JPH0566596B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition device, and particularly to a speech recognition device that recognizes speech produced by consecutively uttering syllables such as words or phrases.

(Constitution of Conventional Example and Its Problems) If voice, which is the most natural means of generating information for humans, can be used as an input means for a human-machine system, the effect would be enormous.

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 speakers during registration increases significantly.

(2) The time required to calculate the similarity or distance between the voice uttered and the standard pattern during recognition 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 expressed as CV and V, respectively; C means a consonant and V means a vowel). That is, monosyllables are registered as a series of feature vectors as standard patterns, and input speech converted into a series of feature vectors during recognition is converted into a series of monosyllables by matching with the standard pattern of monosyllables. It is something to do. In Japanese, the number of monosyllables is at most
There are 101 types, and monosyllables correspond to kana characters, so this method converts (recognizes) any Japanese word or sentence into a monosyllable string.
This will solve all of the problems (1) to (3) above. However, 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 monosyllable units, but a definitive method for doing this reliably has not yet been found. In order to solve these two problems, the current practice is to separate each single syllable and pronounce it, and some devices are in practical use.

However, uttering monosyllables discretely is unnatural and puts stress on the speaker.

(Objective of the Invention) The present invention is small, inexpensive, and easy to register standard patterns even when there are a large number of words or phrases to be recognized, and it is possible to improve recognition accuracy and processing speed. The purpose of this research is to realize a speech recognition device that is easy to use.

(Structure of the Invention) The present invention registers syllables such as V, CV, VV, VCV in advance, and inputs words or phrases by continuously speaking them as a string of these syllables using a word dictionary. It recognizes the input signal, and its configuration includes feature extraction means for converting an input audio signal into a series A of feature patterns, a steady point extraction means for extracting stationary points of the input signal, and each extracted stationary point. is regarded as a vowel, the vowel is identified, and the input vowel string
and a vowel identification means that matches the input vowel string X obtained by the vowel identification means with the vowel string Y ⁿ (n=1, 2,..., N) of the word or phrase to be recognized. a vowel string ^{identification} means for identifying a standard vowel string Y ^no closest to the vowel string X, and a vowel string identifying means for identifying the vowel string vowel string correspondence determining means for determining the correspondence of vowels in the column Y ^no ; and vowel string identification for a partial sequence of the input signal corresponding to a partial interval of the input vowel string X determined based on the correspondence. Of the identified vowel string Y ^no obtained by the means, the specific part Y no corresponding to the partial interval Y ^no = (y ^no _j1 ,...,
y ^no _j1 Cy ^no _j1+1 ^defined by the vowel string y no _j2 ) (j ₁ < j ₂ )
，
By performing matching with standard patterns corresponding to each syllable such as y ^no _j1+1 Cy ^no _j1+2 ,..., y ^no _j2-1 Cy ^no _j2 (C: consonant), etc., the specific partial feature pattern series is created. This is a speech recognition device that is equipped with a determination means that identifies a corresponding syllable string, determines a word or phrase based on the identified syllable string, and outputs the result as a recognition result. , even if there are a large number of them, it is small, inexpensive, and easy to register standard patterns, and can improve recognition accuracy and processing speed.

(Description of Examples) Hereinafter, the word "word" will also represent the word "bunsetsu."

