JPH05165491A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To detect a voice block with high accuracy even under noise and to improve recognizing performance and reliability. CONSTITUTION:An acoustic analysis part 1 acoustically analyzes an input voice and calculates the feature parameter. Based on the feature parameter, a partial pattern detection part 2 detects the partial pattern of the input voice by calculating similarity between the part of the input voice and the standard pattern of a voice previously registered on a keyword standard pattern memory 21. With the position of the partial pattern as a reference, a word border hypothesis generating part 3 calculates the plural block candidates of the input voice and generates the feature pattern for each block. A recognition part 4 recognizes the input voice by calculating similarity between each feature pattern generated by the word border hypothesis generating part 3 and the standard pattern of a voice previously registered in a word standard pattern memory 41. Thus, the high recognizing performance can be presented even under noise.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device.

[0002]

2. Description of the Related Art Speech recognition technology plays an important role in realizing an excellent man-machine interface. In this speech recognition, there is speech section detection as an important pre-processing for improving the recognition performance, and various researches and developments have been conventionally performed. When considering practical use of a voice recognition device, improvement of noise resistance is also an important issue, but voice segment detection is particularly problematic in voice recognition under noise.

In the speech section detection in the conventional speech recognition apparatus, the power time series of the input speech is exclusively obtained, and the start point S is the time when the value of the speech power becomes larger than a predetermined threshold value T1.
In addition, the time point when the voice power value becomes smaller than the predetermined threshold value T2 after the detection of the start point is detected as the end point E.

[0004]

However, in such a voice section detection, since the voice section is uniquely determined, for example, if there is a breathing noise or a tongue noise before and after the actual voice section, this is detected. However, there is a problem that it is detected as a part of the voice section.

On the other hand, in the case of a word voice in which the beginning and end syllables of a syllable are likely to be unvoiced, the power of the unvoiced syllable portion is extremely small, and this portion is easily dropped from the detection of the voice section. There was a problem.

Such an error in voice section detection causes fatal misrecognition in the voice recognition.

Therefore, the present invention has been made in view of the above circumstances, and provides a voice recognition device capable of detecting a voice segment with high accuracy even in a noisy state and improving recognition performance and reliability. With the goal.

[0008]

In order to achieve the above object, the present invention is based on the result of acoustic analysis of the input voice, and calculates the similarity between the part of the input voice and the standard pattern of the corresponding voice A partial pattern detection unit that detects a partial pattern, and a word boundary that generates a plurality of voice section candidates of the input voice based on the position of the partial pattern detected by the partial pattern detection unit and generates a characteristic pattern for each section The present invention is characterized by having a hypothesis generating unit and a recognizing unit that recognizes the input voice by calculating the similarity between each characteristic pattern generated by the word boundary hypothesis generating unit and the standard pattern of the corresponding voice. It is a thing.

[0009]

The operation of the invention having the above construction will be described.

The partial pattern detector detects a partial pattern of the input voice by a similarity calculation based on the acoustic analysis result of the input voice. The word boundary hypothesis generation unit obtains a plurality of voice section candidates of the input voice based on the position of the partial pattern detected by the partial pattern detection unit. Next, the word boundary hypothesis generation unit generates a characteristic pattern for each voice section obtained as a candidate. The recognition unit recognizes the input voice by a similarity calculation based on each feature pattern generated by the word boundary hypothesis generation unit.

In this way, a plurality of voice section candidates are obtained based on the partial pattern obtained by acoustically analyzing the input voice,
Since the input voice is recognized by the similarity calculation, high recognition performance can be exhibited even in the presence of noise.

[0012]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a voice recognition apparatus according to an embodiment of the present invention. The apparatus includes an acoustic analysis unit 1 that acoustically analyzes an input voice to obtain a characteristic parameter thereof, a partial pattern detection unit 2 that detects a partial pattern of the input voice based on the characteristic parameter obtained by the acoustic analysis unit 1, and this portion. A word boundary hypothesis generating section 3 for obtaining a word section candidate (L3) as a plurality of speech section candidates of the input speech based on the position of the partial pattern detected by the pattern detecting section 2 and generating a characteristic pattern for each section. And a recognition unit 4 for recognizing the input voice based on each characteristic pattern generated by the word boundary hypothesis generation unit 3.

