JPS6033599A - Voice 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 Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device capable of inputting arbitrary sentences by recognizing monosyllabic speech.

Conventional configuration and its problems If voice, which is the most natural means of generating information for humans, could be used as an input means for a human-machine system, the effect would be very large.

Conventionally, speech recognition devices based on a specific speaker registration method have been put into practical use. That is, a speaker who intends to use a recognition device converts all the words to be recognized into a series of feature vectors using his/her own voice and registers them as standard patterns in a word dictionary, and then uses the voice uttered during recognition. is similarly converted into a series of feature vectors, which word in the word dictionary is closest is calculated according to a predetermined rule, and the most similar word is taken as the recognition result.

However, this method is good when the number of recognized words is small, but as the number of words increases to hundreds or thousands of words, the following three problems become impossible to ignore.

(1) The burden on the speaker during registration increases significantly.

(2) The time required to calculate the similarity or distance between the voice uttered during recognition and the standard pattern increases significantly, and the response speed of the recognition device becomes slow.

(3) The memory required for the word dictionary becomes very large.

As a method to avoid the above-mentioned drawbacks, there is a method in which the unit of recognition is a consonant+vowel or a monosyllable of a vowel (hereinafter referred to as CV and V, respectively; C means a consonant and ■ means a vowel). That is, monosyllables are registered as a series of feature vectors as a standard pattern, and input speech converted into a series of feature vectors during recognition is converted into a series of monosyllables by matching with the monosyllable standard pattern. It is something to do. In the case of Japanese, there are at most 101 types of monosyllables, and monosyllables correspond to kana characters, so this method converts any Japanese word or communication into a monosyllable string (recognizes ), and the above (1) ~
All problems in (3) will be resolved. but,
Problems in this case include articulatory combination and segmentation. Articulatory coupling is a phenomenon in which when syllables are uttered in succession, each syllable is influenced by the syllables before and after it, and the spectral structure changes depending on the syllables connected before and after it. Segmentation is the process of dividing continuously uttered speech into single syllables, but it is difficult to do this reliably with current technology. In order to solve these two problems, the current practice is to separate each monosyllable into utterances, and some devices are in practical use.

FIG. 1 shows the general configuration of a device that performs monosyllabic speech recognition using pattern matching. 1 is an input terminal for audio signals. 2 is a feature extraction unit that analyzes the input audio signal using a filter bank, FET, LPC, etc., and extracts a series of feature vectors A=a1.2 every few milliseconds. a2...a
i...Convert to aI. 3 is a standard pattern storage unit which converts monosyllabic speech to be recognized in advance into a series of feature vectors by a similar means, and stores standard patterns Rn=bn1bn2...bnj for each syllable.
...bnJn (however, n=1, 2, ...
..., N; N is the part to be stored as the number of standard patterns. 4 is a pattern comparing section, which compares the input pattern A, which is the output of the feature extracting section 2, with each standard pattern Rn stored in the standard pattern storage section 3. The distance D (A, Rn) between the two is calculated. 6 is a determination section,
Based on n=minn [D(A, Rn)], the standard pattern Rn closest to the input pattern is determined. 6 is an output terminal that outputs the determination result as a monosyllable recognition result. Pattern comparison section 4
The pattern comparison in is the so-called DP using dynamic programming.
Matching, linear shift matching, etc. are often used. In addition, by first recognizing the vowel and determining a candidate stage, and then recognizing the consonant part using the standard pattern belonging to that vowel stage, the recognition rate and matching speed are improved. It is.

However, monosyllabic sounds have short durations, such as "shi" and "shi".
Many of them must be distinguished based on subtle differences in consonant parts, such as "ch", making it difficult to achieve a high recognition rate like word sounds.

In order to solve this problem, a method using a word dictionary has been considered. Figure 2 is an example.

