JPH0552516B2 - - Google Patents

Description

[Detailed description of the invention]

FIELD OF INDUSTRIAL APPLICATION The present invention relates to a pattern comparison device that compares an input pattern with a plurality of standard patterns represented by a series of feature vectors and identifies the input pattern, and particularly relates to a pattern comparison device that identifies the input pattern by comparing it with a plurality of standard patterns represented by a series of feature vectors. The present invention relates to an applicable pattern comparison device. 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, word 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 registers in advance all the words to be recognized converted into a series of feature vectors using his/her own voice as a standard pattern in a word dictionary, and then converts the words uttered during recognition into a series of feature vectors. the audio,
Similarly, the feature vectors are converted into a series of feature vectors, the word in the word dictionary is calculated which word is the closest, and the most similar word is taken as the recognition result. However, this method is good when the number of words to be recognized is small, but when the number of words to be recognized increases to hundreds or thousands of words, the following three problems become impossible to ignore. (1) The burden on the speaker in said registration increases significantly. (2) The time required to calculate the similarity or distance between the voice uttered during the 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. OBJECTS OF THE INVENTION The present invention provides a novel pattern comparison device that solves the above three main problems. Structure of the Invention When recognizing word sounds, the basic unit of recognition is a monosyllabic sound, and each monosyllabic sound is stored as a series of feature vectors as a monosyllabic standard pattern. A word dictionary is provided in which is stored as a combination of codes indicating each monosyllable, and based on this word dictionary, the method of combining monosyllables for each word is known, and the monosyllabic standard pattern corresponding to each monosyllable is formed. The combination of the feature vectors is used as the word standard pattern of the word sound, and the input pattern converted into the series of feature vectors of the voice uttered during recognition is calculated to determine which word standard pattern is closest to the word standard pattern. This is the word recognition result. In this way, in the case of Japanese, any word can be expressed as a combination of monosyllables, so the word standard pattern for any word can be represented as a combination of monosyllabic standard patterns, and speakers can There is no need to utter words, only single syllables. Since the number of monosyllables in Japanese is 101, no matter how the number of words increases, it is only necessary to register 101 types of monosyllable sounds. This solves problem (1) above. In addition, as will be described later, the vector relation distance, which requires the most amount of calculation, does not need to be calculated every time it is matched with each of the word standard patterns, but for each frame of input speech with respect to each of the monosyllabic standard patterns. You only need to ask for it once. This does not change even if the number of words increases, so problem (2) above is solved. Furthermore, by storing only the series of feature vectors corresponding to each monosyllable, the word dictionary only needs to store the monosyllable string corresponding to each word as a string of symbols, which increases the number of recognized words. However, the amount of increase in memory is small, and problem (3) above is also solved. Now, let q(w,l) be the name of the lth word clause of the wth word, and let the sound pattern when several monosyllables are uttered in succession be A=a ₁ , a ₂ . . .
a _I , the standard pattern of the nth monosyllable is R ⁿ = b ⁿ ₁ b ⁿ ₂ …
...b ⁿ _Jo , (where w=1, 2, ..., W; l=
1,2,...,Lw;n=1,2,...,N;a _i ,
b ⁿ _j are feature vectors), the standard pattern w of the wth word is w=R ^q(w,1) R ^q(w,2) ……R ^q(w,Lw) = b ^{q( w,1)} ₁ ,b ^q(w,1) ₂ ……b ^q(w,1) _Jq(w,1) b ^q(w,2
) ₁ b ^q(w,2) ₂ ……b ^q(w,2) _Jq(w,2) ……b ^q(w,Lw) ₁ b ^q(w,Lw) ₂
... expressed as b ^q(w,Lw) _Jq(w,Lw) . This shows the connection of patterns. The pattern comparison device of the present invention performs DP matching between this R w and the input audio pattern A,
The distance D (A, R w) between both patterns obtained at that time
The purpose is to find w such that . DESCRIPTION OF THE EMBODIMENTS FIG. 1 is a block diagram showing a first embodiment of the present invention. 1 is the audio signal input terminal, 2 is the input audio signal for frequency analysis, LPC analysis, PARCOR analysis,
This is a feature extraction unit that converts into a series of several sets of numerical values (feature vectors) through correlation analysis or the like. 3 is the nth monosyllable of Japanese N monosyllables (n=1, 2,...
^. _{_} ^_ 4 is a voice section detection step in which a voice section is detected by a well-known method, such as determining the power in each frame from the output of the feature extraction unit 2, and determining a period during which the power exceeds a predetermined threshold as a voice section. Department. 5 is a frame number counting unit that counts the number of frames from the start to the end of the audio section,
The number of the current frame after the start of the voice section is outputted to the terminal 6. 7 has an inter-vector distance calculation unit that calculates the distance d ⁿ (i, _j ) between the input vector a _i of the i-th frame and each vector b ^{n j} constituting the n-th monosyllabic standard pattern R n n=
1, 2, . . . , N; j=1, ² , . The simplest definition of the distance between vectors is the urban distance. That is, each vector is a _i =
(a _i1 , a _i2 , ..., a _in ), b ⁿ _j = (b ⁿ _j1 , b ⁿ _j2 , ...,
b ⁿ _jn
Then, d ⁿ (i, j) = _o 〓 ^k=1 | a _ik − b ⁿ _jk | ...(1). Reference numeral 8 denotes an inter-vector distance storage unit which stores the inter-vector distance d ⁿ (i, j) obtained by the inter-vector distance calculation unit 7. 9 is a word dictionary in which words to be recognized are stored as combinations of codes indicating each monosyllable. Reference numeral 10 denotes a cumulative distance storage unit which stores previous cumulative distances necessary for calculation of cumulative distance, which will be explained next. Reference numeral 11 denotes a cumulative distance calculation unit, which learns how to combine monosyllables for each word based on the code string stored in the word dictionary 9, and calculates the inter-vector distance storage unit 8 according to the order of the combination. This is a cumulative distance calculation unit that calculates the cumulative distance to the current frame from the inter-vector distance stored in the vectors and the cumulative distance stored in the cumulative distance storage unit 10. The calculation result by the cumulative distance calculation section 11 is stored in the cumulative distance storage section 10. Reference numeral 12 denotes a determining unit, which reads out the cumulative distance of the last frame for each word from the cumulative distance storage unit 10 when the input of word speech is completed, and determines the word with the minimum value as the recognition result. 13 is an output terminal for the recognition result. The operation of each part in the above configuration will be explained in more detail below. This description will be made regarding the operation in the i-th frame. In this embodiment, a case will be explained in which the path shown in FIG. 2 is adopted as the DP matching path. That is, when this route is adopted, in a lattice graft in which the horizontal axis is the input frame number i and the vertical axis is the frame number j of the standard pattern,
Standard pattern from coordinates 1, 1 to coordinates i, j
When the cumulative distance between the partial patterns of the 1st to j-th frames of R ⁿ and the partial patterns of the 1st to i-th frames of input pattern A is D ⁿ (i, j), the following recurrence formula is satisfied. do. D ⁿ (i, j) = d ⁿ (i, j) + minD ⁿ (i-2, j-
1+d ⁿ (i-1, j) D ⁿ (i-1, j-1) D ⁿ (i-1, j-2) ...(2) First, the output vector a _i of the feature extraction unit 2 and the monosyllable The distances to all the vectors constituting the standard pattern are calculated as described above by the inter-vector distance calculation section 7 and stored in the inter-vector distance storage section 8. Formula (2)
The distance between vectors required to calculate d ⁿ (i,
j) and d ⁿ (i-1, j), the inter-vector distance storage unit 8 stores the inter-vector distance between the current frame i and the previous frame i-1 of the input pattern as n=1, 2... , N, j=1, 2, ..., J ⁿ . The cumulative distance calculation unit 11 basically performs the calculation of equation (2), and the calculation is performed according to the monosyllable string presented by the word dictionary 9. Now, the case of matching with word w (w=1, 2, . . . , W) will be explained. Word w consists of Lw monosyllables, the lth monosyllable of word w is (w, l), and for word w, (w, 1)
from the coordinates (1,1) of (w,l) to the coordinates (i,
The cumulative distance D ^(w,l) (i,j) to j) is calculated as a continuation of the matching result up to the previous monosyllable (w,l-1), and D ^(w,Lw) (I,J ^Lw ) is the cumulative distance that is the matching result of the input pattern and word w. Therefore, when adopting the matching route shown in Fig. 2,
Since the initial value of the cumulative distance in the lth monosyllable of word w is the cumulative distance to the last two frames of the l-1st monosyllable of word w, the calculation of equation (2) is as follows:
It is easier to understand if the calculation is done separately for 1, 2 and 3jJ ^q(w,1) . Therefore, if the monosyllabic name of monosyllabic (w, l) is q(w, l), then

