JPH0251200B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an audio information filter, and its purpose is to accurately remove noise components from noisy audio information and reproduce clear audio information that is easy to hear. The purpose is to provide an information filter.

In general, it is difficult to avoid noise being superimposed on voice information that is emitted at a point far away from a microphone or in an environment with large ambient noise. However, even in such a noisy environment, there are strong demands to record clear voice information and to increase the recognition rate of voice recognition devices as much as possible. For this purpose, it is necessary to remove as much noise as possible from the voice information on which noise is superimposed.

The present invention has been made in view of such demands, and will be described below with reference to drawings of embodiments.

The present invention provides band dividing means for dividing an input signal into a plurality of frequency bands, the same number of band signal delay means, the same number of delayed signal attenuation means, the same number of band signal average energy measuring means, and one attenuated signal. It is characterized by an audio information filter having an addition output means and a set of control means capable of controlling the amount of attenuation of each of the above-mentioned attenuation means, eliminating the need for means for measuring the degree of periodicity of the signal. Therefore, it has the advantage that the device can be realized with a simpler configuration and that even noise having strong periodicity can be removed.

In the present invention, while voice information has a biased energy distribution on the frequency axis due to its vocal tract characteristics, a so-called formant structure, most of the noise to be removed has such an energy distribution. This study focuses on the fact that the frequency band often exists dispersed over a wide frequency band without having a clear bias. In other words, in voice information with superimposed noise, the ratio of voice energy to noise energy (SN ratio) differs depending on the band on the frequency axis, so this signal is divided into n predetermined bands (channels) on the frequency axis. ), and pass only the m channels in which the audio energy has been expanded, and attenuate the signals of the other channels to improve the S/N ratio.
It becomes possible to perform noise removal.

Next, a more detailed explanation will be given using the drawings. FIG. 1 is a block diagram showing one embodiment of the audio information filter of the present invention. The input signal x _(t) to this device is applied from the input terminal 100, and is first applied to the bandpass filter 20.
It is divided into a plurality of frequency band signals Ui _(t) (i=1, 2, . . . n) by i (i=1, 2, . . . n).
Each band signal Ui _(t) has a delay time τi seconds (i
A delay circuit 30i (i=1, 2,...n)
1, 2, . . . n), and at the same time, it is also guided to an average energy measuring circuit 60i (i=1, 2, . . . n). The average energy measuring circuit 60i calculates the short-term average energy W _{i (} k) (i=
1, 2,...n) (k is a discrete parameter indicating the section number for each fixed period of time, and these sections are hereinafter referred to as frames), and these are sent to the control circuit 70.
Output to 0. The control circuit 700 controls the amount of passage through the attenuator 40i (i=1, 2,...n) using a method described later.
Attenuator 40i is controlled to determine g _i ,
At this time, the response time of each part is set so that each band delayed signal V _i (t) (i=1, 2,...n) outputted from the delay circuit 30i is applied to the attenuator 40i. shall be. That is, by selecting the sum of the response time of the average energy measurement circuit 60i and the response time of the control circuit 700 to be equal to the delay time τ _i of the delay circuit 30i, it is possible to determine the time when the energy measurement becomes a target. Band signal U _i (t)
The damping can be controlled. Furthermore, since the average energy measurement and attenuation control are performed for each frame as described above, the band-specific energy information of the input signal is reflected in the noise removal operation with the frame length as the resolution. in this way,
The output _{Z i (} t) (i=
1, 2,...n) are added by the addition output circuit 500, and output to the output terminal 800 as the final output signal y(t).
is output to.

Next, the method for determining the amount of passage g _i will be described. FIG. 2 is a flowchart of the decision method. In the figure, the overall average energy W _A (j) in each frame of the input signal given in the input signal reading step 1 is compared with a predetermined threshold Th in the condition determination step 2. In this case, the entire band average energy W _A (j) is the threshold
In order to consider the interval exceeding Th and the certain interval before and after it as an interval in which audio exists, if Equation (6) is satisfied in step 2 of condition judgment, the kth frame is determined as audio in step 6. If this is not true, in step 3, the k-th frame is determined to be a section in which no audio exists.

W _A (j)Th ......(6) (However, j=k, k±1, k±2) If it is determined that the kth frame is a frame in which no audio exists, then in step 4, Equation (7) is The noise energy for each band is estimated, and further, in step 5, the amount of passage g _i expressed by equation (8) is determined.