In this figure, blocks given the same numbers as in FIG. 1 perform the same operations as in FIG. 1. 7 is a word dictionary, in which the word Wl to be recognized (l=1.2,..., L; L is the number of registered words) is a symbol string Wl=Cl1Cl corresponding to a monosyllable.
2...ClK...ClKl (ClK is the k-th syllable of the word Wl). 8 is a word comparison section, which inputs a monosyllable string T=A1A2...
...Am...AM (M is the number of syllables of the input word)
When , the word Wl'=Cl'=Cl'1C stored in the word dictionary 7 with the number of syllables equal to the number of syllables of the input word
l'2... Distance D (Am
, Cl′m) for each l′ DW(T, Wl′)=
Calculate ΣMm=1D(Am, Cl'm). 9 is a determination unit, which determines l' such that l'=minl'(T(S, Wl')),
Wl' is determined to be a recognized word. 10 is an output terminal that outputs the recognized word.

As described above, the recognition rate can be improved by using knowledge of the word dictionary. Furthermore, when considering input to a word processor, the word dictionary can be used in common as a dictionary for performing kana-kanji conversion, and there is no need to prepare a word dictionary specifically for speech recognition.

However, the number of words in an instant dictionary is usually over 30,000.
The amount of calculation in the word comparison section 8 can no longer be ignored.

OBJECTS OF THE INVENTION The present invention relates to a monosyllabic speech recognition device that uses a word dictionary to improve the recognition rate of monosyllables. More specifically, the present invention is characterized by improving the speed of matching with the word dictionary. The present invention relates to a speech recognition device.

Structure of the Invention The present invention comprises means for converting an input speech signal into a series of feature vectors, means for dividing the input speech signal into syllables, and means for recognizing a subsequent vowel of each syllable from the series of feature vectors. means for obtaining, as a symbol string, a syllable string of a word or phrase having the same subsequent vowel string as said subsequent vowel string; means for matching said symbol string with a syllable string obtained from said input speech signal; and a result of said matching. and determining means for determining the word or phrase closest to the input voice signal as a recognition result corresponding to the input voice.

The basic idea of the present invention will be explained below.

In monosyllabic speech recognition, vowel recognition is almost certain. Therefore, the input monosyllable CV or ■1 (C is a consonant, ■ is a vowel) subsequent vowel series is ■1V2...
When it is a VM, as a word in the word dictionary to be checked,
The sequence of vowels following the monosyllable that makes up the word is ■1V2
...It is only necessary to select words that are VM. For example, the following vowel in a monosyllable string is |o||o||
If a||a|, "oosaka", "toyonaka", etc. would be selected as the words to be matched.

In this way, for example, in the case of a four-syllable word, if the probabilities of vowel occurrence are equal, the probability of a particular vowel string occurring is (1/5)4=1/625, which means that a four-syllable word is If there are 10,000 words, 4 that corresponds to a certain vowel sequence
There are 16 syllabic words, and the number of words that actually need to be compared and calculated is drastically reduced.

Even if we consider the vowel of the second candidate for some margin (
2/5) 4≒1/39, and similarly a word with 4 syllables becomes 1
If there are 10,000 words, there are 266 four-syllable words that need to be compared.
This will result in a significant decrease. Furthermore, if consonants and consonants are processed in the same way as the vowels described above, the number of comparison calculations can be further reduced.

Since the recognition of these vowels, consonants, consonants, etc. is performed almost completely, not only the amount of calculation is reduced, but also the recognition rate itself is improved.

DESCRIPTION OF THE EMBODIMENT FIG. 3 is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention. Reference numeral 11 denotes an input terminal for audio signals, into which words are input as a chain of monosyllables.

Reference numeral 12 denotes a feature extraction unit similar to that described in the conventional example, which converts input speech into a series of feature vectors as described above. 13 is a power calculation unit, which calculates the output vector sequence of the feature extraction unit 12 by a1a2...ai...
・When aI, the power Pi of the i-th frame is, for example, if ai=(ai1, ai2..., aiμ), then Pi=■ai12+ai22+...+ai
It can be regarded as μ2. Reference numeral 14 denotes a syllable section detection section, which divides the input speech into syllables and detects the start frame and end frame of each syllable from the output of the power calculation section 13. FIG. 4 is an example of this, where the time point when the power exceeds the threshold value 29 is taken as the start frame of a syllable, the time point when the power becomes less than the threshold value 29 is taken as the end frame of the syllable, and the section where the power exceeds the threshold value 29 is taken as the syllable existence section. Also, the section below the threshold value 29 is a constant value tc
If there are more than 1, the interval is considered to be a consonant. The figure is “s
This shows what happens when you say "apporo", and Q
means a consonant. Reference numeral 15 denotes a syllable number counting section, which counts the number of syllables (therefore, the number of moras) by considering a consonant as one syllable. 16 is a vowel standard pattern storage unit, in which the vowel |a
|, |i|, |v|, |e|, |o|, and pixel |N|
Standard patterns are registered in advance.