FIG. 2 is a diagram showing an example of a word set in this embodiment. The word set includes the same keyword "kai". FIG. 3 is a diagram showing, in shades, an example of the characteristic parameter when "nikai" occurs. Each part will be described below.

The acoustic analysis unit 1 obtains a voice power time series as a feature amount used for detecting a voice section from an input voice, and, for example, a frequency-analyzed band as a feature amount used for matching with a keyword standard pattern. The output of the pass filter group is obtained, and these characteristic quantities are sent as characteristic parameters of the input voice to a keyword detecting section 20 of the partial pattern detecting section 2 which will be described later.

The partial pattern detector 2 comprises a keyword detector 20 and a keyword standard pattern memory 21.

In the keyword standard pattern memory 21,
A keyword standard pattern used when the keyword detection unit 20 detects a keyword is registered in advance. The registered content is obtained by acoustically analyzing a known input voice containing a keyword, and for example, as shown in FIG. 9 described later, two keywords “kai” (/ kai /) are included in the elements / ka / and / a.
This is a keyword standard pattern divided into i /.

The keyword detection unit 20 detects a rough occurrence section L1 in which the input voice is included and a keyword included in the section L1 by using the characteristic parameters obtained from the acoustic analysis section 1. After detecting these, the detection result and the characteristic parameter of the input voice from the acoustic analysis unit 1 are output to the word boundary hypothesis generation unit 3 described later.

The detection of the generation section L1 by the keyword detection unit 20 is conventionally performed by setting the generation section L1 from the time point when the value of the voice power exceeds a predetermined threshold value T1 to the time point when it falls below the threshold value T2. It is obtained by the method of detecting the voice section used from. This generation section L
1 corresponds to the section from S to E in FIG.

Further, the keyword detection unit 20 detects a keyword included in the occurrence section L1 by detecting a keyword (for example, "kai") from the obtained occurrence section L1. Specifically, as shown in FIG. 8 to be described later, the segment 22 is sequentially cut out from the occurrence segment L1 while being shifted in the time direction (every frame j), and the characteristic pattern and the keyword standard pattern of this segment are extracted. The section L2 of the keyword "kai" is detected from the difference between the similarity value with (/ ka /, / ai /) and the similarity value. In FIG. 3, the section L2 from KS to KE corresponds to this.

The word boundary hypothesis generating unit 3 uses the position of the keyword (for example, "kai") detected by the keyword detecting unit 20 as a reference to determine the characteristic parameter of the input voice obtained by the acoustic analyzing unit 1. On the other hand, various word section detection parameters are adaptively set, and a plurality of word section candidates (L3) are set as described above. That is, the position of the beginning of the keyword "kai" found by the keyword detector 20 with respect to the input voice is used as the ending candidate of the word section L3, and a plurality of starting candidates (S1, S2) are generated before this.
, ~ SM) and set a plurality of word segment candidates (L3). In FIG. 3, L30 (S1, KS), L31 (S2,
KS) becomes a word section candidate (L3). Further, the generation unit 3 further generates a feature pattern in which each word segment candidate (L3) is time-normalized, and the similarity calculation unit 40 is generated.
I am sending it to. Note that word segment candidates (L3
) May be set before, after, or after the keyword according to the keyword obtained by the keyword detection unit 20.

The recognition section 4 includes a similarity calculation section 40, a word standard pattern memory 41, and a recognition result output section 42.

The similarity calculation unit 40 calculates the similarity between each characteristic pattern of the word segment candidates (L3) obtained as described above and each standard pattern of the recognition target word registered in the word standard pattern memory 41 in advance. Each is calculated and sent to the recognition result output unit 42. The similarity calculation here is performed using a composite similarity method, a mixed similarity method, or the like that has been conventionally proposed.

In the word standard pattern memory 41, standard patterns of recognition target words are registered in advance. The standard pattern of this recognition target word corresponds to the category name of the word voice by acoustically analyzing the input voice whose category name is known in advance and normalizing the occurrence time length of the word section L3 to extract the characteristic pattern. Created with. In the present embodiment, voice characteristic patterns corresponding to "ni" of "nikai" and "san" of "sankai" are registered as word standard patterns, but "nikai" and "san" are registered. A characteristic pattern of the entire word section L3 such as "Kai" may be used as the standard word pattern. In this case, the end position of the keyword "kai" is set as the end candidate of the word section L3.

