JP2007047427A - Sound processor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a disturbing sound suppressing filter which is applicable even to a target sound section and has small distortion of an output sound. <P>SOLUTION: A sound processor has a disturbing sound generation section which decides whether each band component of an input signal is a disturbing sound or target sound based upon sparse property of a sound, extracts only a band whose component is decided as a disturbing sound, and generates a disturbing sound signal. Further, the sound processor has an adaptive processing section characterized in updating a spatial correlation inverse matrix by using the generated disturbing sound signal. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention belongs to a sound source separation technique for restoring only a target sound, for example, from a signal obtained by mixing voice, music, and various noises observed with a plurality of microphone elements.

  Conventionally, a minimum dispersion beamformer has been used as a sound source separation technology that uses multiple microphones to suppress the interference sound by creating a blind spot in the direction of the interference sound and emphasize and extract only the target sound by creating a beam in the direction of the target sound. The law is known. (For example, refer nonpatent literature 1). In the minimum variance beamformer method, the correlation matrix between microphones, called the spatial correlation inverse matrix, is updated from the input signal, and a linear filter that creates a blind spot in the direction of the disturbing sound and a beam in the target sound direction using the spatial correlation inverse matrix. Is generated. In the minimum variance beamformer, if the spatial correlation inverse matrix is updated in the speech section in which the target sound exists, the target sound is erroneously suppressed along with the interference sound. Had to be done only in the section where there is no. For this reason, it is necessary to extract a speech section in which no target sound exists in advance by some method. However, it is difficult to extract a speech section where the target sound does not exist in some way in advance, and if the spatial correlation inverse matrix is updated only in a section where the target sound does not exist, it suddenly occurs in the section where the target sound exists. The generated interference sound cannot be suppressed.

Also, instead of using an adaptive filter, a sound source signal for determining which sound source is input from which of the band-divided output channel signals of that band is based on the parameter value difference between channels for each band. There is a disturbing sound suppression technique that suppresses a detection signal from a sound source that has a determination process and does not sound. (For example, Patent Document 1)
This interference noise suppression technique is a technique based on the voice property of voice sparseness. The sparseness of speech is a property in which the frequency components that exist in a short time of about 40 ms are limited, and the probability that different sound sources hold the same frequency component in a short time is low. Therefore, sound source separation becomes possible by assigning frequency components to any one sound source for each time and frequency. However, in this interference noise suppression technique, when the sound sources hold the same frequency component in a short time, the interference noise suppression performance deteriorates and the sound is easily distorted.

Japanese Patent Laid-Open No. 10-313497

Toshiro Oga, Yoshio Yamazaki, Yutaka Kaneda, “Acoustic systems and digital processing,” IEICE, 1995. Masato Togami, Akio Amano, "Sound source separation method using sound source localization based on sound source duplication determination," Proc. Of the Acoustical Society of Japan Vol.1, pp.441-442, March 2005

In the minimum dispersion beamformer method, the spatial correlation inverse matrix cannot be updated in the section where the target sound exists, and the disturbing sound suddenly generated in the section where the target sound exists cannot be suppressed.
Furthermore, in the technique described in Patent Document 1, when the sound sources hold the same frequency component in a short time, there is a problem that the interference noise suppression performance is deteriorated and the sound is easily distorted.

Representative inventions disclosed in the present application are as follows.
It has a disturbing sound generation unit that determines whether it is a disturbing sound or a target sound for each band component of the input signal, extracts a band determined to be the disturbing sound, and generates a disturbing sound signal. And an adaptive processing unit characterized in that the spatial correlation inverse matrix is updated using the generated interference sound signal, and the interference sound suppression filter generated by using the generated spatial correlation inverse matrix is connected to the microphone array output signal. A sound processing device that performs sound source separation adaptively.

  According to the configuration of the present invention, it is possible to update the spatial correlation inverse matrix only in the band determined to be the interference sound based on the sparseness of the sound even in the section where the target sound exists. It is no longer necessary to previously extract a section in which there exists. Further, in the configuration of the present invention, an output signal is generated using an adaptive filter, and distortion of speech can be prevented.

  Embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a basic configuration diagram of a speech processing apparatus according to the present invention. The microphone array 1 acquires sound signals of a plurality of channels. The acquired sound signals of a plurality of channels are sent to the memory 2 and subjected to various signal processing by the CPU 3. The CPU 3 extracts and uses a sound transmission model from the target sound direction from the storage medium 4 as necessary.

