JPH04340599A - Noise canceller device - Google Patents

Abstract

PURPOSE:To eliminate the disordering of the frequency spectrum of voice by performing spectrum subtraction in the presence of an unsteady noise. CONSTITUTION:The frequency spectrum of a voice input signal is found, the frequency spectrum of a noise is found from an input signal before voice input, and the frequency spectrum of the noise is subtracted from the frequency spectrum of an input signal at the time of the voice input to remove the noise. A noise steadiness calculation part 4 fixes the steadiness of the ambient noise according to the frequency spectrum of the noise. A noise correction part 7 varies the rate of the removal of the noise according to the steadiness.

Description

[Detailed description of the invention]

0001

TECHNICAL FIELD The present invention relates to a noise removal device, and more particularly to a noise removal device in speech recognition technology. For example, it is applied to speech recognition technology under noise.

0002

2. Description of the Related Art A method commonly used as a countermeasure against noise in speech recognition devices is the spectral subtraction method (S.F. Boll, "Suppression").
of Acoustic Noise in Spe
ech Using Spectral Subtra
ction”, IEEE Trans. ASSP Vo
l. 27-2, April. 1979) and adaptive noise canceling (B. Widraw, “Adapti
ve noise canceling: Princi
ples and applications”, Pr.
oc. IEEE, Vol. 63, Dec. There are methods using noise removal techniques such as 1975). However, adaptive noise canceling requires a plurality of input units and feature extraction units, so it has the disadvantage of increasing cost. On the other hand, the spectral subtraction method subtracts the noise frequency spectrum estimated from the input signal before voice input from the voice frequency spectrum.
In the case of non-stationary noise, since the noise spectrum during noise estimation and the noise spectrum during voice input are significantly different, spectral subtraction can actually
This method has the disadvantage that it destroys the frequency spectrum of the voice, reducing the recognition rate.

Note that FIGS. 3(a) to 3(e) show examples of cases in which noise removal cannot be performed appropriately. ) are significantly different, then spectrum subtraction results in a signal like (d), which means that the spectrum (c) before spectrum subtraction is more similar to the input speech spectrum (e) than (d). result.

0004

[Objective] The present invention has been made in view of the above-mentioned circumstances, and an object of the present invention is to provide a noise removal device that eliminates distortion of the frequency spectrum of speech by performing spectral subtraction under non-stationary noise. It was done for a purpose.

[0005]

[Structure] In order to achieve the above objects, the present invention provides (1)
detection means for determining the frequency spectrum of the audio input signal;
A noise removal device having a detection means for determining the frequency spectrum of noise from an input signal before voice input, and a noise removal means for removing noise by subtracting the frequency spectrum of the noise from the frequency spectrum of the input signal at the time of voice input. comprising a calculation means for determining the stationarity of surrounding noise from the frequency spectrum of the noise, and a noise correction means for changing the noise removal rate according to the stationarity determined by the calculation means, and further (2) The calculation means for determining the stationarity of the surrounding noise calculates the stationarity from the time change of the frequency spectrum of the input signal when the noise is input, and the smaller the time change, the higher the stationarity. That is. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining an embodiment of a noise removal device according to the present invention. In the figure, 1 is an acoustic input section, 2 is a frequency spectrum detection section, 3 is an interval detection section, and 4 is an acoustic input section. A noise stationarity calculation section, 5 a noise removal coefficient calculation section, 6 a noise spectrum estimation section, 7 a noise correction section, and 8 a speech recognition section. The acoustic input section 1 converts sound into an electrical signal x(t) using an acoustic/electrical signal converter such as a microphone. The frequency spectrum detection unit 2 detects a time-frequency spectrum X(t,f) of about 10 msec of the electrical signal x(t) obtained by the acoustic input unit 1 using a group of band-pass filters, FFT, etc. .

