 100B of the second embodiment has a configuration in which the frequency selection unit 24, the index calculation unit 26, the first sound source separation unit 31, and the frequency integration unit 42 in the first embodiment are omitted. A frequency analysis unit 22, a sound source separation unit 35, and a waveform synthesis unit 44. Similarly to the first embodiment, the frequency analysis unit 22 uses the K frequency components X1 [k, u] (X1 [1, u] to X1 [K, u]) of the observation signal x1 (t) and the observation signal. X frequency components X2 [k, u] (X2 [1, u] to X2 [K, u]) of x2 (t) are generated.

  Similar to the second sound source separation unit 32 of the first embodiment, the sound source separation unit 35 includes the direction specifying unit 62, the directivity processing unit 64, the coefficient value generation unit 66, and the signal processing unit 68 of FIG. By performing sound source separation using blind spot control type beam forming for the frequency component X1 [k, u] and each frequency component X2 [k, u], K separated components Y1 [k, u] (Y1 [ 1, u] to Y1 [K, u]) and K separated components Y2 [k, u] (Y2 [1, u] to Y2 [K, u]) are generated for each unit period. That is, the operation of the sound source separation unit 35 of the second embodiment is to select all of the K frequencies F [1] to F [K] into the second frequencies FB [1] to FB [N] in the first embodiment. In this case (N = K), the operation is the same as that of the second sound source separation unit 32. Similarly to the first embodiment, the waveform synthesizer 44 generates a separation signal y1 (t) from the K separation components Y1 [1, u] to Y1 [K, u], and K separation components Y2 [ The separated signal y2 (t) is generated from 1, u] to Y2 [N, u]. In the second embodiment, the same effect as in the first embodiment is realized.

  Although the direction specifying unit 62 of the first embodiment uses the separation matrix W [m] for estimating the arrival direction θe1 and the arrival direction θe2, the direction specifying unit 62 of the second embodiment uses the arrival direction θe1 and the arrival direction θe2. A known technique is arbitrarily adopted as a method for identifying the above. For example, the direction specifying unit 62 may be an Ema Takuro and Nozomu Hamada, “FDICA using Time-Frequency Cell Selection for Blind Source Separation”, 2005 RISP International Worksyop on Nonlinear Circuit and Signal Processing (NCSP'05), p.471-474, etc. The direction of arrival θe1 and the direction of arrival θe2 are estimated from each frequency component X1 [k, u] and each frequency component X2 [k, u]. Further, the separation matrix generation unit 54 of the first embodiment is added to the second embodiment, and the direction specification unit 62 uses the separation matrix W [m] generated by the separation matrix generation unit 54 in the same manner as in the first embodiment. A configuration that estimates the arrival direction θe1 and the arrival direction θe2 (that is, a configuration that uses the separation matrix W [m] only for estimation of the arrival direction θe1 and the arrival direction θe2) is also employed.

<C: Modification>
Various modifications are added to the above embodiments. Specific modifications are exemplified below. Two or more aspects arbitrarily selected from the following examples can be appropriately combined.

(1) Modification 1
In each of the above-described embodiments, the method for specifying the arrival direction θe1 and the arrival direction θe2 by the direction specifying unit 62 is arbitrary. For example, a configuration in which a predetermined number of separation matrices W [m] is selected from the M number of separation matrices W [1] to W [M] generated by the separation matrix generation unit 54 and the arrival direction θe1 and the arrival direction θe2 are estimated. Can also be employed. For example, when the direction θ1 of the sound source PS1 and the direction θ2 of the sound source PS2 are known, the arrival direction θe1 and the arrival direction θe2 can be stored in the storage device 14 in advance. The direction specifying unit 62 functions as an element that acquires the arrival direction θe1 and the arrival direction θe2 from the storage device 14. It is also possible to set the arrival direction θe1 and the arrival direction θe2 in accordance with an instruction from the user (for example, an operation for designating a direction by operating the operation element).

(2) Modification 2
The significant index value σ [k] of the first embodiment is not limited to the determinant of the covariance matrix Rxx [k] of the observation vector Xv [k, u]. For example, various indexes (statistics) exemplified in Patent Document 1 can be adopted as the significant index value σ [k].

  For example, in consideration of the tendency that the significance of the learning process is higher as the frequency F [k] has a larger total number of bases in the distribution of the observed vector Xv [k, u], the covariance matrix Rxx of the observed vector Xv [k, u]. The index calculation unit 26 calculates the condition number of [k] as the significant index value σ [k], and the frequency selection unit 24 selects the first frequency FA with M frequencies F [k] having a small significant index value σ [k]. It is possible to sort into [k]. That is, the determinant and condition number of the covariance matrix Rxx [k] are used as an index of the total number of bases in the distribution of the observation vector Xv [k, u]. A configuration is also possible in which the trace of the covariance matrix Rxx [k] is calculated as the significant index value σ [k], and the frequency F [k] having a large significant index value σ [k] is selected as the first frequency FA [m]. Is preferred.

  The learning process of independent component analysis is a process of updating the separation matrix W [m] so that each signal after sound source separation is statistically independent. Therefore, the observation signal x1 (t) and the observation signal x2 (t) It can be evaluated that the significance of the learning process is higher as the frequency F [k] has a lower statistical correlation with. Considering the above tendency, an index of independence between the observed signal x1 (t) and the observed signal x2 (t) is suitable as the significant index value σ [k]. Examples of the independence index include cross-correlation and mutual information. The frequency selection unit 24 selects M frequencies F [k] having high independence between the observation signal x1 (t) and the observation signal x2 (t) (small cross-correlation and mutual information amount) as the first frequency FA. Select [m].