  FIG. 2 is a block diagram of the embodiment of the present invention. The microphone array 1 acquires sound signals of a plurality of channels. The sound signals of a plurality of channels acquired by the microphone array 1 are converted into digital data by the A / D converter 2 and are taken into the memory 2. The sound signal converted into the captured digital data is represented as X (t). The sound signal is sent to the band dividing unit 6, subjected to short-time Fourier transform, and divided for each frequency band. The band-divided signal can be expressed as Xt (f). The band-divided signal is sent to the label generator 7. In the label generation unit 7, first, for each band.

で、音源方向jt(f)を決定する。ここで,Λは方向の推定範囲である。またaj(f)は位置ベクトルと呼ばれるベクトルで、マイクロホンアレイの配置より決定する音の伝達特性である（例えば、非特許文献１参照）。決定した各帯域毎の音源方向が予め定める目的音方向の範囲内にあれば、その帯域を目的音成分であると判定する。そうでなければ、その帯域を妨害音成分であると判定する。音声信号を40ms程度の短い時間で眺めた場合、その短い時間の中で、複数の音源が同じ周波数成分を持つことは確率的に低いという、スパース性という性質が成り立つ。そのため、ラベル生成部７で行う周波数振り分け処理を行うことで、妨害音が優位である信号と、目的音が優位である信号を精度良く分離することが可能となる。背景で述べたように、音声のスパース性に基づき、周波数成分を振り分ける妨害音抑圧技術では、音声が歪んでしまうという問題がある。複数の音源が同じ周波数成分を持つことは確率的に低いとは言っても、複数の音源が同じ周波数成分が持つことはあり得る。そして複数の音源が同じ周波数成分を持つ場合、その周波数成分が妨害音か目的音の一方にしか割り振られず、目的音に割り振られなかった場合、目的音のその周波数成分は失われてしまい、歪みの原因となってしまう。

Then, the sound source direction jt (f) is determined. Where Λ is the direction estimation range. Further, aj (f) is a vector called a position vector, and is a sound transfer characteristic determined by the arrangement of the microphone array (for example, see Non-Patent Document 1). If the determined sound source direction for each band is within a predetermined target sound direction range, the band is determined to be the target sound component. Otherwise, it is determined that the band is a disturbing sound component. When an audio signal is viewed in a short time of about 40 ms, a sparsity property is realized that a plurality of sound sources have a low probability that they have the same frequency component in that short time. Therefore, by performing the frequency distribution process performed by the label generation unit 7, it is possible to accurately separate a signal having a dominant disturbance sound and a signal having a dominant target sound. As described in the background, the interfering sound suppression technique that distributes frequency components based on the sparseness of the sound has a problem that the sound is distorted. Although it is stochastically low that a plurality of sound sources have the same frequency component, it is possible that a plurality of sound sources have the same frequency component. And if multiple sound sources have the same frequency component, that frequency component can only be assigned to one of the disturbing sound or the target sound, and if it is not assigned to the target sound, that frequency component of the target sound will be lost and distorted It becomes the cause of.

  In the present application, the result of the frequency distribution process is used only for the adaptive processing unit and the target direction estimation unit. And the output result is produced | generated using the linear filter which passes all the signals of the target sound direction. For this reason, even when sparsity is not established, it is possible to avoid a situation in which the sound is distorted because the effect of forming the blind spot of the filter after adaptation is slightly affected.

  The interfering sound generation unit 8 outputs the interfering sound using the band component determined by the label generating unit 7 as the interfering sound component. The interference sound output by the interference sound generator 8 is

と表される。適応処理部9では、妨害音生成部8が出力する妨害音を用いて空間相関逆行列を

It is expressed. The adaptive processing unit 9 uses the interference sound output from the interference sound generation unit 8 to calculate the spatial correlation inverse matrix.

で更新する。ここで、βは、忘却係数と呼ばれ、0〜1までの範囲で設定する。βが小さいほど、現在の妨害音の重みを大きくし、空間相関逆行列を更新する。空間相関逆行列の計算は、

Update with. Here, β is called a forgetting factor and is set in the range of 0 to 1. The smaller β is, the larger the weight of the current interference sound is, and the spatial correlation inverse matrix is updated. The calculation of the spatial correlation inverse matrix is

の逆行列としても良い。

It may be an inverse matrix of

Here, in order to update the spatial correlation inverse matrix using the signal output as the interference sound, the spatial correlation inverse matrix is created so as to include a lot of information on the interference sound direction.
The target sound direction estimation unit 10 uses the information on the sound source direction for each band calculated by the label generation unit 7 and uses the corrected delay sum array method (for example, Non-Patent Document 2) within the range of the target sound direction. The sound source direction is estimated, and the estimated sound source direction is set as the target sound direction.

