JPS6078495A - Big vocabulary word recognition equipment - 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] Industrial Field of Application The present invention registers in advance CV syllables or V1CV2i' pronunciations, which are the minimum vocal unit of Japan 1.1),
Recognize the ll'L word uttered by the registered speaker in CV syllable units1
This invention relates to a dog language word speech recognition device that performs instant matching to recognize dog language words.

Structure of the conventional example and its problemsConventional word sounds): -In a simple recognition device, a simple H5 dictionary is registered in units of '15 Sou vector series of input speech. However, in such a 174-format system, the memory required to recognize words with large actual words is recognized. Since the storage space is enormous and the process still takes 11.7 hours, there are two practical problems in applying it to a dog language word speech recognition device. In addition, in order to perform word speech recognition of large words, nine phoneme patterns (for example, 5 vowels /A//i// u//E/1) are registered in advance for each frame of the feature vector of the input speech.
0/, child VS////CZ4//p//l///,/
b//d//q/, /m//n//r/, etc.) and merge the results of the phoneme identification series (for example, continuous phonemes are represented by one phoneme, discontinuous After processing (such as truncating phonemes) to obtain the recognition result of the phoneme sequence, the input speech was compared with a word dictionary to obtain the recognition result. Although such methods have been practical in terms of memory capacity and processing time, the identification rate of phoneme identification has been affected mainly by the deformation of consonants caused by articulatory combinations caused by phoneme combinations and by segmentation. This was due to the lack of clarity (the phoneme intervals were unclear), which caused the recognition rate to drop significantly.

OBJECTS OF THE INVENTION It is an object of the present invention to solve the above-mentioned conventional problems and to provide a speech recognition device for large word drafts that can perform real-time processing and have a high recognition rate.

Composition of the Invention The present invention 8J'J converts the input 1 "fJj・ into a 14!i series of =I Ge characteristic vector JXt,l and power (pt□), and the vowel identification results for each frame. , tυ sound constant' center (□■1°, ■2....i■N and no 1°? section ([:1SS1, 1sE1] [1SS2,
1sE2]... [155M, iSEM] l ; From the beginning of the word to the final IJ, R stationary part center +, VNl-
1: Between 'Jfj l/I' in the range? and end of no-1 section (isEl.

1<2<Ml to lil, 7:7 At the center of the stationary region 1, the previously memorized CV mark 11. A plurality of Cv syllables are output by adding likelihoods in order of priority from the result of matching the Cv syllable. - The range from the center of the constant constant part of the 1'' sound that has been in contact for 1 night, not including the interval, to the center of the constant constant part of 1ζ month + iv8~. VE: (E=S-1-i in January
i, 'f5J19.' by adding likelihood in order of priority from the result of matching the button with the V1CV2 standard button. Individual C■
Outputs two syllables. The above process is performed for each vowel i-regular center, and the final vowel 'i vN' is converted to CVi"f
The series of syllables hi) and Cv2 are combined to obtain the final recognition result.In this case, the likelihood is set to 1 and priority is given to the combinations of the multiple C■ syllables and Cv2 syllables that are output. Adding a generalized likelihood to maximal proverb 1 in order of ranking, multiple (C
A word (consisting of the combination of V syllables and Cv2 syllables) is output. Then, for combined words with a high ranking, for example, after limiting the categories of words in the word dictionary of large words from the series of B-sounds, matching with the word dictionary is performed. By printing the matching result and the weighted sum of the summed likelihoods of the combined words, and using the word dictionary with the highest weighted sum as the recognition result, real-time processing is possible with a simple device configuration. The purpose is to put into practical use a dog language speech recognition device that can obtain even higher recognition rates.

DESCRIPTION OF EMBODIMENTS Examples of the present invention will be described below. 1st
The figure shows an embodiment of the fHj five-side recognition device according to the present invention.