N^ _pi = W _i ………(7) g _i = O ………(8) (However, to avoid complexity, the frame number display (k) in W _i , g _i , and S^i will be omitted from now on. ) On the other hand, if the k-th frame is determined to be a frame in which audio exists, then in step 7, cases are further divided based on the conditional judgment formula W _i N^ _pi (9) shown in equation 9. , if Equation (9) does not hold, the amount of passage g _i shown by Equation (8) is determined in Step 5, and if Equation (9) holds true, Equation (10) is determined in Step 8. The voice energy is estimated based on the

S^ _i = W _i −N^ _pi ………(10) Next, in step 9, the magnitudes of S^ _i are compared and calculations are made in order of the magnitude of S^ _i from the channel that gives the largest S^ _i . Select m channels (O<mn) and perform the step 10 for these channels.
The amount of passage g _i is determined based on equation (11). In addition, the above m
For channels that are not included in the channel,
In step 5, the amount of passage g _i is determined based on equation (8).

g _i = S^ _i /W _i ......(11) Immediately after the start of operation of this device, N^ _M is always determined by formula (7), regardless of whether formula (6) is successful or not. shall be stipulated.

As described above, according to this embodiment, only the signals in which a large weight is applied to channels with high SN ratios for each band are collected in the addition output circuit 500, and as a result, the SN ratio in all bands is improved. can be expected.

The embodiment shown in FIG. 1 can be realized using, for example, the following specific characteristics and circuit. The frequency band filter 20i is set for each 1/3 octave from about 50 Hz to about 7 to 8 KHz so as to cover almost the entire audio band. Gain frequency characteristics are 12dB/
It may be a gentle one of oct. or a relatively sharp one of about 40 dB/oct. The delay circuit 30i has
A delay time of 10 to 20 msec is obtained using a BBD (basket brigade device), and an attenuation circuit 40
i has a maximum of 40~ with a resolution of about 3 dB.
A circuit that can provide attenuation of up to 50 dB may be used.
Furthermore, the average energy measurement circuit 60i uses a full-wave rectification/smoothing/sample-hold circuit, and inputs a total of n signals obtained to a multiplexer to
After combining the signals into channels, the signals may be A/D converted and input to the control circuit 700.

FIG. 3 shows another embodiment of the audio information filter of the present invention. In the figure, the band filter 90i is the band filter 20 at the input section in the corresponding band.
It has the same characteristics as i. By the way, the first
The amount passed through the attenuator 40i in the figure is generally updated discretely at time intervals of the frame length. Therefore, the output signal of the attenuator 40i may undergo a sudden amplitude change, and as a result, the attenuator output signal
Unnatural intermittent sounds may occur in z _i(t) and ultimately in the final output signal y _(t) . Therefore, as shown in FIG. 3, a bandpass filter 90i is connected immediately after the attenuator 40i, and the transient response of this filter 90i is utilized to alleviate the above-mentioned rapid amplitude change. The steeper the filter characteristics, the greater the effect of alleviating sudden amplitude changes in the input waveform caused by the filter's transient response. Even if a sudden change in the amount of passage occurs in a channel, the influence on the final waveform is weakened by adding the signals of all these channels. For the above reasons, the insertion of the bandpass filter immediately after the attenuator in this embodiment is more effective when the amplitude frequency characteristic of the bandpass filter used is steeper, and is only necessary in that case. be.

In this way, it is possible to realize an audio information filter that can accurately remove noise components from audio information containing noise and reproduce clear audio information that is easy to hear.

Next, the operation of the audio information filter according to the present invention will be briefly described with regard to two typical types of noise.

FIG. 4 shows the spectrum of an input signal in which voice and stationary random noise with a flat spectrum are superimposed. By selecting the threshold Th to be larger than the energy of only noise and smaller than the energy of the input signal in which voiced sound and noise are superimposed, equations (6) and (7) described in the example can be obtained. , the voice energy can be easily estimated according to equation (10). In Figure 4
E ₁ , E ₂ , and E ₃ are bands in which the estimated speech energy obtained in this way is non-negative, but among these, there are also bands of a total of m channels in which the estimated speech energy S^ _i is larger than the others. only are extracted. In this way, stationary random noise whose spectrum is relatively flat can be removed.