  In the present application, a sound source outside the range of the target sound direction determined by the label generation unit 7 as the target sound component is determined as a disturbing sound, and a blind spot is formed and suppressed. Conversely, if the sound source exists within the range of the target sound direction, a dead angle is not formed. Therefore, if the sound source is within the range of the target sound direction, it can be extracted without being suppressed. When the target sound direction is limited to one direction, if the actual target sound direction and the target sound direction to be set are slightly deviated, the target sound component may be suppressed. In this respect, by providing a width in the target sound direction, there is an effect that the target sound component is not suppressed even if the actual target sound direction is different from the target sound direction to be set.

  The filter generation unit 11 uses the spatial correlation inverse matrix updated by the adaptive processing unit 9 and the target sound direction estimated by the target sound direction estimation unit 10,

の妨害音抑圧フィルタを生成する。asub(f)は目的音方向の伝達特性である。生成されたフィルタは、目的音方向にビームを当て、空間相関逆行列から計算される妨害音方向に死角を向け、抑圧することで、目的音だけを抽出するための線形フィルタである。この線形フィルタは、入力信号そのものに適用する。

The interference noise suppression filter is generated. asub (f) is the transfer characteristic in the target sound direction. The generated filter is a linear filter for extracting only the target sound by applying a beam in the direction of the target sound, directing the dead angle in the direction of the disturbing sound calculated from the spatial correlation inverse matrix, and suppressing it. This linear filter is applied to the input signal itself.

  The conventional adaptive method of the minimum dispersion beamformer method is adapted so as to suppress a sound source with a loud sound as much as possible. Therefore, the input signal contains a frequency component in which the target sound is dominant and the target sound When the direction of the sound is deviated from the assumed direction, the target sound is adapted to be suppressed.

  In the present application, even if the target sound is included in the input signal, the label generation unit 7 and the interference sound generation unit 8 extract the signal in which the interference sound is dominant (the target sound is hardly included) from the input signal. In addition, since a spatial correlation matrix in which the extracted disturbing sound is preferentially used as a signal is used, a filter that suppresses only the disturbing sound without suppressing the target sound direction can be generated.

  The filtering unit 12 uses the linear filter output from the filter generation unit 11, uses the band division signal obtained by the band division unit 6,

でフィルタリング後の信号を生成し出力する。波形生成部13ではフィルタリング部12が出力したフィルタリング後の信号を逆短時間フーリエ変換し、時間領域の信号を生成し、出力する。
上記実施例は装置構成を説明したが、本願はプログラムとしてコンピュータに読み込むことで実行されるようにしても良い。

To generate and output the filtered signal. The waveform generation unit 13 performs inverse short-time Fourier transform on the filtered signal output from the filtering unit 12 to generate and output a time-domain signal.
Although the above embodiment has described the device configuration, the present application may be executed by being read into a computer as a program.

1 is a basic configuration diagram of a speech processing apparatus of the present invention. The block diagram about embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Microphone array, 2 ... Memory, 3 ... CPU, 4 ... Storage medium, 5 ... A / D conversion part, 6 ... Band division part, 7 ... Label production | generation , 8 ... Interfering sound generation unit, 9 ... Adaptive processing unit, 10 ... Target sound direction estimation unit, 11 ... Filter generation unit, 12 ... Filtering unit, 13 ... Waveform generation Department.

Claims (2)

A microphone array holding at least two microphone elements;
A band division unit for outputting a band division signal obtained by dividing the signal output from the microphone array into a plurality of frequency bands for each channel;
For each band division signal output by the band division unit, a label generation unit that estimates a sound source direction and outputs a label indicating whether the band division signal is an interference sound or a target sound from the sound source direction;
An interference sound generating unit that outputs a band-division interference sound signal from a label output from the label generation unit, and an adaptive calculation that calculates a spatial correlation inverse matrix from the band-divided interference sound signal output from the interference sound generation unit A processing unit;
A filter generation unit that generates a jamming noise suppression filter using the calculated spatial correlation inverse matrix;
And a filtering unit that performs sound source separation by adapting the filter to a signal output from the microphone array. A target sound direction estimating unit that estimates the sound source direction using the sound source direction estimation information calculated by the label generation unit;
The speech processing apparatus according to claim 1, wherein the filter generation unit also uses information on the estimated sound source direction.

Images