In the figure, 100 is a sound processing section, and this sound processing section i
oo is a long value due to the power sequence of the A/D converter 1, the power sequence converting means 2, the feature sequence converting means 3, and the input voice.
s'? I' detects j:,, r,, j3j interval 4a
Voice section detection J-124, voice section 4apC that detects
It is shorter than the power series PC; tl! ! , 1 silence interval detection means 5 which detects high and no 14 intervals 5a, peak 7b of human voice, does not detect force/, , peak power detection means 6a (!: 1 for each vector of the feature vector series ; month ° S - I perform the identification:) sound - present from the identification means 6b, f) month g' ji in frames VC before and after the peak power
From the same-fU sound center of iQ-specific results, 'B) 1η constant constant center 6 C? J-Detect/U sound center detection means 6, human voice! CV syllable 7a or v in the form of a 15-way vector
1cv27'TI quality) 7b's funeral position is recorded in the standard slam memory part T, human power j'1 voice feature vector is recorded in a certain interval '1 stone' is recorded in '4'5r:' Series 1; if part 8. Add 1. to the feature vector sequence of the input voice. - From the beginning of C1 word 4a or the end of silent section ljl 5a! At IU sound center 6 c-1 of l'l liquid, CV syllable standard bang and bang matching is performed, and from IR1 near center 6c to IJ sound center 6c, v1C ■ 21° −
It consists of a bang matching means 9 which outputs a plurality of CV or iJ:CV2 syllables in a priority order based on the clausal standard bang and bang matching based on the likelihood based on the line space distance.

The word matching unit 200 performs 1. CV syllable sequence storage unit 10° with likelihood that sequentially stores a plurality of CV syllables with likelihood for each vowel center 6C output by the bang matching means 9; CV with likelihood;
Candidate word storage unit 11 that totals the likelihoods for each vowel combination of the CV syllable series in the syllable series storage unit, synthesizes words according to the priority order, and stores limited candidate words from the total likelihood; The dog word millet word dictionary 14 classifies and stores the words ``粱'' according to the number of syllables, vowel arrangement, etc., and the large word leaf word dictionary 14 detects the number of syllables and vowel arrangement of candidate words in the candidate word storage section 11. A word matching means 12 performs matching after selecting a category, and the score based on the matching result of the word matching means 12 and the likelihood of the candidate word in the candidate word storage section 11 are weighted and added, and the word with the highest score is recognized. It consists of a weight adding means 13 for adding results.

Reference numerals 16 refer to a voice section detection means 4, a silent section detection means 6, a vowel center detection means 6, a feature series storage section 8, a standard bang storage section 7, and a bang matching means 9 of the acoustic processing section 100.
It comprehensively controls and outputs multiple CV syllables with added likelihood for each vowel center 6c of the input voice, -i/;
Candidate word storage section 11 of word matching section, word matching means 1
2. Weight addition means 13 and large word unit Wt,! i':
r)! By comprehensively controlling i' 14, candidate words are created from combinations of CV syllables for a plurality of CV syllable sequences 1o, which are outputted by the acoustic processing unit and added with likelihoods for each vowel. It is a comprehensive control means that performs matching with a leaf word dictionary l and performs word identification from the matching result and the sum of likelihood weights of candidate words.

The motion Takehara Jll of this example will be explained below.

The beginning 4a and the ending 4a of the voice section are detected by the voice section detecting means 4 by -C.

The vowel center detecting means 6 detects the beginning of the word 4a for the input speech 16.
If the power sequence that is set after a certain threshold by the peak power detection means 6a continues for a certain frame length or more, and furthermore, the vowels identified by the vowel identification means 6b continue for a certain frame length or more with the same insertion. vowel center (〔,■
1. i■2. ...,,V5] is detected as 16C. Each time a vowel center 6C is detected, the silent interval detecting means 6 detects whether a silent interval exists between it and the previous vowel center, and the silent interval start, end [
1SS1°1sE1]5a.

From the beginning of the word or the end of the silent section to the vowel center 1. v,
: For the section with 100 cm, matching is performed using the CV41 semi-pattern 7a and the matching means 9, and the CV
Add likelihood to syllables and prioritize them? (The lift output 9a is used, and for adjacent vowel centers where there is no silent interval, matching is performed using the ■1C■2 mark slam 7b and the bang matching means 9V, and CV2 syllables are similarly added to the likelihood. An example of the processing result of the audio processing section for the input voice /Amagasaki (Amagasaki)/ is shown in FIG. 3-J'.

Next, the principle of operation of the word matching section 200 will be explained below using FIGS. 4 and 5.

Assuming that the processing result of the acoustic processing unit 100 is as shown in FIG.
・Determined as shown. Once a candidate word is determined, the condition of the candidate word (for example, the score (SWD, heat and l) is determined by the single notating means 12, and then the score (SW,...) and 1 are written.
] 1 of the money side of the belly (S□;'s offer addition (swD, +ks,)
] The most effective single ii, 't 1'i'' 4.1! Let the candidate be the recognized word WD.
Are you going to get caught? +Ii Single Wn Ka, W 1-/A/MA/
G A/b A/H1/.

