JP3091244B2 - Noise removal device and speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise elimination device and a speech recognition device using the noise elimination device, and more particularly, to a technology for removing ambient noise components mixed into a speech signal from speech input to the speech recognition device. And speech recognition using the technology in a noisy environment, such as in an office, in a car, or in a factory.

[0002]

2. Description of the Related Art In a speech recognition device, when ambient noise is mixed in an inputted speech, the recognition rate is remarkably lowered. Therefore, removal of the ambient noise is an important issue when the speech recognition device is put into practical use. As a method of denoising is to use two acoustic input of speech input and the noise input unit known in such Sho 63-2 6 2695 and JP-A-1-239596. This is a method of removing a noise component corresponding to a feature amount of a signal input from a noise input unit from a feature amount of a signal input from a voice input unit. However, in this method, noise removal is performed by subtracting a feature amount of noise obtained from a signal input from the noise input unit from a feature amount of voice including noise input from the voice input unit. If the noise input at the noise input unit is significantly different, for example, there are a plurality of noise sources,
For example, when two noise sources are near the noise input unit, appropriate noise removal cannot be performed, and the speech recognition rate may be reduced instead.

FIG. 4 shows an example in which noise removal is not properly performed. As shown in FIG. 4 , a noise component (a) included in a voice input signal and a spectrum (b) of the noise input signal are shown. ) Is significantly different, the result is as shown in (d) when noise is removed, and the spectrum (c) before the noise removal is more similar to the spectrum (e) of the voice than (d). .

[0004]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned circumstances, and in particular, has been made to prevent improper noise removal when the noise between the voice input unit and the noise input unit is significantly different. It was made for the purpose.

[0005]

To achieve the above object, the present invention provides (1)
A first sound input unit for inputting speech including noise,
A second sound input unit for inputting noise, a first feature extraction unit for obtaining a feature amount of a speech signal containing noise input from the first sound input unit, and the second sound input unit A second feature extraction unit for obtaining a feature amount of the noise signal input from the second unit; and a noise component from the feature amount of the noise-containing speech signal by subtracting the feature amount of the noise signal from the feature amount of the speech signal containing noise. A noise elimination unit having a noise elimination unit for elimination, a characteristic amount of an input signal of the first sound input unit in a section where no voice input is performed,
Estimating a noise component of the feature amount of the audio signal including noise, if the feature amount of the noise component and the noise signal are largely different, and child cease to remove the noise, Or, (2) the 1 A noise removing device, a pattern creating unit that creates an audio input pattern from a feature amount of a voice obtained by the noise removing device, and a standard pattern memory that stores a standard pattern of the voice registered in advance. , and this for performing recognition processing by using the reference pattern and the input pattern obtained by the pattern preparing unit the stored in the standard pattern memory
It is characterized by. Hereinafter, a description will be given based on an example of the present invention.

FIG. 1 is a block diagram for explaining an embodiment of the first aspect of the present invention, and FIG. 2 is a flowchart for each frame for explaining an operation principle thereof. A first sound input unit 1 for inputting speech including noise, and a second sound input unit 2 for inputting noise
A first feature extraction unit 3 for obtaining a feature amount of a speech signal including noise input from the first sound input unit; and a feature amount of a noise signal input from the second sound input unit. The second feature extraction unit 4, the speech section detection unit 5, the noise similarity calculation unit 6, the noise removal determination unit 7, and the feature amount of the noise signal is subtracted from the feature amount of the speech signal containing noise. A noise removing unit that removes a noise component from a feature amount of a speech signal including noise, wherein a feature amount of an input signal of the first sound input unit in a section where no speech input is performed is obtained; This is estimated as a noise component of a feature amount of a speech signal including noise, and when the noise component greatly differs from the feature amount of the noise signal, the removal of the noise is stopped.