Now, FIG. 1 is a functional block diagram showing a first embodiment of the present invention. 1 is an audio signal input terminal. 2 is a feature extraction unit that uses, for example, a 20-channel filter bank and processes one frame for 10 msec.
Then, 20 values (feature patterns) are obtained every 10 msec in the output. That is, the input audio signal is a series of feature patterns A = (a ₁ , a ₂ , ..., a, ...,
a _L ). Here, a=(a ₁ , a ₂ ,
..., a ₂₀ ) is the feature pattern obtained in the th frame, and L is the number of frames of input audio. 3 is a power calculation unit, and if the power of the th frame is P, then P = √ ² ₁ + ² ₂ +...+ for each frame.
a ² ₂₀ is calculated. Reference numeral 4 denotes a voice section detecting section, which detects the beginning and end of the input voice signal from the change pattern of P. In other words, a threshold value is set to determine whether there is a sound or not, and when the interval above this threshold continues for a predetermined period of time (e.g. 30 msec), the time when this threshold is exceeded is considered to be the start point of audio, and the time when the sound is below this threshold is A certain period of time (for example,
300 msec) or more, it is possible to use a method such as setting the point in time when the value falls below this threshold as the end of the audio. Reference numeral 5 is a vowel standard pattern storage unit which stores characteristic patterns of the constant portion of each vowel in advance. A buffer memory 6 temporarily stores the characteristic pattern of the input audio signal from the start to the end detected by the audio section detection section 4. Reference numeral 7 denotes a stationary point detection section, which reads out the contents of the buffer memory 6 and detects a stationary point. The stationary point can be detected, for example, by calculating the variance of feature patterns of several frames (for example, 5 frames) before and after each frame, and detecting the frame with the minimum variance. That is, if this variance in the th frame is σ ² , then the sequence of characteristic patterns of the input signal A = (a ₁ , a ₂ , ..., a, ..., a _L ), a
= (a ₁ , a ₂ ,…, a ₂₀ ), σ ² = ₂₀ 〓 ^k=1 _1+N 〓 ^q=lN (a _qk − _k ) ² (1) _k = 1/2N+1 _1+N 〓 ^q=lN a _kq (2) N=5 (3) Given as follows. A stationary point (frame) storage section 8 stores a stationary point (frame) sequence detected by the stationary point detection section 7. Reference numeral 9 denotes a vowel pattern comparing section, which regards each of the stationary point (frame) strings stored in the stationary point (frame) storage section 8 as a vowel center frame and compares the characteristics of the stationary points (frames). The distance (or similarity, hereinafter, the degree of similarity between the pattern and the standard pattern (characteristic pattern) corresponding to each vowel in the vowel standard pattern storage section 5,
“Similarity” is represented by “distance”. In other words, "small distance" means "high similarity"). Reference numeral 10 denotes a vowel identification section, which takes the vowel that gives the minimum value among the outputs of the vowel pattern comparison section 9 as the vowel identification result of the stationary point (frame). 11 is a vowel/consonant determination result storage unit,
The vowel string (input vowel string) obtained by the vowel identifying section 10 and the portion of the silent section detected by the speech section detecting section 4 that is determined to be a consonant are stored. Here, detection of a consonant is determined based on the time length of the silent section based on the above definition. For example, this section may be set as a consonant for 100 msec to 250 msec. Furthermore, the vowel/consonant determination result storage section 11 also stores vowel strings identified by the vowel string identification section 15, which will be described later. Reference numeral 12 denotes a standard vowel string storage unit which stores the vowel strings (standard vowel strings) of words to be recognized, that is, words stored in the word dictionary unit 22 (described later) (hereinafter, the vowel strings will also include consonants). are stored to avoid duplication. Reference numeral 13 denotes an inter-vowel distance storage section, in which an inter-vowel distance determined in advance using the vowel standard pattern stored in the vowel standard pattern storage section 5 is stored. 14
is a vowel string comparison section that reads out the input vowel string stored in the vowel/consonant determination result storage section 11 and matches it with each standard vowel string stored in the standard vowel string storage section 12. This is what we do. Here, matching can be performed by well-known DP matching. That is, the nth standard vowel string is Y ⁿ = (y ⁿ ₁ , y ⁿ ₂ , ..., y ⁿ _j , ..., y ⁿ _Jo ), and the input vowel string is X = (x ₁ , x ₂ , ..., x _i ,…, x _I )
(J ⁿ , I are the number of vowels in the standard vowel string and the number of stationary points in the input vowel string, respectively), and d ⁿ (i, j) is the i-th input vowel x _i and the j-th standard vowel y ⁿ _j . When the distance is g (i, j) = ming (i-3, j-1) + 3・d ⁿ (
i, j) g (i-2, j-1) + 2d ⁿ (i, j) g (i-1, j-1) + d ⁿ (i, j) g (i-2, j-2) + 2d ⁿ (i, j) g (i-1, j-2) + d ⁿ (i, j) If we solve the recurrence formula as g ⁿ (1, 1) = d ⁿ (1, 1), then X and Y The distance D(X, Y ⁿ ) of ⁿ is D(X, Y ⁿ )=g(I, J ⁿ ) (5). Here, d ⁿ (i, j) is the i-th vowel x i of the input vowel string X and the j-th vowel _y of the n-th standard vowel string Y ⁿ among the contents of the inter-vowel distance storage section 13. It is given by reading the inter-vowel distance with _j . In the manner described above, the distance D (X, Y ⁿ ) between the input vowel string X and the standard vowel string Y ⁿ is determined and output. Further, the vowel string comparison unit 14
is, g(i, j) given by equation (4) is the grid point (i-3, j-1), (i-2, j-1), (i-1,
j-1), (i-2), j-2), (i-1, j-2)
The starting point g ⁿ (1, 1) is sequentially memorized from which grid point the transition is from, and when equation (5) is determined,
The path (hereinafter referred to as the "optimal path") from to the end point g ⁿ (I, J ⁿ ) is output together with the distance D (X, Y ⁿ ) given by equation (5). Note that various forms have been proposed for the recurrence formula (5), and only one example is shown here. As described above, the vowel string comparison unit 14 compares all standard vowel strings Y ⁿ (n=
1, 2,...,N) to output the distance and optimal path. 15 is a vowel string identification section, which identifies the standard vowel string Y ^no that gives the minimum value among the outputs of the vowel string comparison section 14, takes it as the identification result vowel string Y ^no , and selects Y ^no and the accompanying optimal path. (identification result optimal path) is output. 16 is a vowel correspondence determination unit which sequentially ascends the identified optimal path from the terminal grid point (I, J ^no );
It is determined which stationary point (i.e., vowel) in the stationary point sequence the grid point (i ^no , j ^no ) on the identification result optimal path corresponds to. By the way, Figure 2 shows "A
``I ZU WA KA MA TSU'', eight stationary points are detected in the stationary point detection unit 7, and the vowel comparison unit 9 detects eight stationary points for each stationary point.
and the vowel identification unit 10 identifies the vowel,
An input vowel string
FIG. 4 is a diagram showing the transition of lattice points when the identification result vowel string Y ^no "A A U" is obtained. The vowel correspondence determination unit 16 traces this transition in reverse. That is, first, lattice point (I, J ^no )=
The starting point of the transition to (8,7) is the grid point (7,5), and the transition to the grid point (7,5) is from the grid point (5,3), and the transition to the grid point (5,3) is from the grid point From (2, 2), each vowel in the identification result vowel string Y ^no and the input vowel string Determine the correspondence with each vowel. That is, in the input vowel string X, the first vowel "A", the second vowel "I", the fifth vowel "U", the seventh vowel "A", and the eighth vowel " Each of "U" is the first vowel "A", the second vowel "I", the third vowel "U", and the fifth vowel of the identified vowel string Y ^no .
