JP3114757B2 - Voice recognition device - Google Patents

Abstract

PURPOSE:To provide the voice recognizing device improving reject ability against noise by considering a pitch cycle peculiar for voices in the case of a recognition processing so as to recognize only the voices in the input sound mixing noise and voices. CONSTITUTION:This device is composed of a recognition part 1 to output voice recognizing candidates in the blocks of a signal estimated as voice parts in the input audio signal, pitch extraction part 2 to extract the pitch cycle from the input audio signal, and judge part 3 to output the voice recognized result from the voice recognizing candidates to the signal of the blocks outputted by the recognition part 1 and the pitch extracted result of the signal in that block by the pitch extraction part 2.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device for recognizing speech input by input means. Today, voice recognition devices are increasingly being used to input data into computers, make reservations by telephone, control the progress of work in steelworks and car factories, and the like.

For this reason, in order to recognize only speech from input speech in which speech and noise are mixed, a speech recognition device having a high noise rejection capability is required.

[0003]

2. Description of the Related Art FIG. 3 shows a conventional example. The input unit 101 is a sound collection device such as a microphone, and the input sound is converted into an analog signal, and then the analog signal is converted into a digital signal by the A / D converter 102.

[0004] The feature extracting unit 103 divides the digitized signal at a fixed time, further divides a predetermined frequency range into several frequency bands, and divides each frequency band at each divided time. The feature is extracted from the signal value for each. The section detection unit 104 detects a section of a signal that satisfies a threshold for a preset signal value as a recognition processing section that performs a process of recognizing as speech. In the process of detecting the recognition processing section, if the signal value changes from a state that satisfies the threshold to a state that does not satisfy the threshold, and after a certain time elapses, a state that satisfies the threshold again, a time when the threshold is not satisfied Is within a preset time,
The section of the signal that does not satisfy the threshold is also detected as the recognition processing section.

[0005] The dictionary 105 stores a standard pattern of speech created in advance, and stores in the dictionary 105 the features of the signal extracted by the feature extraction unit 103 in the recognition processing section detected by the section detection unit 104. The collation unit 106 calculates distances to all the standard patterns that have been set. The standard pattern having the smallest distance among the distances calculated by the distance calculation performed by the matching unit 106 is output as a recognition candidate for the signal of the recognition processing section detected by the section detection unit 104.

Thereafter, the minimum distance obtained by the collating unit 106 in the recognition judging unit 108 is stored in the distance threshold storing unit 10.
7, it is determined whether or not the minimum distance satisfies the threshold. If it is determined that the minimum distance satisfies the threshold, the signal of the recognition processing section detected by the section detection unit 104 is recognized as speech. Then, the recognition candidate output by the matching unit 106 is output as a recognition result. If it is determined that the threshold is not satisfied, the signal in the recognition processing section is rejected as noise and re-input of speech to the input unit 101 is performed. Becomes

[0007]

As described above, the pattern of the signal of the recognition processing section detected by the section detection section 104 obtained by the recognition determination section 108 and the pattern of the recognition candidate output by the matching section 106 Is the distance threshold storage unit 10
If it is determined that the threshold for the distance stored in 7 is satisfied, the recognition candidate is output as a recognition result, and if it is determined that the threshold is not satisfied, it is rejected as noise.

Accordingly, by setting the threshold value for the distance stored in the distance threshold value storage unit 107, the section detection unit 1
It is determined whether the signal of the recognition processing section detected in 04 is recognized as speech or rejected as noise. The threshold value for the distance stored in the distance threshold value storage unit 107 is obtained experimentally or empirically. If the threshold value is set low, even if the signal in the recognition processing section where the section is detected is noise, the matching unit 106 sets the threshold value. Since the obtained minimum distance satisfies the threshold for the distance, the matching unit 106
In some cases, the recognition candidate obtained in step (1) is incorrectly recognized as speech.

When the threshold value for the distance is set high, even if the signal in the recognition processing section is a voice, the matching unit 1
Since the minimum distance obtained in step 06 does not satisfy the threshold for distance, it may be rejected as noise. Therefore, the threshold for completely separating voice and noise together with the case where the threshold for distance is set low is set. Is impossible to set.

According to the present invention, there is provided a speech recognition apparatus for improving noise rejection capability by considering a pitch period peculiar to speech in recognition processing in order to recognize only speech in an input in which noise and speech are mixed. The purpose is to provide.

[0011]

FIG. 1 is a diagram illustrating the principle of the present invention. In the figure, reference numeral 1 denotes a recognition unit, which outputs speech recognition candidates for a section of a signal estimated to be a speech part from an input speech signal. Reference numeral 2 denotes a pitch extraction unit that extracts a pitch cycle from the input audio signal.

