JPS6285300A - Word voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] A speech section is detected from the speech power of the input speech using a preset threshold value, speech feature parameters are extracted, and comparisons with speech feature parameters of words registered in advance are performed. In a speech recognition device that recognizes speech by matching, it calculates multiple speech sections using multiple preset speech power thresholds, checks the difference in the position of the start and end of each section, and if the difference between both is small, selects one section. This configuration is configured to perform comparison/verification with the registered word speech pattern only for the registered word speech pattern, 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.

Extraction of voice feature parameters is performed using bandpass filters 1 and 2, for example.
,3. ...- is prepared, and a bandpass filter is applied for each Ioms (frame) of input words. - is stored in the format shown in FIG. 5, and this is called an a-word pattern.

This human voice 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 word sections using a single power threshold, depending on the relationship between the power of the input speech and the threshold, (1) Noise sections may also be detected. If it is included in the voice section (2
) If a segment does not contain low-power sounds such as devoiced syllables, this can cause misrecognition or even jets.

Therefore, by detecting word sections using multiple power thresholds, extracting a word pattern for each, and determining the degree of similarity with the registered word pattern, we can select the correct speech section from among the detected plurality of speech sections. It has been confirmed that there is a high possibility that this will be included, and that a high recognition rate can be obtained.

(For example, Proceedings of the Acoustical Society of Japan, October 1980, Insertion, Placement, Sato: Simple word speech recognition method using multiple thresholds.) Figure 6 shows speech segment detection using this multiple thresholds. This shows the effectiveness of the method.

In FIG. 6, (a), fb), ((+1) indicate three word patterns that are registered.

Figure fd) shows the power pattern for input A, where the correct speech interval ■ is detected using the highest threshold L3 and is correctly recognized with the highest degree of similarity to the registered word in Figure (a), and the lower threshold Ll , L2, the noise section is included in the speech section and the speech sections ■ and ■ are detected, indicating that there is a possibility that the words may be mistakenly recognized as the registered words in FIG.

Figure 6(e) shows the power pattern for input B,
The correct speech interval ■ is detected using the lowest threshold L1, and the degree of similarity to the registered word fbl in the same figure is high, and it is correctly recognized, and the highest threshold L2. This indicates that there is a possibility that the word Ll may be mistakenly recognized as the registered word tc> in the same figure.

However, in this method using multiple thresholds, as shown in Fig. 6(d)
In the case of input A shown in the figure, even though the intervals between ■ and ■ are almost the same, and in the case of input B shown in the same figure (e), even though the intervals between ■ and ■ are almost the same. First, each word pattern had to be found and matched against the registered pattern.

For this reason, the processing required a considerable amount of time.The present invention aims to provide a new word speech recognition method that solves this problem.

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

In FIG. 1, I 1.12. ..., Im is a section detection unit that detects a voice section using m voice power thresholds.

2 are m section detection units 11.12. . . . is a section selection unit that selects only necessary ones from m voice sections detected at 1 m.

Section detection unit 1+, 12. ...-, 1m, the input audio power is the respective threshold value L1, L2 .・-, Lm
The first point exceeding 31. S2. −． Starting edge detection units 111, 112. that detect S+n. -=, l1m s and the last point El, E2°·-I Em. -・-, consisting of 12m.

The end detection units 12+, 12z, -, and 12m detect the start end, and if the input audio power decreases and exceeds the threshold value, but rises again within a predetermined time and exceeds the threshold value, the end detection units 12+, 12z, -, and 12m do not detect the end point, but detect 1 of the audio section. ! Processing is also performed to consider it as a continuation.

The section selection section 2 includes section detection sections 1 +, 12. −・−,
Detected start ends 31, 32 of 1 m. ...-9S... and terminal E+.

E2. ...-, perform the following inspection on Em.

1st -sj l<Qt lEi -Ej l<Q However, Ql and Q2 are predetermined thresholds, and i
, j is i=1.2. ..., m −1i J =
This is performed for all combinations of 2+3, . . . -, m (i<j).

