JPS6044680B2 - Audio information filter - Google Patents

Description

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

In general, it is difficult to avoid superimposition of noise 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, we want to record clear audio information. There is a strong demand for increasing the recognition rate of speech recognition devices as much as possible. To this end, it is necessary to remove as much noise as possible from the noise-superimposed audio information, but there are only a few technologies that can fully meet this requirement. In particular, when removing noise from speech information on which noise is superimposed, (1) If the noise has periodicity, (I
It is difficult to effectively remove just the noise in i) when the noise is sudden, {iii) when the energy of the noise is greater than the voice energy in all bands where the main energy of voice exists, etc. It was hot. The present invention
The input signal is analyzed into each band energy by the I-channel bandpass filter, and these are divided at a certain time. The statistical appearance probability of the board-like energy pattern obtained when dividing the spectrum into time series for each interval (this is called a frame) is different when the input signal is only audio and when noise is superimposed on the audio. We are focusing on the fact that the situation is different depending on the situation. Book.
A conceptual diagram of such a spectrum time series obtained when implementing the invention is shown. 1 in the figure is the channel number parameter,
k is a frame number parameter, and these exist up to 11 and k, respectively. Now, if we focus on one substrate (called a cell) in the kth frame of the 1st channel, this cell is the one that is used for the kth frame of the band signal analyzed by the bandpass filter of the first channel. It can be seen that the short-time average energy is the portion written after being quantized by a predetermined energy threshold. Therefore, FIG. 1 is an energy matrix with K rows and 1 column, and the entire input signal is characterized by this pattern. If we take a relatively small region (called a unit mesh) consisting of consecutive NN channels and consecutive MM frames and consider that this region represents the local characteristics of the input signal, we can create a unit mesh. Pattern of quantized energy that appears (unit pattern)
It is possible to locally understand the characteristics of the input signal. Specifically, while fixing the n1 cell to band 1 and moving the unit mesh one frame at a time in the frame direction, find the cumulative frequency of the unit pattern that occurs in each case, and divide it by the total frequency of occurrence to calculate the probability of occurrence. Find the density function Pi(γ). Here, γ is an index indicating the type of unit pattern, and when the unit mesh consists of NN channels and MM frames, γ=1, 2,...L
nNx′″M (where L indicates the total number of quantization levels).The set of appearance probability density functions obtained is {Pi(γ)Ii=1,2,...I −n, +
1}. When implementing this, during actual noise removal processing, unit patterns related to each channel and frame are sequentially measured for the noise-superimposed audio signal, and based on the index of the obtained unit pattern, Referring to the standard P1(γ) written in the memory in advance, if this Pi(γ) is smaller than a predetermined probability density threshold θi, it is considered an illegal pattern and the band signal of the corresponding unit mesh part is attenuated. The pattern is attenuated, and if it is larger than the threshold value θi, it is input to the addition output circuit as a legal pattern without being attenuated. FIG. 2 is a block diagram of an apparatus for carrying out the above processing.

The input signal x(t) to this device is applied from the input terminal,
First, a plurality of frequency band signals Ui(t) (1=1, 2, . . . n) are filtered by a band filter 201 (1=1, 2, . . .
n). Each frequency band signal Ui(t) is guided to a delay circuit 301 (1=1, 2...n) with a delay time γi (1=1, 2,...n) seconds, and at the same time the average energy Measurement circuit 601 (1=1, 2...
・It is also guided by n). The average energy measuring circuit 60i measures the short-time average energy Wi(k) (1=1, 2...n) of each frequency band signal Ul(t) for each frame,
Each predetermined threshold for quantization λl(t)=1,
2, . At this time, the value of each threshold value λl may be normalized for each corresponding unit mesh. The control circuit 700 receives this pattern, reads Pi(γ) from the storage circuit 900 according to the corresponding pattern index γ, and controls the attenuator 401 (1=1, 2) according to the predetermined probability density threshold θi.・・・・・・
・The attenuation amount of n) is controlled, but at this time the delay circuit 301
The response time of each part must be set so that each band delay signal UI(t) outputted from the attenuator 401 is applied to the attenuator 401 in the same way. More specifically, when Pi(γ) of a unit pattern generated in an arbitrary channel and arbitrary frame is less than θi, the question remains as to which direction frame of the band signal Ui(t) should be attenuated. . To deal with this, it is practical to attenuate only the signal in the portion corresponding to one cell in the center of the unit mesh. By doing this, if an abnormality occurs in the first channel, frame k, due to noise superimposition, as shown in Fig. The signal Ui(t) in the portion corresponding to the range of cells may be attenuated. Immediately, thi,
5 In addition to the I, k cells themselves being attenuated due to the abnormality of the 1, k cells, the 1, k cells are also part of unit meshes such as B, C, D, and E in the figure; There is a good possibility that the unit pattern is also an illegal pattern. Therefore, B, . The cells indicated by c, d, and e may also be attenuated, and as a result, there is a possibility that the signal U1(t) in the portion corresponding to the unit mesh indicated by A in the figure is all attenuated. The output signal Zi(t) from the attenuator 401 controlled in this way (1=1, 2,...
, n) are added by the addition output circuit 500, and the output terminal 800
The output signal y(t) is output to. Note that in the above embodiment, the control circuit 700 calculates the standard appearance probability density frequency Pl(γ
), the threshold value θi of the appearance probability density of the unit pattern read into the control circuit 700 is compared in size, and the attenuator 401 is controlled based on the result. All the determined legal patterns may be stored, and the control circuit 700 may additionally perform a process of approximating the actual pattern with the legal pattern to remove noise.