Moreover, FIG. 5 shows the spectrum of an input signal in which voice and resonant noise with a concentrated spectrum are superimposed. Since the Q of the noise spectrum is large, even if the peak energy of the noise on the frequency axis is greater than the peak energy of the voice, the energy of the voice will be greater when comparing the energy in all bands. Therefore, by appropriately selecting the threshold Th, it is possible to estimate N^ _pi , and by using equation (9) described in the embodiment, N^ _pi is compared with the input signal energy W _i for each frame/channel. Judgment becomes possible. As a result, in the band indicated by F ₁ in FIG. 5, equation (9) does not hold, so the amount of passage through the attenuator is set to zero, making noise removal possible.

As described above, according to the present invention, noise can be accurately removed from noisy audio information with a simple circuit configuration, and clear audio information that is easy to hear can be reproduced.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the audio information filter of the present invention, and FIG. 2 is a flowchart for explaining a method for determining the amount of information passed through the filter.
FIG. 3 is a block diagram showing another embodiment of the audio information filter of the present invention, and FIGS. 4 and 5 show a stationary random noise with a flat spectrum and a concentrated spectrum, respectively, to explain the operation of the same filter. FIG. 3 is a frequency spectrum diagram of audio including resonance noise. 100 is an input terminal, 20i (i=1,...,n)
is a bandpass filter, 30i (i=1,...,n) is a delay circuit, 40i (i=1,...,n) is an attenuator, 5
00 is an addition output circuit, 60i (i=1,...,n)
is an average energy measuring circuit, 700 is a control circuit, 8
00 is an output terminal, and 90i (i=1,...,n) is a band filter.

Claims (1)

[Scope of Claims] 1. Band dividing means for dividing an input signal into a plurality of frequency bands; Delay means for delaying each band signal divided by the band dividing means by an arbitrary time; an attenuating means for passing an arbitrary amount of each of the band signals; and an addition output means for adding and final outputting the respective band signals attenuated by the attenuating means;
The apparatus comprises an average energy measuring means for receiving the output of the band dividing means and measuring the average energy of each band signal, and a control means capable of controlling the amount of passage through the attenuation means, An audio information filter characterized in that the amount of passage through the attenuation means is controlled based on the measured value of the average energy measuring means. 2. In the audio information filter according to claim 1, the average energy is measured for each short period and for each band from an input signal containing a mixture of audio and noise, and based on the obtained average energy value, Audio information characterized in that an estimated value of audio energy and an estimated value of noise energy are calculated, and an amount of passage through an attenuation means is determined based on the calculated estimated value of audio energy and estimated value of noise energy. filter. 3. In the audio information filter according to claim 2, the short-term average energy of the input signal over the entire frequency band is determined, and the section where this value exceeds a certain value and the certain sections before and after that are determined as sections where speech exists. Segmentation is performed by regarding other sections as sections where there is no voice, and the average energy for each band in the section where there is no voice is the estimated noise energy N^ _pi (where, i indicates the band number), and in the section where voice exists, the estimated value of the voice energy S^ _i is determined based on equation (1), and then
An audio information filter characterized in that the amount of passage q _i of the attenuation means is determined based on equations (2) and (3). S^ _i = W _i −N^ _pi ………(1) However, when S^ _i < o, set S _i = o. Here, i: frequency band number. (i=1, 2..., n) W _i : Average energy measured in an arbitrary interval. In the section considered to be a voice section, g _i = S^ _i / W _i ......(2) However, in each arbitrary section, in bands other than the total m band that gives the largest S^ _i , g _i = o. o<mn In the section where there is no voice, g _i =o (3) for all bands. 4. In the audio information filter according to claim 3, when calculating the estimated noise energy value N^ _pi , on the series of short-term average energy W _A in the entire frequency band measured in each fixed interval. Then, when the subsequent W _A of each adjacent W _A gives a smaller value than the immediately preceding W _A for the first time, determine the estimated value N^ _pi of the noise energy for each band using equation (4). A voice information filter featuring: N^ _pi =W _i(h) +W _i(h+1) /2 ………(4) However, h+1 indicates the number of the interval when W _A gives a smaller value than the previous W _A for the first time, W _i (h+
1) indicates the average energy of band i in the section whose section number is h+1. 5. In the audio information filter according to claim 1, before inputting the attenuated band signal to the addition output means, a new filter having characteristics equivalent to the band dividing means corresponding to each band of each band signal is used. An audio information filter characterized in that the band is divided again by a band dividing means, and a new band signal obtained by this operation is inputted to the addition output means.