When the above two equations both hold true, the speech interval (Sj, E
j) is excluded from the candidate sections, and only the remaining speech sections are output as candidate sections.

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.

lsi −5j I+lEi −Ej l<Q) In either case, if both the starting end position difference and the ending end position difference are smaller than the same value, one is excluded.

[Operation] With the above configuration, it is possible to reduce the amount of processing by excluding almost the same speech sections while maintaining the advantage of using a plurality of threshold values.

[Example] The present invention will be described in more detail below with reference to Examples shown in FIGS. 2 and 3.

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

In FIG. 2, 1 +, 12. -11m is a section detection section, 2 is a section selection section, and these are the same objects as in FIG.

3 is a power calculation unit which calculates the power of the input voice, and each section detection unit 1 +, 12. ...-, input to 1m.

4 is a microphone for audio input, 5 is an audio amplifier, and the amplified output is sent to the power calculation unit 3, sl, 62 . f3
Input to L'-+6n bandpass filter.

6+, 62, 63. ---, fHl is a bandpass filter (B
PF), which is a filter that passes through each frequency band obtained by dividing the audio band into n pieces.

7 +, 72.7 L·-170 are rectifiers and smoothers, and the respective bandpass filters 6 +, 62.63.・
..., rectifies and smoothes the output from 6n, and outputs an envelope output.

8 is a word speech pattern generation unit, which includes a rectifier/smoother 7;
+, 72, 73. - By the output of -17n, a word sound pattern as shown in FIG. 5 is generated for the section output by the section selection section 2.

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, turn the switch 11 upward, input the word sound from the microphone 4, and generate the word sound pattern by the output of the section detection section, for example 12, which has a standard power threshold. This is stored in the registered pattern storage section 9.

In the case of voice recognition, the switch 11 is pushed down and the voice recognized from the microphone 4 is sent to each section detection section 1+.

12, . . . , 1+m is used to detect sections, and the section selection unit 2 selects only the necessary sections, generates a word speech pattern for this, compares 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 the same figure (a), the three thresholds Ll, L2゜L3 are used to extract voice sections ■, ■, and ■, but the differences between the beginning and end of the sections ■ and ■ are both small. Therefore, as shown by the broken line, section ■ is excluded.

In the same figure (in the case of input B of bl, voice sections ■, ■, and ■ were detected using the three thresholds Ll, L2 and L3, but since the starting and ending differences between section ■ and section ■ are both small, the dashed line As shown, the section ■ indicates that it has been excluded.

[Effects of the Invention] As explained above, according to the present invention, while maintaining the advantage of using multiple thresholds, almost the same speech sections are excluded,
The processing speed can be improved by reducing the amount of processing, 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 present 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 sounds. , 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, I 1.1 z, -, 1m is the section detection unit, 111.11
2. . . , 11m is a starting end detection unit, 121, 122 .・
, 12m is a termination detection section, 2 is a section selection section, 3 is a power calculation section, 4 is a microphone, 5 is an amplifier, 61.62, 63. ---. 611 is a bandpass filter, 7
I, 72.7 ], ..., 7o is a rectifier/smoother, 8
9 indicates a word speech pattern generation unit, 9 indicates a registered pattern storage unit, 10 indicates a collation unit, and 11 indicates a switch.

Claims (1)

[Claims] Speech sections are detected from the voice power of the input voice using a preset threshold, voice feature parameters are extracted, and voice recognition is performed by comparing and collating with the voice feature parameters of words registered in advance. The speech recognition device includes a section detection section that detects a speech section exceeding a plurality of preset speech power thresholds, and a section detection section that compares the positional difference between the start and end of each section detected by the section detection section with a preset threshold. , a section selection section that performs a process of excluding one section and outputting the remaining section when both the start end position difference and the end end position difference are within the threshold, and the section selection section selects the section selected by the section selection section. A word speech recognition method characterized in that a speech feature parameter for a speech section is compared and collated with a speech feature parameter of a registered word.