17 is a vowel frame detection unit, which includes a syllable interval detection unit 1
The frame position corresponding to the vowel is detected from the start and end frames of each syllable detected in step 4 and the series of feature vectors extracted by the feature extraction unit 12.

Since the vowel part is stationary, the frame that should be claimed is the i-th frame.
Frame i can be detected as the frame i in which the sum of the variances of each component of the feature vector from the -r frame to the i+r frame (r is a constant) is minimal. That is, the input feature vector of the i-th frame is ai=(ai1, ai2,...
..., aij, ......, aiμ), then mij=1/2r+1Σi+rk=i-r(anj-m
ij) In 2, from the last frame of each monosyllable, v
As i is increased, the frame where vi becomes minimum can be taken as the frame at the center of the vowel stationary part. A buffer memory 18 stores a series of feature vectors extracted by the feature extraction section 12 for each monosyllable from the monosyllable start frame detected by the speech section detection section 14 to the end frame. Reference numeral 19 denotes a vowel pattern comparing section which reads the feature vector corresponding to the frame detected by the vowel frame detecting section 17 from the buffer memory 18, compares it with each vowel standard pattern in the vowel standard pattern storage section 16, and calculates the distance to each one. Calculate. For example, ai=(ai
1, ai2, .
2,...,vνμ) (where ν=1,2,・
..., ) can be set as diν=■ΣμK=1(aik−vνk)2. 20 is a vowel determination unit, ν=min[d
iν] ν is the vowel corresponding to vν as the vowel recognition result.

Reference numeral 21 denotes a vowel/consonant determination result storage section, which stores the vowels determined by the vowel determining section 20 and the consonants detected by the syllable interval detecting section 14 in the order of their occurrence.

Reference numeral 22 denotes a monosyllabic standard pattern storage unit, which stores standard patterns corresponding to each monosyllable that have been converted into a series of feature vectors. Reference numeral 23 denotes a monosyllabic pattern comparing section, which compares the input pattern stored in the buffer memory 18 with the monosyllabic standard pattern stored in the monosyllabic standard pattern storage section 22, and compares each input pattern with the monosyllabic standard pattern stored in the monosyllabic standard pattern storage section 22. It calculates the distance to a monosyllabic standard pattern. At this time, the monosyllable standard pattern to be matched is limited to monosyllables having the vowel determined by the vowel determining section 20 as a subsequent vowel. Furthermore, the range to be compared for each single syllable is from the start frame of that single syllable to the constant part of the vowel. This is exactly the part that contains consonant information. As a method of comparison and matching, well-known methods such as linear shift matching and DP matching can be used. If DP matching is used, the result will be as follows. The nth monosyllabic standard pattern is Rn=bn1bn
2...bni...bnJn, monosyllabic input pattern A=a1a2...ai...
When aI (where I and Jn are the input pattern and the center frame of the vowel stationary part of the standard pattern, respectively) and dn (i, j) are the distances between the vectors of ai and bnj, the recurrence formula becomes g (1, 1) =2dn(1,1), the distance D(A,Rn) between A and Rn becomes D(A,Rn)=g(I,J). Here, dn(i,j) is ai=(ai1,aj
2,...aiμ)bnj=(bnj1, bni
2. ...bnjμ), then dn(i,j)
=Σμk=1|aik−bnjk| is generally set. Furthermore, various forms of the above recurrence formula have been proposed, and only one example thereof is shown here. A distance storage section 24 stores the distance calculated by the monosyllabic pattern comparison section 23. Monosyllable string A1A2...
Am... When a word consisting of AM is input, the distance storage unit 24 sets D(Am, Rn) to 1≦m≦M, R
n ■ Memorize everything about SAm. However, let SAm be a set of monosyllabic standard patterns having the same following vowel as Am.

