JPH0367279B2 - - Google Patents

Description

[Detailed Description of the Invention] [Summary] Detects speech sections from the speech power of input speech using a preset threshold, extracts speech feature parameters, and compares and verifies them with speech feature parameters of words registered in advance. In a speech recognition device that recognizes speech using a plurality of preset speech power thresholds, it calculates multiple speech sections, checks the difference in the position of the start and end of each section, and if the difference in both is small, determines the position of one section. This configuration is configured to perform comparison/verification with the registered word speech pattern only, thereby making it possible to maintain the advantages of multiple threshold values and reduce the amount of processing.

[Industrial Field of Application] The present invention relates to a word recognition method in a speech recognition device, and particularly to a speech segment recognition method.

[Prior art] A method of registering speech for each word, comparing the speech feature parameters of the input speech word with the feature parameters of each registered word, and recognizing the one with the highest degree of similarity as the recognition result. In a speech recognition device, it is first necessary to detect the speech section of a word from input speech.

In order to detect that word speech has been input and extract it as a word speech section, the power of the input speech is cut by a certain threshold, the section exceeding the threshold is detected as one word section, and its speech parameters are The word is extracted and compared with that of the registered speech, and the most similar word is used as the recognition result word.

For example, if the power of the input voice is as shown in FIG. 4, a section exceeding a certain threshold is detected as a word section, and its speech feature parameters are extracted.

To extract speech feature parameters, for example, prepare bandpass filters 1, 2, 3, etc., and extract 10ms of the input word.
The outputs of the bandpass filters 1, 2, 3, . . . for each frame are stored in a format as shown in FIG. 5, and this is called a word pattern.

This input speech pattern is compared and collated with the word patterns of each registered word stored in a similar format to find the one with the highest degree of similarity.

[Problems to be Solved by the Invention] As mentioned above, when detecting a word section using a single power threshold, depending on the relationship between the power of the input speech and the threshold, (1) (2) When a noise section is included in the speech section, and (2) When a speech section does not include low-power speech such as a devoiced syllable, this can lead to misrecognition or rejection.

Therefore, by detecting word intervals using multiple power thresholds, extracting word patterns for each, and determining the degree of similarity with the registered word patterns, we can select the correct one from among the detected multiple speech intervals. It has been confirmed that there is a high possibility that voice sections are included, and that a high recognition rate can be obtained. (For example, Proceedings of the Acoustical Society of Japan, October 1988, Kanda, Sakuraba, Sato: Simple word speech recognition method using multiple thresholds.) Figure 6 shows the speech interval detection method using multiple thresholds. There is evidence that it is effective.

A, b, and c in FIG. 6 show three registered word patterns.

Figure d shows the power pattern for input A, and with the highest threshold L ₁ ...Thresholds L ₂ and L ₃ detect that the noise section is included in the voice section and is a voice section, and the registration shown in figure b This indicates that there is a possibility that it may be mistakenly recognized as a word.

Figure 6e shows the power pattern for input B, where the correct speech section is detected with the lowest threshold L ₃ and correctly recognized with the highest degree of similarity to the registered word in Figure 6b, and the higher thresholds L ₁ , L ₂ indicates that there is a possibility that the word may be mistakenly recognized as the registered word in c of the same figure.

However, in this method using multiple thresholds, the sixth
In the case of input A shown in Figure d, even though the intervals are almost the same, input B shown in Figure e
In the case of , even though the sections of are almost the same, it was necessary to find word patterns for each and match them with the registered patterns.

Therefore, the processing required a considerable amount of time.

The present invention aims to provide a new word speech recognition device that solves this problem.

[Means for Solving the Problems] FIG. 1 shows a basic block diagram of the word speech recognition device of the present invention.

In FIG. 1, _{1 1} , 1 ₂ , . . . , 1 _n are section detection units that detect voice sections using m voice power thresholds.

Reference numeral 2 denotes a section selection section that selects only necessary ones from the m speech sections detected by the m section detection sections 1 ₁ , 1 ₂ , . . . , 1 _n .

The section detection units 1 ₁ , 1 ₂ , ..., 1 _n detect the first points S ₁ , S ₂ , ..., exceeding the respective thresholds L ₁ , L ₂ , ..., L _n ,
Starting edge detection unit 11 ₁ , 11 ₂ , ..., 11 that detects S _n
n and the last points E ₁ , E ₂ , . _{. .} , _{E n} .

After detecting the start end, the end detection units 12 ₁ , 12 ₂ , ..., 12 _n detect the start end and, even if the input audio power decreases and falls below the threshold, it increases again within a predetermined time and exceeds the threshold, it does not detect the end and detects the audio. It also performs processing to consider it as a continuation of the section.

The section selection section 2 includes section detection sections 1 ₁ , 1 ₂ ,..., 1
_n detected start ends S ₁ , S ₂ , ..., S _n and end ends E ₁ ,
The following tests are performed for E ₂ , ..., E _n .

|Si−Sj|<l ₁ |Ei−Ej|<l ₂ However, l ₁ and l ₂ are predetermined thresholds,
i, j are i=1, 2,..., m-1; j=2,
3,..., m (i<j).

When both of the above two equations hold true, the speech section (Sj, Ej) is excluded from the candidate sections, and only the remaining speech section is output as a candidate section.

A word speech pattern is generated for each candidate section output from the section selection unit 2, and compared and verified with the registered word speech pattern.