W2=/A/MA/ZA/SA/Hi/, W3=/A/
MA/GA/SA/Ki/・・・・
・The sum of the respective likelihoods is 811”'2153...
It is shown that it will become a mountain, and from this, Otsume', 1
Amagasaki (/ A 7M
A / G A / S A / K i / ) and 1
Inexpensive (/A/TA/TA/KA/i
/ ) indicates that it has been selected as a word dictionary candidate, and the score s'wD, with respect to the candidate word.

3 and the sum of the likelihood (S,) indicates that the maximum value of the weighted addition is Amagasaki.

Effects of the Invention The present invention performs matching for each vowel stationary part center of input speech with a CV syllable or v1Cv2 syllable standard bang registered in advance, and combines the recognition results into multiple CVs in a priority order with a likelihood based on the matching distance. Or output as CV2 syllable, and when the center of the final vowel stationary part is detected,
A plurality of candidate words based on combinations of recognition results for each vowel center are output in priority order based on the sum of their respective likelihoods. Then, the categories of the Inu-Kankan word dictionary are limited based on the number of syllables and vowel arrangement of the candidate word, and the maximum value of the score of the matching result with the candidate word and the weighted addition of the likelihood of the candidate word is used as the word recognition result. By doing so, it becomes possible to realize real-time processing with a simple device configuration and to put into practical use a dog language word speech recognition device that can obtain a higher recognition rate.

[Brief explanation of drawings]

Fig. 1 is a schematic diagram of a large word speech recognition device according to an embodiment of the present invention; )” is uttered, and Figure 3 is the final output of the acoustic processing unit, which is the multiple likelihood + J” for each vowel center.
■i? FIG. 4 is a diagram showing the clause recognition result series, which is based on the embodiment shown in FIG.
11 From the weighted sum of the matching result and the likelihood of the candidate word, single 1. l
Fig. 6 is a diagram showing a method for identifying F 1fjR. The candidate word is “Amagasaki” 1-1! FIG. 3 is a diagram showing word recognition results for "gai". F2... Electrical series conversion means, 3... Characteristic series conversion means, 4... Voice section detection means, 5...
...Silent section detection means, 6...I;month'? ~Center detection means, 7...Standard bang memory, memory section, 8.
...Feature series storage unit, 9...Bang matching means, 10...Likelihood CV syllable series record 1
．． (Yu section, 11...Candidate ili word storage section, 12.
...Single; t1 matching means, 13...
・Weight addition means, 14...Large word single edict 1'l・(11
:, 15... General control means, 10Q-... Sound processing) 11S, 200... Single, ? 1% polymerization part. Agent's name Jfl'l! I - Toshio Nakao and 1 other person (procedural amendment ceremony) February 2, 1980 1 Display of 1 item Showa 58-1 '11 License No. 186481
Month 2 invention name for human language 11'! , ,T72.71. No I/Recognition device 3
To make corrections, 1 point, H+, 4LI allowed L1
+11t Personnel No: ii Person 1υi 111 Kadoma City Dai'j, Barpage 1006 Address name I゛1. (582) Matsudo Dentomi; Sangyo Co., Ltd.? I-E-1, 1111 Lower Toshihiko 4 agent 〒,'+71 Nozu ``人l:I＜ IljBar Teizanjin'j・Bar↓°
+H10o6it ground 1・≧1・electronic: so:; industry +4
, I have corrected the dimensions as shown in Figure 3 of the drawing in Ceremony 11.・Work Ri

Claims (1)

[Claims] Feature sequence converting means converts the input audio into a sequence of feature vectors (Xtil); power value sequence converting means converts the input audio into a sequence of power values 1Pti1;
, l, and a vowel stationary part center detecting means for detecting the center of the seasonal/sound stationary part from the result of the level determination of the characteristic vector (Xt yo) and the phrasal sound identification result of the feature vector (Xt yo). Between adjacent W″ sound constant center centers, Cv2(
C: consonant, ■: vowel) syllable or v1cv2 syllable mark 'lfA bang and bang matching is performed, and for each center of the vowel stationary part, the Rg nfk result of the Cv2 syllable is calculated based on the distance based on the bang matching result. The recognition result as a word is the C■2 syllable recognition result for each vowel stationary part center, and the likelihoods are summed for all combinations, and the priority order is based on the summed likelihood. Then, when outputting a plurality of words corresponding to the combination and matching the plurality of words with the word dictionary, the total likelihood is added with a weight of τ1 to the Masochinoda result to obtain the recognition result. -, T, 7C1′I1. A speech recognition device for Chinese and Chinese words.