The recognition result output section 42 is a similarity calculation section 40.
All the degrees of similarity obtained in (1) are compared and the highest or highest category names or category numbers are output as a result.

The keyword detection, which is a feature of this apparatus, will be described in more detail with reference to FIGS. 4 to 10.
FIG. 4 to FIG. 7 show a flowchart for detecting the keyword “kai”. FIG. 8 shows the positions of the segments 22 when the keyword standard patterns “/ ka /” and “/ ai /” are created. In FIG. 9, the segment 22 is shifted in the time direction from the occurrence segment L1 and sequentially cut out, and the similarity S between the characteristic pattern of this segment and the keyword standard patterns "/ ka /" and "/ ai /" is calculated. It is illustrated that the calculation of ^(ka) j and S ^(ai) j is performed for each frame j. FIG. 10 is a diagram showing the processing steps of FIGS. 4 to 7 for the characteristic pattern of “nikai”. The operation of the keyword detecting unit 20 will be described with reference to FIGS. 4 to 7. First, as a preparation process, each parameter is initialized (S1). That is, as shown in FIG.
kacnt = 0, kaatta = 0, kaj = 0, ai
cnt = 0, aiatta = 0, aij = 0, kaia
Perform tta = 0. Next, the detection processing start frame j
Is set to S (starting point of the rough generation section L1 described above) (S2), and / ka / detection processing is performed (S3).
). This processing is performed on the frame j as shown in FIGS.
Between the input feature pattern and the keyword standard pattern “/ ka /” and “/ ai /” in S ^(ka) j,
S ^(ai) j is calculated, and if S ^(ka) j is larger than the constant S _KA and SD _j (= | S ^(ka) j, S ^(ai) j |) is larger than the constant S _DF (S301), kacnt is incremented (S303). Otherwise, kacnt is cleared to zero (S302) and the process returns. If kacnt is incremented, k is larger than the constant KA.
Set the aatta flag to 1 (S304, S305),
Assuming that / ka / has been detected, the central frame position kaj (= j-kacnt / of the section of the segment / ka /
2) is calculated (S306). After the above processing, ka
When atta = 1 and kacnt = 0 [/ ka / section detection is completed] (S4), / ai / detection processing is performed in the same manner as / ka / detection processing (S5, FIG. 7). That is,
Similarity S between the input feature pattern in frame j and the keyword standard patterns “/ ka /” and “/ ai /”
^(ka) j, S ^(ai) j are calculated such that S ^(ai) j is larger than the constant S _AI and SD _j (= | S ^(ka) j, S ^(ai) j |) is the constant S _AI.
If it is larger than _DF (S501), aicnt is incremented (S503). Otherwise, aicnt is cleared to zero (S502) and the process returns. If aicnt is incremented and aicnt is larger than the constant AI, the flag of aiatta is set to 1 (S504, S
505), / ai / is assumed to be detected and segment /
The central frame position aij (= j-aic in the section of ai /
nt / 2) is calculated (S506). Step S4
If kaatta = 1 and kacnt = 0 [/ ka / section detection is completed], the frame is incremented (S9), and the process is repeated from step S3 to j = E (S10). After the processing of step S5, / ai
If / is detected and aiattta = 1, then kaia
If tta = 1 (S6, S7), then kaiatta
= 1 and aicnt = 0 [/ ai / section detection ends] (S8), the detected central frame positions of the sections 22 of / ka / and / ai / are ka.
It is stored in j, aij, and the process ends. If not, it is determined that the "kai" detection has failed, and processing such as reject is performed and the processing ends (S11).

In FIG. 10, the actual keyword "kai" detection processing is performed along the flow of the above processing, and / ka /, / a
The positions of the segments 22a and 22b for detecting i / and the similarity S ^(ka) j, S ^(ai) j and SD for each frame
_j (= | S ^(ka) j, S ^(ai) j |) is displayed.