Reference numeral 3 denotes a judgment unit which outputs a speech recognition result from a speech recognition candidate for the signal of the section output by the recognition unit 1 and a pitch extraction result of the signal of the section by the pitch extraction unit 2. Reference numeral 4 denotes a period range storage unit, which stores a period range for the pitch period extracted by the pitch extraction unit 2.

[0013]

According to the present invention, as in the prior art, an input voice is converted into a signal, the characteristics of the signal are extracted, and a section in which the signal value satisfies a preset value is detected as a recognition processing section. The distance between the feature of the signal of the section and the standard pattern stored in the dictionary is calculated, the standard pattern with the minimum distance is selected as a recognition candidate, and it is determined whether or not the minimum distance satisfies a threshold for distance.

According to the first aspect of the present invention, a pitch period indicating a vibration period of a vocal cord when a voice is uttered is extracted from a signal of the recognition processing section detected in the section. When extracted, the signal in the recognition processing section is determined to be a voice candidate, and when no pitch period is extracted, it is determined to be noise. If it is determined that the minimum distance satisfies the threshold for distance, and if the signal of the recognition processing section is determined to be a speech candidate by processing related to pitch, the recognition candidate is output as a speech recognition result of the recognition processing section. I do.

The minimum distance satisfies a threshold value for the distance. If the signal in the recognition processing section is determined to be noise by the processing related to pitch, the signal in the recognition processing section is determined to be noise and rejected. I do. Therefore,
By setting the threshold value for the distance low, even if the signal of the recognition processing section detected by the section detection is noise, the minimum distance obtained by calculating the distance between the characteristic of the signal of the recognition processing section and the standard pattern is the distance. However, since the pitch period is hardly extracted from the noise, the signal in the recognition processing section is likely to be rejected as noise.

According to a second aspect of the present invention, a pitch cycle is extracted as a result of the pitch extraction processing in addition to the pitch extraction processing of the first aspect, and the pitch cycle is set to a preset cycle range with respect to a preset pitch cycle. If it is within, the signal in the recognition processing section is determined to be a speech candidate, and if it is outside the range of the pitch cycle, it is determined to be noise. If it is determined that the minimum distance satisfies the threshold for distance, and if the signal of the recognition processing section is determined to be a speech candidate by processing related to pitch, the recognition candidate is output as a speech recognition result of the recognition processing section. I do.

The minimum distance satisfies a threshold value for the distance. If the signal of the recognition processing section is determined to be noise by the processing related to the pitch, the signal of the recognition processing section is determined to be noise and rejected. I do. Therefore,
According to the second aspect of the present invention, even if the signal of the recognition processing section detected by the section detection in claim 1 is noise and the pitch cycle is extracted from the signal of the recognition processing section, the pitch cycle is not changed. Is often outside the set cycle range for the pitch cycle, the signal in the recognition processing section is rejected as noise. Therefore, it is possible to realize a speech recognition apparatus having a rejection ability with higher accuracy than the first aspect.

According to a third aspect of the present invention, a pitch cycle set in the second aspect is determined by extracting a pitch cycle and determining whether the pitch cycle is a voice. It can be changed according to the cycle. Since the distribution range of the pitch cycle differs depending on the voice input person, setting the pitch cycle setting cycle range to the pitch cycle of the voice input person makes it possible to recognize only the voice of a specific input person.

According to a fourth aspect of the present invention, the pitch period is extracted in the second aspect, and even if the extracted pitch period is within a preset period range of the pitch period, the extracted pitch period is determined. The set cycle range of the pitch cycle is automatically changed so as to be centered.
Although the pitch range of the pitch period extracted from the voice uttered by a human is limited, the pitch period range of each voice input person is different, so the pitch extracted in the recognition processing section in the voice input by the voice input user By changing the set cycle range of the pitch cycle in accordance with the cycle of the above, it becomes possible to perform optimal speech recognition processing according to the speech input person.

[0020]

Embodiments of the present invention will be described below in detail with reference to the drawings. First, a first embodiment will be described. FIG.
FIG. 3 is a diagram showing an embodiment of the present invention. Speech input by the input means 201 and conversion to an analog signal, conversion from an analog signal to a digital signal by the A / D converter 202, feature extraction of the signal by the feature extractor 203, recognition of the signal by the section detector 204 The detection of the processing section, the calculation of the distance between the standard pattern of the dictionary 205 by the matching unit 206 and the feature of the signal of the recognition processing section, and the recognition candidate output of the standard pattern of the minimum distance by the distance calculation are performed as in the above-described conventional example. , And a detailed description thereof will be omitted.

A pitch extracting section 207 shown in FIG. 2 extracts a pitch period of a signal output from the A / D converting section 202. In the recognition processing section detected by the section detection section 204, the noise / speech determination section 208 firstly outputs the pitch extraction section 2
At 07, it is determined whether or not the pitch period has been extracted. When it is determined that the pitch period has been extracted, it is determined whether or not the pitch period is within a set period range for the pitch period stored in the period range storage unit 209.

