JP2005257748A - Sound pickup method, sound pickup system, and sound pickup program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the S/N ratio of a voice signal superposed with noise. <P>SOLUTION: A sound pickup method includes a noise-suppressing processing for suppressing the noise included in an input signal and intensifying a target signal; a voice section degree calculating processing for calculating the voice section likeness in each section of the input signal; an original sound addition rate determining processing for determining the rate for adding an original sound to the target signal suppressing the noise, on the basis of the voice section likeness calculated by the voice section degree calculating processing; and an original sound addition processing for adding the original sound to the target signal suppressing the noise, in accordance with an original sound addition rate determined by the original sound addition rate determining processing. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention calculates the likelihood of a speech segment at each time Δt for a signal subjected to noise suppression processing, and automatically sets the original sound addition rate according to the likelihood of the speech segment, thereby maintaining noise while maintaining speech quality. The present invention relates to a sound collection method, a sound collection device, and a sound collection program for performing suppression.

A technique has been conventionally proposed in which distortion of a processed sound is reduced by adding an original sound to a signal subjected to noise suppression processing, thereby improving the quality of hearing (Non-Patent Document 1).
In the original sound addition method disclosed in Non-Patent Document 1, a process of keeping the original sound addition rate constant over the entire section (all sections in the time axis direction) is performed.
Junko Sasaki, Yoichi Haneda, “Examination of low distortion single-input noise reduction method considering masking effect,” sound theory, 2-3-11, pp. 525-526, (1997.3)

However, for example, in an environment where the noise is extremely large (an environment where the noise is larger than the target speech), if the original sound addition rate sufficient to improve speech quality is always added at a constant value, the noise will remain large. Therefore, there is a problem that it is difficult to improve the S / N ratio while maintaining the quality.
An object of the present invention is to propose a sound collection method, a sound collection device, and a sound collection program that can improve the S / N ratio over the entire interval in the time axis direction while maintaining the quality of the voice interval.

The present invention calculates the likelihood of a speech section for a signal after noise suppression processing, compared to the conventional method using a constant original sound addition rate in all sections, and keeps the speech quality as much as possible in the speech (likely) section. However, it is intended to improve the S / N ratio improvement amount in the entire section as compared with the conventional art.
Specifically, as a first embodiment of the present invention, noise suppression processing for suppressing noise included in an input signal and emphasizing a target signal, and a speech section for calculating the likelihood of a speech section for each section of the input signal are described. The original sound addition rate determination processing for determining the rate of adding the original sound to the target signal subjected to the noise suppression processing based on the sound interval characteristic calculated by the voice interval degree calculation processing, and the original sound addition rate determination processing A sound collection method including original sound addition processing for adding original sound to a target signal subjected to noise suppression processing according to the original sound addition rate is proposed.

In this first embodiment, the speech interval degree calculation process proposes a sound collection method in which a section in which the power of the input signal is greater than or equal to a predetermined value is determined as a speech section and a section below the predetermined value is determined as a noise section.
Furthermore, in the first embodiment, the voice interval degree calculation process determines a section in which the power of the input signal is equal to or higher than the first set value TS as a voice section, and the second set value in which the power of the input signal is smaller than the first set value. A section below TN is determined as a noise section, and when the power of the input signal is between the first set value TS and the second set value TN in the previous period, the sound collection is determined according to the power of the input signal. Suggest a method.

Furthermore, in the second embodiment of the present invention, the speech interval degree calculation processing includes at least a pair of speech input means installed at positions where the distances to the target sound source and the noise source are different from each other. The target signal including the noise signal captured by the voice input means is band-divided for each channel, the band components of each band-divided channel are compared between channels for each same band component, and input according to this level comparison result The voice signal component or the noise signal component is determined in association with the distance difference between the sound source and the target sound source and the noise source, and the power of the frequency band component determined as the voice signal component is integrated. We propose a sound collection method that determines the sound quality.
Furthermore, in the third embodiment of the present invention, a smoothing process for gradually changing the change is added to the change of the original sound addition rate, thereby preventing the original sound addition rate from changing suddenly and distorting the sound. We propose a sound collection method that never gives.