  Further, it can be evaluated that the significance of the learning process is higher as the number of types of sound (number of sound sources) included in the observation signal x1 (t) and the observation signal x2 (t) is larger. Considering the tendency (the central limit theorem) that the kurtosis (cartesis) of the intensity distribution of the observation signal x1 (t) and the observation signal x2 (t) decreases as the number of acoustic mixtures increases, the observation signal x1 (t) or The kurtosis of the intensity distribution (probability distribution) of the observation signal x2 (t) can be adopted as the significant index value σ [k]. The frequency selection unit 24 first selects M frequencies F [k] having a low kurtosis (a large number of acoustic mixtures) in one or both of the observation signal x1 (t) and the observation signal x2 (t). Sort to frequency FA [m].

  It is also possible to calculate a significant index value σ [k] from a plurality of types of indexes (for example, two or more types of indexes selected from the above examples). For example, a configuration is adopted in which the weighted sum of the above-described plural types of indices (for example, the determinant of the covariance matrix Rxx [k] and the trace) is calculated as the significant index value σ [k].

  However, the configuration (index calculation unit 26) that uses the significant index value σ [k] for selection between the first frequency FA [m] and the second frequency FB [n] can be omitted. Specifically, K frequencies F [k] can be selected regardless of the observation signal x1 (t) or the observation signal x2 (t). For example, the frequency F [k] (for example, odd-numbered frequency F [k]) selected from the K frequencies F [k] at a predetermined interval is selected as the first frequency FA [m] and the remaining frequency F is selected. A configuration is adopted in which [k] (for example, even-numbered frequency F [k]) is selected as second frequency FB [n]. In addition, the frequency F [k] with high significance of the learning process is known in advance from the situation such as the acoustic characteristics assumed for the observation signal x1 (t) and the observation signal x2 (t) and the contents of the learning process. Then, the frequency F [k] can be selected as the first frequency FA [m] and the remaining frequency F [k] can be selected as the second frequency FB [n].

(3) Modification 3
In each of the above embodiments, the configuration in which the sound Si from the two sound sources PSi (PS1, PS2) is collected by the two sound collecting devices PMj (PM1, PM2) is exemplified. The total number of devices PMj is changed as appropriate. However, the total number of sound collection devices PMj needs to be equal to or greater than the total number of sound sources PSi.

(4) Modification 4
For example, the sound processing device 100 (100A, 100B) receives the observation signal x1 (t) and the observation signal x2 (t) transmitted from a terminal device such as a mobile phone or a personal computer via a communication network such as the Internet. Can be employed. The sound processing apparatus 100 generates a separation signal y1 (t) and a separation signal y2 (t) from the observation signal x1 (t) and the observation signal x2 (t) in the same manner as in the first embodiment and the second embodiment to generate a terminal. Send to device. Configuration in which each frequency component X1 [k, u] and each frequency component X2 [k, u] are transmitted from the terminal device to the sound processing device 100 (the frequency analysis unit 22 is mounted on the terminal device and is added to the sound processing device 100) In which the separated components Y1 [k, u] and the separated components Y2 [k, u] are transmitted from the sound processing device 100 to the terminal device (the waveform synthesis unit 44 is mounted on the terminal device). And a configuration that is not mounted in the sound processing apparatus 100).

100A, 100B: Acoustic processing device, 12: Arithmetic processing device, 14: Storage device, 22: Frequency analysis unit, 24: Frequency selection unit, 26: Index calculation unit, 31: First sound source separation , 32... Second sound source separation unit, 35... Sound source separation unit, 42... Frequency integration unit, 44 ...... waveform synthesis unit, 52 ...... signal processing unit, 54 ...... separation matrix generation unit, 62 ...... Direction specifying unit, 64... Direction processing unit, 66... Coefficient value generation unit, 68... Signal processing unit, PS 1, PS 2 .. Sound source, PM 1, PM 2.

Claims (4)

An acoustic processing device that processes a plurality of observation signals obtained by collecting a mixed sound of sounds coming from a plurality of sound sources by a plurality of sound collecting devices,
Direction specifying means for specifying the direction of arrival of sound for each of the plurality of sound sources;
Directional processing means for generating a directional signal for each of the plurality of arrival directions specified by the direction specifying means by executing blind angle control type beam forming that forms a dead angle of sound collection in the arrival direction for the plurality of observation signals. When,
Coefficient value generating means for generating a processing coefficient value corresponding to the ratio of the amplitude of one directional signal to the sum of amplitudes of a plurality of directional signals generated by the directional processing means for each frequency;
A sound processing apparatus comprising: first signal processing means for causing a processing coefficient value of the frequency to act on each frequency component of the observation signal. Frequency sorting means for sorting a plurality of frequencies into a first frequency and a second frequency;
Separation matrix generating means for generating a separation matrix for each of the first frequencies from the components of the first frequencies in the plurality of observation signals;
Second signal processing means for generating a first separated signal by applying a separation matrix of the first frequency to each first frequency component in the plurality of observation signals;
The directional processing means generates the directional signal for each arrival direction from the components of the second frequencies in the plurality of observation signals,
The coefficient value generating means generates the processing coefficient value for each of the second frequencies,
The acoustic processing device according to claim 1, wherein the first signal processing unit generates a second separated component by applying the processing coefficient value of the second frequency to the component of the second frequency in the plurality of observation signals. The acoustic processing apparatus according to claim 2, wherein the direction specifying unit estimates an arrival direction of sound for each of the plurality of sound sources from the separation matrix generated by the separation matrix generation unit for each first frequency. Comprising index calculation means for calculating, for each frequency, a significant index value indicating the significance of learning processing for generating a separation matrix from components of each frequency in the plurality of observation signals;
The acoustic processing apparatus according to claim 2 or 3, wherein the frequency sorting unit sorts the plurality of frequencies into a first frequency and a second frequency according to a significant index value of each frequency.

Images