More specifically, the sound input unit 1 for sound input converts sound into an electric signal x (t) using an acoustic-electric signal converter such as a microphone installed near the mouth of the speaker. I do. Note that not only voice but also ambient noise is mixed in this x (t). The sound input unit 2 for noise input uses a sound-electric signal converter such as a microphone installed at a position where it is difficult to input a voice uttered by a speaker from the sound input unit 1 for voice input, and reduces ambient noise. It is converted into an electric signal n (t) (t is a variable indicating time). The voice input feature extraction unit 3 uses a band-pass filter group, FFT, or the like, and outputs a short-time frequency spectrum X (t) of about 10 msec of the electric signal x (t) obtained by the voice input audio input unit 1. , F) (f is a variable indicating frequency). The noise input feature extraction unit 4 uses a band-pass filter group or FFT or the like to generate an electric signal n obtained by the noise input sound input unit 2.
(T) Short-time frequency spectrum N of about 10 msec
A feature amount of noise such as (t, f) is extracted. The voice section detection unit 5 detects whether or not voice input is being performed. For example, the voice section detects a section in which the short-time average value of the absolute value of x (t) is equal to or larger than a predetermined threshold. Is used. As another method, there is a method in which a voice input switch is provided, and a section in which the switch is turned on is during voice input. The noise similarity calculator 6 estimates a noise component included in the feature X (t, f) of the voice input signal, and the voice section detector 5 determines a section (t ₁₎ determined not to be the voice input section. To t ₂ ) voice input feature extraction unit 3
Is obtained as a noise component Y (t, f) in the speech input signal, and is similar to N (t, f) extracted by the noise input feature amount extraction unit 4. There is a method of obtaining the degree as in the following equation (1).

[0008]

(Equation 1)

Note that k in the above equation (1) is a coefficient for performing a level correction obtained from a level ratio between two input sections in a section where no voice is input. The noise removal determining unit 7 extracts the noise feature extracted by the noise input feature extracting unit 4 and the noise input feature extracting unit 3 from the noise similarity Z obtained by the noise similarity calculating unit 6. It is determined whether or not the noise components included in the feature amount are significantly different. For example, when the noise feature amount Z is equal to or larger than a predetermined threshold value, it is determined that the noise component is greatly different, and the noise removal is not performed. There is a way. The noise elimination unit 8 performs noise elimination when the noise elimination determination unit 7 determines that the noise is to be eliminated. Feature X
Using (t, f) and the noise input feature N (t, f), S (t, f) obtained by S (t, f) = X (t, f) -kN (t, f) ) Is a noise feature amount from which noise has been removed.

[0010]

FIG. 3 is a block diagram for explaining one embodiment of the invention described in claim 2. In the drawing, reference numerals 1 to 8 in FIG. 3 operate in the same manner as in the embodiment shown in FIG. Voice recognition unit 9
Is the speech feature amount S (t,
f) performs voice processing using "binary TSP".
Development of a Word Speech Recognition System Using ”(Yasuda et al., The Institute of Electrical Engineers of Japan, C108, October 1988, pp.8)
58 to 865), but can be realized by using other known speech recognition systems.

[0012]

[0013]

(1) The function and effect corresponding to the first aspect: when the noise component included in the voice signal and the feature amount of the noise signal are largely different, the noise removal is stopped, so that it is inappropriate. No noise is removed. Since no noise is removed, inappropriate noise removal is not performed. (2) A function and effect corresponding to the second aspect : a speech recognition apparatus in which the recognition rate does not decrease even when the noise component included in the speech signal and the feature amount of the noise signal are significantly different by combining the speech recognition unit. Can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining an embodiment of the invention described in claim 1;

FIG. 2 is a flowchart for each frame for explaining the operation principle of the embodiment shown in FIG. 1;

FIG. 3 is a block diagram for explaining one embodiment of the invention described in claim 2 ;

FIG. 4 is a waveform diagram for explaining an example of a case where noise removal cannot be performed properly.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Speech input sound input part, 2 ... Sound input sound input part, 3 ... Speech input feature extraction part, 4 ... Noise input feature extraction part, 5 ... Speech section detection part, 6 ... Noise similarity calculation part , 7
... Noise removal determination unit, 8. Noise removal unit, 9.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/20 G10L 21/02 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A first sound input unit for inputting speech including noise, a second sound input unit for inputting noise, and a noise input from the first sound input unit. A first feature extraction unit for obtaining a feature amount of the audio signal;
A second feature extraction unit for obtaining a feature amount of a noise signal input from the audio input unit of the first embodiment; and a feature amount of the noise signal obtained by subtracting the feature amount of the noise signal from the feature amount of the speech signal containing noise. A noise removing unit for removing a noise component, wherein a characteristic amount of the input signal of the first sound input unit in a section where no voice input is performed is obtained, and the characteristic amount of the input signal of the noise-containing audio signal is obtained. A noise component that is estimated to be a noise component of a noise signal and stops performing noise removal when the feature amount of the noise component is significantly different from that of the noise signal. 2. A noise removing apparatus according to claim 1, a pattern creating section for creating a speech input pattern from a feature amount of the speech obtained by said noise removing apparatus, and a standard pattern of speech registered in advance. And a standard pattern memory for storing voice data, and performing a voice recognition process using the input pattern obtained by the pattern creation unit and the standard pattern stored in the standard pattern memory. apparatus.