  The present invention calculates a physical quantity representing the likelihood of a voice section for a received signal collected by mixing a target signal and a noise signal, and automatically adjusts the original sound addition rate according to the likelihood of the voice section, It is possible to output a voice with less distortion in the voice section and a signal having a large S / N ratio improvement amount (less residual noise) in the noise section. As a result, it is possible to output a voice signal that is easier to hear and has less noise than the conventional method of uniformly determining the addition rate of the original sound in all sections.

Each embodiment which is the best mode for carrying out the present invention will be described in detail below.
FIG. 1 shows a first embodiment of the present invention. The first embodiment is a sound collecting device that operates using the sound collecting method proposed in claim 1 of the present invention, and this corresponds to the sound collecting device proposed in claim 6.
The sound collection device 100 proposed in the first embodiment of the present invention includes a voice input means 1 composed of a microphone, a noise suppression means 2, a voice segment degree calculation means 3, an original sound addition rate determination means 4, and an original sound addition. Consists of means 5.

The voice input means 1 receives the target signal S1 (t) and the noise signal S2 (t) from the target sound source S and the noise source N. Here, in order to simplify the description, the description will be made with one noise source N. However, in general, there may be a plurality of noise sources N. A signal x (t) (hereinafter referred to as an original sound signal) in which a noise signal is superimposed on a target signal is output from the audio input means 1.
The original sound signal x (t) output from the voice input unit 1 is input to the noise suppression unit 2. The noise suppression means 2 suppresses noise using a general method, for example, spectral subtraction. The signal subjected to noise suppression processing is shown as S1 ′ (t).
The noise-suppressed signal S1 ′ (t) is supplied to the speech interval degree calculating means 3, and the speech interval degree calculating means 3 calculates the “speech interval degree” representing the likelihood of the speech interval.

In the sound collection method proposed in claim 2 of the present invention as a method for calculating the likelihood of a speech section, the power (Pow) of the noise-suppressed signal S1 ′ (t) is calculated, and if the power exceeds a certain value TS, it seems to be a speech section. , And it is determined that it is a voice section. Conversely, if the power is equal to or less than TS, it is determined to be a noise section.
In the sound collecting method proposed in claim 3 of the present invention, for example, certain values TS and TN are set (TS> TN), and when Pow> TS is satisfied, it is determined that the likelihood of a speech section is 100%. In the case of TN <Pow <TS, it is determined that the closer the value of Pow is to TS, the higher the likelihood of a speech segment is. For example, let (Pow-TN) / (TS-TN) be an amount (sound interval degree) that represents the likelihood of a speech interval.

As shown in FIG. 2, when calculating the signal power Pow to calculate the speech interval, the original signal x (t) may be used instead of the noise-suppressed signal S1 ′ (t). The merit of the configuration of FIG. 2 is that, for example, if a speech interval is calculated after noise suppression processing is performed on a signal whose noise is not so large (for example, the S / N ratio is about 10 dB), the processing delay is increased. It becomes long, and is exhibited in an environment where it is inappropriate for communication, for example. In such an environment, using the received sound signal for calculating the speech interval degree allows noise suppression processing and speech interval degree calculation to be performed in parallel (although the speech interval degree calculation accuracy is somewhat degraded, instead) In addition, the processing delay can be shortened.
In the original sound addition rate determining means 4, the original sound addition rate is dynamically determined according to the voice interval degree calculated by the voice interval degree calculating means 3. Here are two specific methods. First, a description will be given of a case in which each section is classified into two types, that is, a voice section and a noise section by the voice section degree calculation means 3.
In this case, for the section determined to be a voice section (Pow> TS), the original sound addition rate α (t) is set high (for example, 0.3) for the purpose of placing importance on the voice quality. For the section determined to be a noise section (Pow <TN), the original sound addition rate α (t) is set low (for example, 0.05) for the purpose of giving priority to the S / N ratio improvement amount.