In addition, the control circuit 700 compares the change characteristics of the pattern appearance probability when noise is superimposed, which is stored in the storage circuit 900, with the actual pattern through real-time processing, and determines the presence or absence of noise based on the difference between the two. You may perform a process to remove it. As described above, according to the present invention, it is possible to selectively pick up a signal portion having the same energy pattern (legal pattern) that often appears when measuring only audio from a signal superimposed with noise. On the other hand, signal parts whose energy patterns become abnormal (illegal patterns) due to the superposition of noise can be discarded, so noise components can be appropriately removed and clear, easy-to-listen audio information can be reproduced. There are some things that can realize a voice information filter that can do this.

[Brief explanation of drawings]

FIG. 1 is a conceptual diagram of a spectrum time series measured in carrying out the present invention, FIG. 2 is a block diagram showing an embodiment of the audio information filter of the present invention, and FIG. 3 is an explanatory diagram of its operation. 100...Input terminal, 201 (1=1,...,
n)...Band filter, 301 (1=1,...,
n)...Delay circuit, 401 (1=1,...,n)
... Attenuator, 500 ... Addition output circuit, 601
(1=1,...,n)... Average energy measurement circuit, 700... Control circuit, 800... Output terminal, 900... Unit pattern appearance probability density storage. circuit.

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; attenuating means for attenuating each of the band signals by an arbitrary amount; addition output means for adding and final outputting the respective band signals attenuated by the attenuating means; comprising an average energy measuring means for measuring the average energy of the signal at regular time intervals, and a control means capable of controlling the control amount of the attenuation means at regular time intervals; The appearance probability density of each pattern formed by a multidimensional time series over one or more time intervals of the quantized average energy obtained at fixed time intervals from one or more band signals is calculated by eliminating noise. For the input signal with superimposed noise, the appearance probability density obtained in the same pattern as the pattern of each multidimensional time series obtained by the same method as above is calculated in advance for long-duration speech that does not contain sound. When the probability density is smaller than a predetermined threshold value for each band, it is regarded as an illegal pattern and the band signal in the band and time interval corresponding to the multidimensional time series is attenuated by the attenuation means. and when it is larger than a threshold value, it is regarded as a legal pattern and the band signal is inputted to the addition output means without attenuating it. 2. In the audio information filter according to claim 1, one or more times of the average energy obtained and quantized at regular time intervals from one or more band signals obtained in advance by the band dividing means. Analyze each pattern formed by a multidimensional time series over an interval, and from among all the patterns that have appeared, select all the patterns that have appeared with a sufficiently good degree of approximation in terms of probability of occurrence. A series of patterns that can be approximated (legal patterns) are extracted, and for input signals with superimposed noise, all patterns generated from each multidimensional time series obtained using the same method as above are extracted from the legal patterns. 1. An audio information filter characterized in that a signal portion corresponding to each interval consisting of a frequency band and a time interval on a legal pattern after the approximation is controlled by an attenuation means. 3. In the audio information filter described in claim 2, it is understood in advance how the appearance frequency distribution of legal patterns changes when a signal with superimposed noise is input, and the appearance frequency distribution is sequentially measures the temporal transition of A voice information filter characterized in that the noise superimposed section and the non-noise superimposed section are subjected to different processing operations according to the determination.