[Table] becomes. However, the initial conditions are D ^(w,0) (-1,0)=0 d ^q(w,1) (0,1)=0 D ^(w,0) (i,0)=∞ D ^{(w, 0)} (i, -1) = ∞ D ^{(w, l)} (-1, j) = ∞ D ^{(w, l)} (0, j) = ∞ J _q(w,0) = 0. The results of the above calculations are sequentially stored in the cumulative distance storage unit 10, but as is clear from equation (2) or FIG. 2, the past cumulative distance necessary to calculate the i-th frame is Since the values are only for the -1 frame and the i-2 frame, the cumulative distance storage unit 10 only needs to store the cumulative distances for the previous and two previous frames. Furthermore, as a result of the above calculation, the final value for the intervening word in the i-th frame for each word is
D ^(w,Lw) (i, J ^q(w,Lw) ) is also stored in the cumulative distance storage unit 10. The processing of the i-th frame has been described above, and the count value of the frame number counting unit 5 is set to i. Therefore, the above processing requires only one frame.
This is done each time the voice section ends, i.e., i=I.
Then, the cumulative distance storage unit 10 stores the final cumulative distance D ^(w,Lw) (I, J ^q(w,Lw) ) for each word, and the speech interval detection unit 4 detects the end of the speech. Then, this D ^(w,Lw) (I, J ^q(w,Lw) ) becomes w=1,...
The determination unit 12 finds w=argmin[D ^(w,Lw) (I, J ^q(w,Lw) )] w and takes w as the recognition result. Here, argmin[f(x)] means x that minimizes f(x). FIG. 3 is a flowchart showing the operation of the above embodiment, and this flowchart can also be followed when implementing it by software. Steps (100) to (105) are the parts for initialization. Steps (106) to (115) represent processing in the i-th frame, and steps (107) to
(109) is the part that calculates the distance between vectors, steps (110) to (115) are the part that calculates the cumulative distance, step (111) is the part that performs initialization, and step (113) is the part that calculates the distance between vectors. Cumulative distance for,
Steps (114) to (115) are steps for calculating the cumulative distance for 3jJ ^q(w,l) . Step (118) is a part for finally determining the word with the smallest cumulative distance, and corresponds to the calculation performed by the determining unit 12 in FIG. Next, a second embodiment will be described. This is an improvement on the first embodiment. In other words, when consecutive monosyllables are made, patterns near the boundaries of monosyllables become ambiguous, so it is necessary to match the standard pattern with the beginning and end of each monosyllable free. By allowing frames to be skipped and matched appropriately,
It becomes possible to perform matching with higher accuracy. This is the cumulative distance calculation unit 1 in FIG.
This can be easily realized by slightly changing the method of calculating the cumulative distance in 1. That is, the calculation of the recurrence formula in the cumulative distance calculating section 11 is changed as follows. Let the end point free sections in the head-tail part of the monosyllabic speech pattern of the standard pattern be δ ₁ frame and δ ₂ frame, respectively. That is, the matching start frames for each monosyllabic standard pattern n _are
The frames between the frames are set as frames, and the matching end frame is set as a frame between the J ⁿ -δ ₂ to ^{J n} frames, and each optimal frame is selected. In this case as well, if the constraint conditions shown in FIG. 2 are adopted for the matching path, the cumulative distance D ^(w,l) (i,j) will be changed as follows. In other words, when calculating the cumulative distance of the lth monosyllable,