The extracted pitch period is stored in the period range storage 2
If the signal is within the set cycle range for the pitch cycle stored in step 09, the signal in the recognition processing section is determined to be a speech candidate. Is determined. When the noise / speech determining unit 208 determines that the recognition processing section detected by the section detecting unit 204 is a voice candidate, the matching unit 20
The recognition determination unit 211 determines whether or not the minimum distance obtained by the distance calculation in Step 6 meets the threshold for the distance stored in the distance threshold storage unit 210. If the minimum distance is satisfied, the recognition unit 206 outputs the result. The recognition candidate is output as a recognition result.

If the recognition determining section 211 determines that the minimum distance does not satisfy the threshold for the distance stored in the distance threshold storage section 210, the recognition processing section is determined to be noise and rejected. You. When the noise / voice determination unit 208 determines that the pitch period has not been extracted in the recognition processing section by the pitch extraction unit 207, the minimum distance obtained by the matching unit 206 is stored in the distance threshold storage unit 210. Even if the recognition determining unit 211 determines that the threshold value for the present distance is satisfied, the recognition processing section is determined to be noise and is rejected.

If rejected as noise, it will be re-input to the input means 201 as in the prior art. Next, a second embodiment will be described. The second embodiment differs from the first embodiment in that the set cycle range of the cycle range storage unit 209 can be changed.

As in the first embodiment, after the pitch period of the signal is extracted by the pitch extracting section 207, the section detecting section 2
The noise / speech determination unit 208 determines whether or not a pitch cycle has been extracted in the recognition processing section detected in step 04. If it is determined that the pitch cycle has been extracted in the recognition processing section, the noise / speech determination section 208 further determines the pitch cycle of the recognition section. Is within the set cycle range for the pitch cycle stored in the cycle range storage unit 209.

In this embodiment, the set cycle range for the pitch cycle set in the cycle range storage unit 209 can be changed, and the set cycle range of the pitch cycle preset in the cycle range storage unit 209 can be changed by the voice input user. By setting the range to a set cycle range corresponding to the pitch cycle, only a specific voice input person can perform the voice recognition processing. Next, a third embodiment will be described.

The third embodiment differs from the first embodiment in that the set cycle range for the pitch cycle of the cycle range storage unit 209 is changed according to the pitch cycle of the input voice. The description of the same processing as in the first and second embodiments will be omitted. In the present embodiment, when the signal of the recognition processing section detected by the section detection section 204 is determined to be a speech by the noise / speech determination section 208, the noise / speech determination section 208 determines whether the signal corresponds to the pitch cycle preset in the cycle range storage section 209. The set cycle range for the pitch cycle in the cycle range storage unit 209 is changed in accordance with the pitch cycle extracted from the recognition processing section.

[0028]

As described above, according to the present invention,
Since speech recognition processing with high noise rejection capability can be performed on input speech containing noise, recognition processing is performed on noise, thereby reducing the possibility of outputting incorrect recognition results. This makes it possible to improve the accuracy rate of the recognition result.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing an embodiment of the present invention.

FIG. 3 is a diagram showing a conventional example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Recognition part 2 Pitch extraction part 3 Judgment part 4 Period range storage part 101,201 Input means 102,202 A / D conversion part 103,203 Feature extraction part 104,204 Section detection part 105,205 Dictionary 106,206 Collation part 107 , 210 Distance threshold storage unit 108, 211 Recognition determination unit 207 Pitch extraction unit 208 Noise / voice determination unit 209 Cycle range storage unit

Continuation of the front page (56) References JP-A-4-115299 (JP, A) JP-A-2-238493 (JP, A) JP-A 1-159697 (JP, A) JP-A 64-40997 (JP) , A) JP-A-63-163495 (JP, A) JP-A-63-123100 (JP, A) JP-B-61-29517 (JP, B2) Furui "Digital Audio Processing" (1985-9-25) Tokai University Press p. 17-18 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-17/00

Claims (1)

(57) [Claims] A recognition unit for estimating a section including a voice signal from an input voice signal and outputting a voice recognition candidate for the signal in the section; a pitch extraction unit for extracting a pitch period from the input voice signal; A speech recognition device comprising: a decision unit that outputs a speech recognition candidate output by the recognition unit as a speech recognition result when a pitch cycle extracted by the pitch extraction unit is within a set cycle range of a cycle range storage unit. When the pitch cycle extracted by the pitch extraction unit is within a set cycle range preset in a cycle range storage unit, the cycle range set in the cycle range storage unit is changed to the extracted pitch cycle. Changing to a set cycle range based on the speech recognition device.