The identified vowel string corresponds to the th vowel "A" and the 7th vowel "U", and corresponds to the 3rd vowel "U" and the 4th vowel "I" of the input vowel string There is no vowel in Y ^no (insertion), and there is no vowel in the input vowel string X that corresponds to the sixth vowel "A" in the identified vowel string Y ^no (dropout); There is no vowel in the identification result vowel string Y no corresponding to the sixth vowel "I" in the identification result vowel string Y ^no , and at the same time there is no vowel in the input vowel string X corresponding to the fourth vowel "A" in the identification result vowel string Y ^no . It is determined that there are no vowels (insertion and omission occur simultaneously).
(Here, if "insertion and omission occur simultaneously" and if the vowel determined to be "dropped" in the identified vowel string Y ^no is determined to be "inserted" in the input vowel string A case in which a vowel is incorrectly identified by the vowel identification unit 10 is distinguished by the transition state of the optimal path.In other words, regarding the transition from the lattice point (5, 3) to the lattice point (7, 5), As shown in Figure 2, the case where there is a direct transition from lattice point (5, 3) to lattice point (7, 5) is a case where "insertion and dropout occur simultaneously," and the transition from lattice point (5, 3) to lattice point ( 6,
The case where the transition occurs from 4) to grid point (7, 5) is the case where "misidentification occurs at grid point (6, 4)." ) The correspondence between each vowel in the input vowel string X and the identification result vowel string Y ^no determined as above and the identification result vowel string Y ^no are stored in the vowel/consonant determination result storage unit 11. . Reference numeral 17 is a specific part determination unit which determines the correspondence between each vowel ⁱⁿ the input vowel string ^no is read, and the input vowel string X and the identification result vowel string Y ^no correctly correspond (identified) vowels (i.e., the dropped vowels,
the inserted vowel, the vowel other than the misidentified vowel), for example, in FIG. 2, the interval from the first vowel "A" to the second vowel "I" in the input vowel string , the section from the second vowel "I" to the fifth vowel "U", the section from the fifth vowel to the seventh vowel "A", the section from the seventh vowel "A" to the fifth vowel "A" Each section up to the eighth vowel "U" is determined to be a specific portion. however,
If the initial vowel of a word is incorrect, the section from the beginning of the word to the correctly identified vowel is used as the specific part,
If the word-final vowel is incorrect, the section from the last vowel to the word-final vowel among the correctly identified vowels is set as the specific portion Y ^no . Reference numeral 18 is a syllable standard pattern storage unit which stores a series of characteristic patterns for syllables such as V, CV, VV, VCV, from the beginning of the word to the constant vowel part for V and CV, and the constant part of the preceding vowel for VV and VCV. to the stationary part of the following vowel is uttered in advance by the speaker as a standard pattern and registered. Reference numeral 19 denotes a syllable pattern comparison unit which, for the specific part Y ^no determined by the specific part determining unit 17, starts from the frame corresponding to the preceding vowel y ^no _j1 that defines the specific part, and compares the subsequent vowel y ^no The partial feature pattern sequence ending at the frame corresponding to _j2 is read from the buffer memory 6, and the specific portion Y ^no =y ^no _j1 , . . . stored in the syllable standard pattern storage section 18 is read out.
y ^no _j2 (j ₁ < j ₂ ) defined by y ^no _j1 Cy ^no _j1+1 , y ^no _j1+1 C
y ^no _j1+2 ,
…, y ^no _j2-1 Cy ^no _j2 , y ^no _j1 y ^no _j1+1 , y ^no _j1+1 y ^no _j1+2 ,
...,y ^no _j2-1
A syllable standard pattern series (compound syllable standard pattern series) in which standard patterns corresponding to each syllable such as y ^no _j2 (C: consonant) are combined in various ways to correspond to the vowel string of the specific part ^no (for example, y ^no _j1 Cy ^no _j2 ,y ^no _j1 y ^n
o _j2 ,
y ^no _j1 C ₁ y ^no _j1+1 C ₂ y ^no _j2 , y ^no _j1 Cy ^no _j1+1 y ^no _j2 , etc.). For example, in FIG. 2, the partial feature pattern sequence corresponding to the specific part from the first vowel "A" to the second vowel "I" in the input vowel string is the compound syllable standard pattern "A・C".・Matched with "I" (C: consonant).
This can be done using the well-known DP matching. That is, the compound syllable standard pattern "A.
The compound syllable standard pattern series corresponding to “C・I” is R
= (r ₁ , r ₂ , ..., r〓, ..., r〓), and the partial feature pattern sequence ¹ = (a ¹ ₁ , a ¹ ₂ , ..., a ¹ 〓, ..., a ¹ 〓,
δ(τ, λ) is the τ-th of the partial feature pattern series ¹ .
When the distance between the th feature pattern ^a1〓 and the λth standard feature pattern r〓 of the compound syllable standard pattern series R is (τ, λ)=min(τ-1, λ-2)+δ( τ,
λ−1)+δ(τ,λ) (τ−1,λ−1)+δ(τ,λ) (τ−2,λ−1)+δ(τ,λ) (6) The recurrence formula becomes (1 , 1) = 2δ(1, 1), the distance Δ( ¹ , R) between ¹ and R becomes Δ( ¹ , R) = (T, Λ) (7). Here, δ(τ, λ) is a〓=(a〓 ₁ , a〓 ₂
，
…, a〓 ₂₀ ), r〓=(r〓 ₁ , r〓 ₂ , …, r〓 ₂₀ ),
It is generally given by δ(τ, λ) = ₂₀ 〓 ^p=1 ｜a〓〓−r〓〓｜ (8). Further, various forms of the above recurrence formula have been proposed, and only one example thereof is shown here. In the above manner, various consonants C (including the consonant C in some cases where there is no C) sandwiched between the preceding vowel "A" and the following vowel "I" in the partial feature pattern series are obtained. The distance for the complex syllable standard pattern series R having the same complex syllable standard pattern series is determined and output together with the numbers of the standard syllables constituting the corresponding complex syllable standard pattern series. Similarly, specific parts of the second vowel "I" and the fifth vowel "U" of the input vowel string, the fifth vowel
The specific parts of the 7th vowel "U" and the 7th vowel "A", and the specific parts of the 7th vowel "A" and the 8th vowel "U", respectively, are compared with the compound syllable standard pattern series. The distance is determined and output together with the number of the standard syllable that constitutes the corresponding complex syllable standard pattern sequence. Reference numeral 20 is a syllable identification unit, which calculates the minimum value of the distances output from the syllable pattern comparison unit 19 for each of the specific portions, and determines the standard syllable number constituting the composite syllable standard pattern sequence that gives the minimum value. (identification syllable number). 21 is a syllable string storage unit,
The identified syllable number obtained by the syllable identifying section 20 is stored. A word dictionary section 22 stores a series of syllable numbers constituting words to be recognized. For example, for the word "Osaka",
It memorizes a sequence of numbers corresponding to the three syllables "OO", "OSA", and "AKA". Reference numeral 23 denotes an inter-word distance calculating section, which performs matching between the identification syllable number string stored in the syllable string storage section 21 and the word syllable number string stored in the word dictionary section 22. For example, this is as follows. That is, the case where a syllable with the same number exists in the corresponding position in the identification syllable number string and the word syllable number string is set as "1", and the case where a syllable with a different number exists is set as "0". The sum of the number strings is calculated and normalized by the number of syllables that make up the word to obtain the distance between words. The inter-word distance 23 outputs a word number corresponding to the inter-word distance. Reference numeral 24 denotes a word determination unit 24, which determines the minimum value of the distance between words and outputs the word number that gives the minimum value as a determination result. 25 is an output terminal, and the determination result is outputted from the output terminal 25.