In place of the above formula, the following formula may be used for testing.

|Si−Sj|+|Ei−Ej|<l ₃ In either case, if both the start end position difference and the end end position difference are smaller than a certain value, one is excluded.

[Operation] With the above configuration, while maintaining the advantage of using multiple thresholds, almost the same speech sections can be excluded,
The amount of processing can be reduced.

[Example] According to the example shown in FIGS. 2 and 3 below,
The present invention will be explained more specifically.

FIG. 2 is a block diagram of an embodiment of the invention.

In FIG. 2, 1 ₁ , 1 ₂ , . . . , 1 _n are section detection units, and 2 is a section selection unit, which are the same objects as in FIG. 1.

3 is a power calculation unit which calculates the power of the input voice and inputs it to each section detection unit 1 ₁ , 1 ₂ , . . . , 1 _n .

4 is a microphone for audio input, 5 is an audio amplifier, and the amplified output is sent to power calculation units 3 and 6.
₁ , 6 ₂ , 6 ₃ , ..., 6 _o bandpass filters.

6 ₁ , 6 ₂ , 6 ₃ ,..., 6 _o are bandpass filters (BPF)
It is a filter that passes through each frequency band obtained by dividing the audio band into n pieces.

7 ₁ , 7 ₂ , 7 ₃ , ..., 7 _o are rectifiers and smoothers, and the respective bandpass filters 6 ₁ , 6 ₂ , 6 ₃ , ...,
6 Rectify and smooth the output from _o to produce an envelope output.

8 is a word speech pattern generation unit, which performs rectification and
Based on the outputs _of the smoothers ₇₁ , ₇₂ , ₇₃ , .

Reference numeral 9 denotes a registered pattern storage unit, which is a storage device that initially stores the voice pattern generated by the word voice pattern generation unit 8 for a desired word.

Reference numeral 10 denotes a collation unit, which compares and collates the voice pattern input from the word voice pattern generation unit 8 with each pattern in the registered pattern storage unit 9, and outputs the word with the highest degree of similarity as a recognition result.

When registering a word for the first time, switch 11
Tilt it up and input the word audio from microphone 4,
An interval detector with a standard power threshold, e.g. 1 ₂
Based on the output, a word sound pattern is generated and stored in the registered pattern storage section 9.

In the case of voice recognition, the switch 11 is turned down, the voice input from the microphone 4 is detected by the zone detection sections 1 ₁ , 1 ₂ , ..., 1 _n , and the zone selection section 2
The system selects only the necessary sections, generates a word speech pattern for this, compares and collates it with the registered word speech pattern, and outputs the recognition result.

FIG. 3 is a diagram showing an example of processing by the section selection section.

In the case of input A in figure a, there are three thresholds L ₁ , L ₂ ,
A vocal interval, , was detected by L ₃ , but
Since both the start end difference and the end end difference between the sections are small, the sections are excluded as shown by the broken lines.

In the case of input B in figure b, three thresholds L ₁ , L ₂ ,
A vocal interval, , was detected by L ₃ , but
Since both the start end difference and the end end difference between the sections are small, the sections are excluded as shown by the broken lines.

[Effects of the Invention] As explained above, according to the present invention, the processing speed can be increased by excluding almost identical speech sections and reducing the amount of processing while maintaining the advantage of using multiple thresholds. can be improved, and its practical effects are extremely large.

[Brief explanation of drawings]

Fig. 1 is a block diagram of the principle of the present invention, Fig. 2 is a block diagram of an embodiment of the invention, Fig. 3 is a diagram showing an example of processing of the section selection section, and Fig. 4 is a diagram showing the power pattern of word speech. , FIG. 5 is a diagram illustrating a word speech pattern, and FIG. 6 is a diagram showing the effect of the speech section detection method using multiple threshold values. In the drawing, 1 ₁ , 1 ₂ , ..., 1 _n are section detection parts, 11 ₁ , 11 ₂ , ..., 11 _n are start detection parts, and 1
2 ₁ , 12 ₂ , ..., 12 _n is the end detection section, 2 is the section selection section, 3 is the power calculation section, 4 is the microphone, 5 is the amplifier, 6 ₁ , 6 ₂ , 6 ₃ , ..., 6 _o is the band filter, 7
₁ , 7 ₂ , 7 ₃ ,..., 7 _o is a rectifier/smoother, 8 is a word speech pattern generation section, 9 is a registered pattern storage section,
Reference numeral 10 indicates a collation unit, and reference numeral 11 indicates a switch.

Claims (1)

[Claims] 1. A section in which the voice power of the input voice exceeds a preset threshold is detected as a voice section, and voice feature parameters within the voice section are extracted, and voice feature parameters of words registered in advance are determined. In a word speech recognition device that performs speech recognition by comparing and collating words to find the word with the highest degree of similarity, section detection detects each speech section that exceeds a speech power threshold for each of a plurality of preset speech power thresholds. and compares the positional difference between the start and end of each voice section detected by the section detecting section with a preset threshold, and when both the start and end position differences are within the threshold, one of the a section selection section that performs a process of excluding speech sections and outputting only the remaining speech sections, and comparing the speech feature parameters of the speech sections outputted by the section selection section with the speech feature parameters of the registered word. A word speech recognition device characterized in that it is configured to perform speech recognition by matching.