原音付加手段５においては、原音付加率決定手段４で決定された原音付加率α（ｔ）の値を原信号ｘ（ｔ）に乗算し、その信号と雑音抑圧された信号Ｓ１′（ｔ）を加算して出力信号ｙ（ｔ）として出力する。 Next, the voice interval degree calculation means 3 sets certain values TS and TN (TS> TN), and the original sound addition rate determination means when the likelihood of the voice interval is expressed as (Pow-TN) / (TS-TN). The operation example of 4 is described.
When Pow> TS is satisfied, the maximum value α _max (for example, α _max = 0.5) of the original sound addition rate is set as the addition rate α (t) on the assumption that the speech interval likelihood is 100%. When Pow <TN is satisfied, the noise section likelihood is 100%, and the minimum value α _min (for example, α _min = 0.1) of the original sound addition rate is set as the addition rate α (t). In the case of TN <Pow <TS, the original sound addition rate α (t) is determined by the method shown in FIG. That is, by obtaining a straight line passing through two points (TN, α _min ) and (TS, α _max ) on a graph with Pow on the horizontal axis and the original sound addition rate α (t) on the vertical axis, An original sound addition rate α (t) corresponding to Pow is calculated. The above operation is shown in Equation (1) using a programming language.

The original sound addition means 5 multiplies the original signal x (t) by the value of the original sound addition ratio α (t) determined by the original sound addition ratio determination means 4 and the signal S1 ′ (t) whose noise is suppressed. Are added and output as an output signal y (t).

Next, a second embodiment of the present invention will be described. The second embodiment of the present invention corresponds to the sound collection method proposed in claim 4 and the sound collection device proposed in claim 7.
Here, an example of the speech segment degree calculation means 3 will be described. In the first embodiment, the signal S1 ′ (t) or the original signal x (t) that has been noise-suppressed by the existing noise suppression means 2 is used as the signal used for calculating the speech interval degree. When the signal S1 ′ (t) is used, for example, when the noise suppression process of the existing noise suppression unit 2 does not work well, the calculation accuracy of the speech interval degree calculation unit 3 may be adversely affected. Further, when the original signal x (t) is used, it is difficult to calculate the speech interval degree when the S / N ratio of the received signal is extremely bad (for example, the S / N ratio is 0 dB or a negative value). Met.

In the second embodiment, FIG. 4 shows an example of a method for calculating a speech segment degree with the highest possible accuracy without being influenced by the noise suppression means 2 and the S / N ratio of the original signal x (t). . The input means 1A and 1B of the speech interval degree calculating means 3 are composed of, for example, two or more microphones. The band dividing means 6 performs frequency analysis on the signals from the input means 1A and 1B. For example, Fourier transform is used for frequency analysis. The inter-channel level difference calculating means 7 calculates the level difference ΔA (ω) between channels (the input means 1A side and 1B side are referred to as inter-channel) in each frequency component. ΔA (ω) is defined by equation (2).
ΔA (ω) = 20 log ₁₀ [| X ₁ (ω) | / | X ₂ (ω) |] (2)
The sound source signal determination means 8 determines whether each frequency component is a target signal component or a noise signal component based on the value of the inter-channel level difference ΔA (ω). For example, as shown in FIG. 4, the target sound source 3 is arranged closer to the input means 1A than the input means 1B, and conversely, the noise source N is arranged closer to the input means 1B than the input means 1A. If it is, the frequency component satisfying ΔA (ω) ≦ 0 is determined to be a noise signal component.

In the sound source signal selection means 9, based on the determination result of the sound source signal determination means 8, the target signal component is multiplied by a certain gain value g _S and the noise signal component is multiplied by a certain gain value g _N. A control expression in the sound source signal selecting means 8 is shown in Expression (3) using a program language.
if ΔA (ω) ≧ 0 then S ^ ₁ (ω) = g _S · X ₁ (ω)
elseif ΔA (ω) ≦ 0 then S ^ ₁ (ω) = g _N · X ₁ (ω) (3)
For example, g _S = 1.0 and g _N = 0.0 are given as examples of gain values.

