JPH0431896A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To reduce the occurrence of speech recognizing errors due to wrong voice detection by performing matching to voice sections having difference detecting levels for comparison. CONSTITUTION:A voice detecting section 2 is set so that the section 2 can regard a somewhat high noise level (for example, 10 times) as a voice section and another voice detecting section 6 is set so that the section 6 can regard a somewhat low noise level (for example, 2 times) as a voice section. A matching section 4 performs matching to the voice section detected by the section 2 and another matching section 7 performs matching to the voice section detected by the section 6. A discriminating section 8 compares the recognized results of the sections 4 and 7 with each other and discriminates that which result is closer to a standard pattern and outputs a discriminated result. Therefore, dropping of recognition rate caused by wrong voice detection can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and more particularly to a speech recognition device that aims to improve the reduction in recognition rate caused by a speech detection section.

[Conventional technology]

A block diagram of a conventional speech recognition device is shown in FIG.
In the figure, 1 is a microphone, 2 is a voice detection unit that determines a voice section, 3 is a voice analysis unit that extracts feature vectors that are voice feature parameters, 4 is a matching unit that matches standard patterns and feature vectors, and 5 is a standard This is a standard pattern memory that stores patterns. The voice input to the microphone 1 is input to the voice detection section 1 and the voice analysis section 2. The audio detection unit 1 detects the level of input audio, and determines an effective audio section based on the level of the input audio. The speech analysis unit 3 digitally converts the input speech and extracts a feature vector for each analysis frame. The matching unit 4 matches the feature vector for each analysis frame in the voice section output by the voice detection unit 1 with the standard pattern stored in advance in the standard pattern memory 5, and determines the standard pattern closest to the input voice. and output it as a recognition result.

[Problem to be solved by the invention]

The above-described conventional speech recognition device has a drawback that when inputting speech with a weak beginning or end of a word, the beginning or end of the word cannot be accurately detected due to surrounding noise, resulting in incorrect speech recognition results.

For example, in the case of the sound "san", the sound level of the "su" part at the beginning of the word is low, so if the surrounding noise level is high, the "su" part cannot be considered as speech, and the sound level of the "su" part at the beginning of the word is low.
Only that part is detected as audio.

Additionally, if you set a low voice detection threshold in an attempt to accurately detect such voices, non-voice parts, such as the "tongue-clicking sound" before the voice is uttered, will also be detected as voice. However, this method has the disadvantage that the result of speech recognition is incorrect.

[Means to solve the problem]

The device of the present invention includes a speech detection section that detects a speech section of input speech, a speech analysis section that obtains a feature vector of input speech, a standard pattern memory section that stores a standard pattern, and a speech detection section that detects a speech section of input speech. A speech recognition device comprising a matching section that performs matching with a matching section to recognize input speech, includes a plurality of speech detection sections that detect speech sections having mutually different speech detection levels, and a plurality of speech detection sections corresponding to the plurality of speech detection sections. The pattern includes a plurality of disposed matching sections or one matching section that is used in a time-sharing manner, and a judgment section that compares the recognition results of the plurality of matchings by the matching section, judges the one closest to the standard pattern, and outputs the result. .

〔Example〕

Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a speech recognition device according to the present invention, and takes as an example a case in which there are two speech detection sections and two matching sections. In FIG. 1, 1 is a microphone, 2.6 is a voice detection section having different detection levels, 3 is a voice analysis section, 4.7 is a matching section, 5 is a standard pattern memory, and 8 is a judgment section. The basic operation of each section in FIG. 1 is the same as that in FIG. 4, but the voice detection levels of voice detection section 2 and voice detection section 6 are different.

In this embodiment, the voice detection unit 2 is set to consider a level that is somewhat higher than the noise level (for example, 10 times) as a voice section, and the voice detection unit 6 is set to consider a level that is smaller than the noise level (for example, 2 times) as a voice section. However, it is set to be treated as a voice section. The matching unit 4 performs matching on the voice section detected by the voice detection unit 2, and the matching unit 7
The 0 determination unit 8 that performs matching for the voice section detected by the voice detection unit 6 compares the recognition result output by the matching unit 4 and the recognition result output by the matching unit 7,
Determine which result is closer to the standard pattern and output the result.

Next, the operation of this embodiment will be explained using FIGS. 2 and 3. FIG. 2 shows the speech waveform of "San" under low noise, and FIG. 3 shows the speech waveform of "San" under high noise. In FIG. 2.3, Ll is the voice detection level of the voice detection section 2, and L2 is the voice detection level of the voice detection section 6. or,
tl and tl' are voice sections detected by the voice detection section, and t2°t2' is a voice section detected by the voice detection section 6. As shown in FIGS. 2 and 3, the speech detection section 2 can accurately detect the speech section under low noise, and the speech detection section 6 can accurately detect the speech section under high noise. In the case shown in FIG. 3, the voice detection section 2 detects the part "n", and the voice detection section 6 detects the part "r sun". These voices are compared with the standard patterns in the standard pattern memory section 5 by matching sections 4 and 7, respectively. Since the standard pattern memory section 5 stores the standard pattern r-san J, the matching section 7 outputs "san" as the recognition result. The matching unit 4 outputs the standard pattern closest to "n", for example "yon", as a recognition result. The determining unit 8 compares the degree of similarity between "san" and "san" and the degree of similarity between "n" and "yon", and determines "san" with the greater degree of similarity as the recognition result.

In this embodiment, a plurality of matching sections are provided, but this is for performing matching in real time. When there is only one matching section 1, one matching section is used for time division.

〔Effect of the invention〕

As explained above, the present invention has a plurality of voice detection units having different voice detection levels, and a plurality of matching units corresponding to each voice detection unit, or one matching unit for time-sharing use, to match voice segments with different detection levels. By performing matching and comparing the results, it is possible to significantly reduce errors in speech recognition due to errors in speech detection.

FIG.

1...Microphone, 2.6...Audio detection section, 3
...Voice analysis section, 4.7...Matching section, 5...
- Standard pattern memory section, 8... Judgment section, Ll... Detection level of the voice detection section 2, L2... Detection level of the voice detection section 6, tl, tl'... Voice of the voice detection section 2 Detection interval, t2. t2'...Audio pressure detection section of the audio detection unit 6.

Claims (1)

[Claims] A speech detection section detects speech sections of input speech, a speech analysis section obtains feature vectors of input speech, a standard pattern memory section stores standard patterns, and performs input by matching the feature vectors with standard patterns. A speech recognition device including a matching section that recognizes speech includes a plurality of speech detection sections that detect speech sections having mutually different speech detection levels, and a plurality of speech detection sections arranged corresponding to the plurality of speech detection sections. One matching part for split use,
A speech recognition device comprising: a determining section that compares recognition results of a plurality of matchings by the matching section, determines the one closest to a standard pattern, and outputs the same.