The effectiveness of the above keyword "kai" detection processing, which is a feature of the present invention, will be described with reference to FIG. S
The values of ^(ka) j and S ^(ai) j are large to some extent even at frame positions other than / kai /. Therefore, S ^(ka) j
, S ^(ai) j values are stable and highly accurate / kai
It is easily expected that / will be difficult to detect.
However, when attention is paid to the difference SD _j between S ^(ka) j and S ^(ai) j, it can be seen that a larger and more stable value appears at the positions of / ka / and / ai /. This is because by learning the patterns of / ka / and / ai / when creating a keyword standard pattern, the / ka / interval (22
In a), not only does S ^(ka) j increase, but S ^(ai) j
As a result, the effect of reducing the value of SD _j appears, and SD _j has a stable and large value. Similarly, in the section (22b) of / ai /, not only S ^(ai) j increases but also S ^(ka) j is reduced, so that SD _j has a stable and large value. Also, a section other than / kai /, for example, / n
In the i / section (22c), the value of S ^(ai) j is large,
^Since the value of ^(ka) j also becomes large and SD _j becomes small, the effect that erroneous detection can be avoided is obtained. That is, S
^{In addition to (ka)} j −S ^(ai) j, the difference SD _j between S ^(ka) j and S ^(ai) _j
By using (= | S ^(ka) j, S ^(ai) j |), it becomes possible to detect stable and highly accurate / kai / easily.

According to the voice recognition apparatus of the above embodiment, when the vocabulary set to be voice-recognized includes the same keyword or a partial pattern forming a voice, the keyword or the partial pattern is detected in advance. After that, by detecting the desired voice section based on that position, it is possible to realize a much more highly accurate voice section detection as compared with the conventional voice section detection method, and to add before and after the desired voice section. It is also excellent in terms of noise resistance because it can deal with noise and unnecessary words that are generated, and it becomes possible to detect the basic keywords that are detected with high accuracy and stability, improving the recognition performance and reliability of the device. It is possible to obtain a great effect in practical use.

The present invention is not limited to the above embodiment,
Various modifications can be made without departing from the spirit of the invention.
For example, in the present embodiment, only "kai" is used as a keyword, but "gai" and the like may be added to the keyword as necessary, and by adding a keyword such as "10", words on the 10th floor and above It can be applied to have multiple keywords. In addition, "kai (/ ka
"i /)" is divided into two elements, / ka / and / ai /, but it may be divided into arbitrary plural elements such as / ka /, / a /, / ai /.

Further, even when the vocabulary set to be voice-recognized does not include the same keyword or a partial pattern forming a voice, the keyword of the present invention is used as means for detecting one or more specific keywords. The detection method is effective.

[0032]

As described above, according to the present invention, a plurality of voice segment candidates are obtained based on the partial pattern obtained by acoustically analyzing the input voice, and the input voice is recognized by the similarity calculation. As a result, it is possible to provide a voice recognition device with which it is possible to detect a voice segment with high accuracy even in a noisy condition, and to improve recognition performance and reliability.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram showing an embodiment of a voice recognition device of the present invention.

FIG. 2 is a diagram showing an example of a word set in the present embodiment.

FIG. 3 is a diagram showing an example of a characteristic parameter of voice in gray scale in the present embodiment.

FIG. 4 is a flowchart showing a process of detecting a keyword “kai”.

FIG. 5 is a flowchart showing a process of detecting a keyword “kai”.

FIG. 6 is a flowchart showing a process of detecting a keyword “kai”.

FIG. 7 is a flowchart showing a process of detecting a keyword “kai”.

FIG. 8: Standard keyword patterns (“/ ka /” and “/
It is a figure which shows the position of the segment at the time of creating ai / ").

FIG. 9 shows the calculation of the similarity S ^(ka) j, S ^(ai) j in the frame j.
It is the figure which showed a mode that it performed for every.

FIG. 10 is a diagram illustrating an example of “kai” detection processing.

[Explanation of symbols]

 1 acoustic analysis unit 2 partial pattern detection unit 3 word boundary hypothesis generation unit 4 recognition unit

Claims (1)

[Claims] 1. A partial pattern detection unit for detecting a partial pattern of an input voice by calculating a degree of similarity between a portion of the input voice and a standard pattern of the voice corresponding thereto based on an acoustic analysis result of the input voice, and this partial pattern detection. Is generated by the word boundary hypothesis generating unit that obtains a plurality of voice section candidates of the input voice based on the position of the partial pattern detected by the section and generates a feature pattern for each section. A voice recognition device, comprising: a recognition unit that recognizes the input voice by calculating the degree of similarity between each characteristic pattern and the corresponding standard pattern of the voice.