The power integration means 10 integrates the power of the signal S ₁ (ω) weighted by the sound source signal selection means 9 over the entire frequency band. The integrated power Pow value is sent to the original sound addition rate determination means 4. The original sound addition rate determination means 4 determines the addition rate by the same control as described in the first embodiment using the integrated power Pow value. The voice interval degree calculating means 3 shown in FIG. 4 uses the direction information of the sound source for the original signal, determines whether each frequency component is a voice signal or a noise signal, and is determined as a voice spectrum. The band determined as a noise spectrum is suppressed by multiplication by a gain value of 0, so that the S / N ratio of the original signal x (t) is substantially reduced. Thus, the speech interval degree is calculated after improving the S / N ratio. For this reason, it is possible to improve the accuracy of calculating the speech interval degree for a signal having a poor S / N ratio, compared to the method using the original signal as it is. In addition, since the speech interval degree can be calculated independently of the existing noise suppression processing, the speech interval degree can be calculated without being affected even when the performance of the existing noise suppression processing is poor.

Next, a third embodiment of the present invention will be described. The third embodiment corresponds to the sound collection method proposed in claim 5 and the sound collection device proposed in claim 8.
A configuration example is shown in FIG. In the third embodiment, when the original sound addition rate α (t) is changed in the original sound addition rate determining means 4 described in the first embodiment, the original sound addition rate smoothing is performed to smooth the time change of the original sound addition rate α (t). The configuration is characterized by adding means 11.
In order to smooth the time change of the original sound addition rate α (t), the difference between the original sound addition rate α _pre (t) at the previous time and the original sound addition rate α (t) at the current time is calculated, and the difference The original sound addition rate α (t) at the current time is determined according to the polarity (+, −). An example of the determination method is shown in Formula (4) using a programming language.
if (α _pre (t) <α (t)) then α _smooth (t) = α _pre (t) + atk (α (t) −α _pre (t))
else then α (t) _smooth = α _pre (t) + rls (α (t) −α _pre (t))
(4)
Here, atk and rls are values satisfying 0 <atk <1 and 0 <rls <1. With the above operation, the original sound addition rate α (t) does not change significantly every time, and the time change becomes smooth because it only slightly increases or decreases from the previous time, and the sound quality distortion after processing is small. Become. The direction of the slight increase or slight decrease is determined by the difference between the original sound addition rate α (t) at the current time and the addition rate at the previous time, and the original sound addition rate α (t) at the current time is compared with the previous time. If it is increasing, it will increase slightly, and if it is decreasing, it will decrease slightly. The invention of the third embodiment can be used together with the second embodiment.

In this way, by using the above means, it is possible to calculate a physical quantity representing the likelihood of a speech section from a signal mixed with noise and automatically adjust the original sound addition rate according to the likelihood of the speech section based on the value. Compared with the conventional method of uniformly determining the original sound addition rate for all sections, distortion reduction is realized in the voice section, and the S / N ratio improvement amount is secured in the noise section, making it easy to hear in all sections, and remaining. A signal with less noise can be extracted and collected.
The sound collecting method of the present invention described above is realized by causing a computer to execute the sound collecting program according to the present invention. The sound collection program according to the present invention is written in a computer-readable program language, recorded on a recording medium such as a magnetic disk or a CD-ROM, and installed on the computer from the recording medium or installed through a communication line. The The installed sound collecting program is decoded and executed by a central processing unit CPU provided in the computer.

  The sound collection method and sound collection device according to the present invention can be applied to, for example, a sound collection device of a voice recognition device, and can improve the recognition rate of the voice recognition device by improving the quality of a voice signal to be collected. it can.

The block diagram for demonstrating 1st Embodiment of this invention. The block diagram for demonstrating the modification of 1st Embodiment shown in FIG. The graph for demonstrating the operation | movement of 1st Embodiment shown in FIG. The block diagram for demonstrating 2nd Embodiment of this invention. The block diagram for demonstrating 3rd Embodiment of this invention.