Reference numeral 25 is a word dictionary in which words to be recognized are stored in a form expressed as a string of syllable symbols. 26 is an inter-word distance calculation unit which calculates the distance between a word input as a monosyllable string and a word in the word dictionary 25 from the distance stored in the distance storage unit 24.

The words to be compared and verified against the word dictionary 25 are the values in the syllable number counting section 15, that is, the number of syllables of the input word and the subsequent vowels (including pellicles and consonants) indicated in the vowel/consonant determination result storage section 21. Restricted to words with the same subsequent vowel sequence as the sequence. Now, let this limited set of words be Sw, and W
l■SW word Wl is Cl1Cl2...Cl
If it is made up of a syllable string m...ClM, the inter-monosyllabic distance D (Am, Clm) between the monosyllables Am and Clm is stored in the distance storage unit 24 according to the above explanation. , input word T=A1A2...Am...
...AM and the word WlCl1Cl2 in the word dictionary...
...Clm...Distance to ClM DW (T, W
l) is DW(T, Wl)=ΣMm=1D(Am, Clm
). Reference numeral 27 denotes a word determination unit, which determines Wl to be a recognized word based on l=min[DW(T,Wl)]Wl■SW. 28 is an output terminal for the recognition result.

Although the present embodiment has been described as recognizing on a word-by-word basis, it is of course possible to recognize on a phrase-by-phrase basis. In that case, it is possible to consider a noun with attached words, verbs, adjectives, conjugated forms of adjectives, etc. as the above-mentioned words and register them in the word dictionary. Since the memory capacity of the dictionary increases significantly, nouns are registered in the word dictionary without stems or adjuncts, and when the word distance calculation unit 26 performs comparison and matching,
It is also possible to create various clauses using rules.

In particular, when the device of the present invention is used as an input to a word processor with a kana-kanji conversion function, the word dictionary for kana-kanji conversion can be shared, and the function for creating the adjunct words is already provided, which is very convenient.

Further, in this embodiment, a consonant is detected from the silent interval length, but a consonant can also be input by saying "tsu". In this case, in the word dictionary, the consonant should be
All you have to do is replace it with ``tsu'', and it is easy to distinguish whether it is ``tsu'' or a consonant as a matter of language processing.

Further, although the present invention has been described with respect to the case where the utterance is divided into monosyllables, it is sufficient if the utterance can be divided into single syllables, and if this division can be performed even when uttered continuously, the present invention The principle can be applied as is.

Effects of the Invention According to the present invention, not only single syllables are recognized, but whole words are recognized, and by limiting the words to be compared and matched by vowel strings, the recognition rate and matching speed are significantly improved. This is a significant improvement.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a conventional monosyllabic speech recognition device.
FIG. 2 is a block diagram showing an improved example of the conventional example, FIG. 3 is a block diagram showing a speech recognition device in an embodiment of the present invention, and FIG. 4 explains a part of the moving object of the device of the present invention. FIG. 11...Audio signal input terminal, 12...
Feature extraction unit, 13...Power calculation unit, 14...
. . . Vocal section detection unit, 15 . . . Syllable number counting unit, 16 . . . Vowel standard pattern storage unit, 17.
. . . Vowel frame detection section, 18 . . . Buffer memory, 19 . . . Vowel pattern comparison section, 2
0...Vowel/sound determination section, 21...Vowel/sound judgment section
Consonant determination result storage unit, 22... Monosyllabic standard pattern storage unit, 23... Monosyllabic pattern comparison unit,
24... Distance storage unit, 25... Word dictionary, 26... Word distance calculation unit, 27...
. . . Word determination unit, 28 . . . Recognition result output terminal.

Claims (1)

[Claims] means for converting an input audio signal into a sequence of feature vectors;
means for dividing an input speech signal into syllables; and means for recognizing a subsequent vowel of each syllable from the series of feature vectors;
means for obtaining a syllable string of a word or phrase having the same subsequent vowel string as the subsequent vowel string as a symbol string; and means for matching the symbol string with a syllable string obtained from the input speech signal. . A speech recognition device characterized by comprising: determining means for determining, as a result of this matching, the word or phrase closest to the input speech signal as the recognition result corresponding to the input speech.