In this embodiment, recognition is performed in units of syllables such as V, VV, CV, and VCV, but the present invention can also be applied to units of speech such as demi-syllables and diphons.

(Effects of the Invention) According to the present invention, even when a single syllable is uttered in succession, a stationary point is extracted, the stationary point is regarded as a vowel, the vowel is identified, and the identified vowel is recognized as a series. Matching is performed with the vowel strings constituting the word to be used, correcting misidentifications, insertions, omissions, etc. in the vowel strings, and matching the input partial pattern with syllable standard patterns such as V, CV, VCV, VV, etc. By doing so,
The words and syllable standard patterns to be compared and matched can be appropriately limited, resulting in significant improvements in recognition rate and matching speed.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, and FIG. 2 is a diagram for explaining the operation of a vowel string comparison section. 1...Audio signal input terminal, 2...Feature extraction unit,
3... Power calculation unit, 4... Voice section detection unit, 5...
... Vowel standard pattern storage unit, 6... Buffer memory, 7... Steady point detection unit, 8... Steady point (frame) storage unit, 9... Vowel pattern comparison unit, 10...
Vowel identification unit, 11... Vowel/consonant determination result storage unit, 12... Standard vowel string storage unit, 13... Vowel distance storage unit, 14... Vowel string comparison unit, 15... Vowel string identification unit, 16 ...Vowel correspondence determination section, 17...
Specific portion determination section, 18...Syllable standard pattern storage section, 19...Syllable pattern comparison section, 20...Syllable identification section, 21...Syllable string storage section, 22...Word dictionary section, 23...Word distance Calculation unit, 24...word determination unit, 25...output terminal.