Explanation of symbols

1, 1A, 1B Voice input means 7 Channel level difference calculating means
2 Noise suppression means 8 Sound source signal judgment means
3 Voice interval degree calculation means 9 Sound source signal selection means
4 Original sound addition rate determining means 10 Power integrating means
5 Original sound addition means 11 Original sound addition rate smoothing means
6 Band dividing means 100 Sound collecting device

Claims (9)

Noise suppression processing that suppresses noise contained in the input signal and emphasizes the target signal;
A voice segment degree calculation process for calculating the likelihood of a voice segment in each segment of the input signal;
An original sound addition rate determination process for determining a rate of adding the original sound to the target signal subjected to the noise suppression process based on the likelihood of the voice interval calculated in the voice interval degree calculation process;
Original sound addition processing for adding the original sound to the target signal subjected to the noise suppression processing according to the original sound addition rate determined in the original sound addition rate determination processing;
A sound collection method comprising:   2. The sound collection method according to claim 1, wherein the voice interval degree calculation processing determines a section where the power of the input signal is a predetermined value or more as a voice section, and determines a section below the predetermined value as a noise section. Sound method.   The sound collection method according to claim 1, wherein the voice segment degree calculation processing determines a section in which the power of the input signal is equal to or greater than a first set value TS as a voice section, and the power of the input signal is greater than the first set value. A section having a small second set value TN or less is determined as a noise section, and when the power of the input signal is between the first set value TS and the second set value TN in the previous period, the sound is determined according to the power of the input signal. A sound collection method characterized by determining the likelihood.   The sound collection method according to claim 1, wherein the speech segment degree calculation processing includes at least a pair of input means installed at positions where the distances to the target sound source and the noise source are different from each other. The target signal including the noise signal captured by the input means is band-divided for each channel, the band components of each band-divided channel are compared between channels for each same band component, and the input according to the level comparison result The voice signal component or the noise signal component is determined in association with the distance difference between the means and the target sound source and the noise source, the power of the frequency band component determined as the voice signal component is integrated, and the integrated value of this power A sound collection method, wherein the sound quality is determined by the method.   5. The sound collection method according to claim 1, wherein a difference between an original sound addition rate at the current time determined by the original sound addition rate determination process and an original sound addition rate determined at a previous time is calculated, A sound collection method characterized by gradually changing an original sound addition rate toward a target original sound addition rate at a current time according to polarity. Noise suppression means having a function of suppressing noise with respect to the output signal of the voice input means and emphasizing a target signal;
A voice interval degree calculating means for calculating the likelihood of a voice interval at a certain time of the output signal of the voice signal input means or the noise suppression means;
Original sound addition rate determining means for determining the original sound addition rate α (t) (0 <α (t) <1) based on the likelihood of the voice interval calculated by the voice interval degree calculating unit;
Original sound adding means for adding the output signal of the voice input means to the output signal of the noise suppressing means after multiplying the original sound addition ratio α (t) determined by the original sound addition ratio determining means (0 <α (t) <1) When,
A sound collecting device comprising: The sound collecting device according to claim 6,
A plurality of microphones provided apart from each other as the sound input means, and a band dividing means for dividing a plurality of channel signals output from the plurality of microphones as the sound interval degree calculating means, and the band dividing means For each same band of each channel signal divided in step 1, a level difference calculation unit between channels for each band to calculate a level difference for each same band component between channels;
Sound source signal determination means for determining which one of the above-mentioned band-divided channel signals of each band is a signal input from which sound source based on the level difference between channels for each band component;
Based on the determination of the sound source signal determination means, sound source signal selection means for selecting a signal input from the target sound source from each of the band-divided channel signals,
Power integrating means for integrating the power of the output signal of the sound source signal selecting means;
Original sound addition rate determining means for determining the original sound addition rate based on the power value integrated by the power integration means;
Original sound addition means for adding the original sound signal output by the voice input means to the target signal output by the noise suppression means according to the original sound addition ratio determined by the original sound addition ratio determination means;
A sound collecting device comprising:   8. The sound collecting device according to claim 6, wherein a difference between the original sound addition rate at the current time determined by the original sound addition rate determination means and the original sound addition rate determined at the previous time is obtained, and a value of the difference is obtained. And an original sound addition rate smoothing means for gradually changing the original sound addition rate toward the target original sound addition rate at the current time according to the polarity.   6. A sound collection program that is written in a computer-readable program language and causes the computer to execute the sound collection method according to claim 1.

Images