[0007] The section detecting section 3 detects whether or not voice input is in progress, and uses, for example, a method in which a section in which the x(t) is greater than or equal to a predetermined threshold is determined to be voice input. . Further, as another method, a method is used in which an audio input switch is provided and the period in which the switch is turned on is set as audio input, but it is also possible to implement it using other methods. The steady/unsteady determination section 4 includes the section detection section 3
The stationarity of surrounding noise is determined using the input signal detected in the interval not during voice input. For example, using the short-time frequency spectrum X (t, f) detected by the frequency spectrum detection unit 2 Then, D(t,f)=X(t,f)−X(
t-Δ,f) (Δ: 10msec) After finding the change for each band,

[0008]

[Math 1]

The sum E(t
), and calculate the average value Z of this E(t) during time T

00
10]

[Math 2]

##EQU1## is determined as a value indicating the stationarity of the noise. The noise removal coefficient calculation unit 5 calculates the noise removal coefficient α from the value Z indicating the noise stationarity determined by the noise stationarity calculation unit 4. The noise removal coefficient α is set so that the larger Z is, in other words, the more stationary the noise is, the larger it becomes (0≦α≦1
). The noise spectrum estimator 6 calculates the short-time frequency spectrum
(t, f) to estimate the frequency spectrum of surrounding noise,

0012

[Math 3]

A method of taking the average over time T is used. The noise correction unit 7 removes noise components from the input signal using the noise removal coefficient α calculated by the noise removal coefficient calculation unit 5, and Y(t,f)=X(t,f)−α・Like N(f),
Frequency spectrum of input audio X(t,f) for each band
The input signal is corrected by subtracting the value obtained by multiplying the noise spectrum N(f) estimated by the noise spectrum estimator 6 by the noise removal coefficient α from the noise spectrum estimator 6. The speech recognition section 8 includes a noise correction section 7
Speech recognition is performed using the output signal Y(t,f) of ``Development of a word speech recognition system using binary TSP'' (Yasuda et al., IEEJ Transactions Vol. C108, 1932).
October 3rd issue p. Although the speech recognition system described in 858-865) is used, other speech recognition systems can also be used.

FIG. 2 is a flowchart for explaining the operation of the noise removal apparatus according to the present invention. Below, each step will be explained in order. Step 1: First, input the acoustic signal X(t). Step 2: Detect the frequency spectrum X(t,f). Step 3: Determine whether it is in the voice section. Step 4: If it is during the voice section in step 3, noise correction is performed. The calculation formula is as follows. Y (t, f) = X (t, f) - α・N (f) step 5
; Performs voice recognition processing. Step 6: If it is not a voice section in step 3, the noise spectrum N(f) is estimated. Step 7: Calculate the noise removal coefficient α.

[0015]

[Effects] As is clear from the above description, the present invention has the following effects. In other words, if the spectral subtraction method is used to remove noise when there is a lot of ambient noise, the noise spectrum during noise estimation and the noise spectrum during speech input will be significantly different, making it difficult to perform appropriate noise removal. However, in the present invention, spectral subtraction is performed according to the stationarity of the surrounding noise. Since we are changing the proportion of
There is less possibility of misrecognition than with conventional methods.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a noise removal device according to the present invention.

FIG. 2 is a flowchart for explaining the operation of the noise removal device according to the present invention.

FIG. 3 is a diagram illustrating an example where noise removal is not performed appropriately.

[Explanation of symbols]

1... Acoustic input section, 2... Frequency spectrum detection section, 3... Section detection section, 4... Noise stationarity calculation section, 5... Noise removal coefficient calculation section, 6... Noise spectrum estimation section, 7... Noise correction section, 8
...Speech recognition section.

Claims (2)

[Claims] 1. Detection means for determining the frequency spectrum of a voice input signal; detection means for determining the frequency spectrum of noise from an input signal before voice input; and detection means for determining the frequency spectrum of the noise from the frequency spectrum of the input signal at the time of voice input. In a noise removal device having a noise removal means that performs noise removal by subtraction, a calculation means for determining the stationarity of surrounding noise from the frequency spectrum of the noise, and according to the stationarity determined by the calculation means, 1. A noise removal device comprising: noise correction means for changing the rate at which noise is removed. 2. The calculation means for determining the stationarity of the surrounding noise is characterized in that the calculation means calculates the stationarity from the time change of the frequency spectrum of the input signal when the noise is input, and the smaller the time change is, the higher the stationarity is. 2. The noise removal device according to claim 1.