Claims (1)

[Claims] 1 The input audio signal is converted into a series of feature patterns A=(a ₁ ,
a ₂ , ..., a, ..., a _L ); a steady point extracting means that extracts the stationary points of the input signal; Identify and input vowel string X = (x ₁ , x ₂ ,...,
x _i , ..., x _I ) (where I is the number of stationary points of the input speech); and a vowel string of the word or phrase to be recognized as the input vowel string X obtained by the vowel identifying means. (hereinafter referred to as standard vowel string)
Y ⁿ = (y ⁿ ₁ , y ⁿ ₂ , ..., y ⁿ _j , ..., y ⁿ _J n) (however, n (=
1, 2, ..., N) is the class of standard vowel strings, and J is the number of standard vowel strings) to identify the standard vowel string Y ^no (identification result vowel string) closest to the input vowel string X. a vowel string identifying means for determining a correspondence between vowels in the input vowel string X and the identification result vowel string Y ^no based on a matching result between the input vowel string X and the identification result vowel string Y ^no ; a determining means and a partial interval of the input vowel string X determined based on the correspondence relationship=(x _i1 ,...,
x _i2 ) (i ₁ < i ₂ ) A partial sequence of the feature pattern of the input signal (specific partial feature pattern sequence) =
(a _t1 ,..., a _t2 ) (t ₁ < t ₂ ), of the identified vowel string Y no obtained by the vowel string identification means, the specific portion ^no corresponding to the partial interval ^no = (y ^no _j1 ,
…, y ^no _j2 ) (j ₁ < j ₂ ) y ^no _j1 Cy ^no defined by the vowel sequence
_j1+
1 ,y ^no _j1+1 Cy ^no _j1+2 ,…,y ^no _j2-1 Cy ^no _j2 ,y ^no _j1 y ^no _j1
+1 ，y ^no _j1+1
y ^no _j1+2 ,..., y ^no _j2+1 y ^no _j2 (C: consonant), etc., are matched with standard patterns corresponding to each syllable, and a syllable string corresponding to the specific partial feature pattern sequence is obtained. 1. A speech recognition device comprising: a determining means for determining a word or a phrase based on the identified syllable string obtained and outputting the result as a recognition result.