[Table] becomes. FIG. 4 is a flowchart of the operation of this second embodiment, and steps with the same numbers as in FIG. 3 perform the same processing as in FIG. 3. Step (117') sets the free interval of the terminal point to J ^q(w,l) −δ ₂ 〜J ^q(w,l) , so D that appears in the calculation of the cumulative distance when l=1 ^q(w,0) (i-1, J ^q(w-0) -k) with k=0, 1, 2,
..., to make it ∞ over δ ₂ .
Step (118) is the i-th syllable (w, l-1)
This is the part in which the cumulative distance up to frame -1 is determined as the minimum value within the free section of the end point.
Step (119) is the process when j = 1, 2, Steps (120) and (121) are the process when 3jδ ₁ , Steps (122) and (123) are the process when j _{= 1} + 1j
This is the part that performs the processing when J _q(w,l) . By doing the above, it is possible to realize the free matching of the starting end. As described above, in the second embodiment, the recognition rate can be improved because unstable parts of monosyllable joints can be skipped appropriately, but in order to perform more fine-grained matching, We propose a method to introduce weights. In other words, in normal matching, all frames to be matched are matched with uniform weights, but in each pattern, important parts that better represent its characteristics are given large weights, and By performing matching with a small weight on the parts that are not the same, it becomes possible to sufficiently identify patterns that are close to each other and are therefore likely to be confused. FIG. 5 shows a third embodiment that is realized as a more reliable recognition device by introducing weights to the first embodiment. The difference from the first embodiment shown in FIG. 1 is that a weight count storage section 14 is added, and the operation of the cumulative distance calculation section 11 is performed using this weight count. When employing the matching paths shown in FIG. 2, the weights for each path can be set as shown in FIG. When weighting is performed in this way, no matter how you choose the matching path between the n-th standard pattern and the input pattern, the sum of the weights along that path is _Jo 〓 ^j=1 H, where M is the number of frames of the input pattern. ⁿ (j)+M, which is constant for the standard pattern and input pattern. The calculation in the cumulative distance calculating section 11 is as follows.

【table】

【table】

According to [Table], the recognition result is the word for which the cumulative similarity of each word up to the final frame divided by the sum of the weights for that word is the minimum (I is for all words. It can be omitted since it is common.) At this time, if the weighted sum _Jo 〓 ^j=1 H ⁿ (j) for each monosyllable is kept constant, I
Since w is constant during matching with all standard patterns (words), w can be obtained as follows. w= ^argmin _w [D ^q(w,Lw) (I, J ^q(w,Lw) ))/Lw] FIG. 7 is a flowchart showing the operation of the third embodiment. Step (200) ~
(201) is the part where d ⁿ (1, 1) is calculated in advance. Steps (202) to (207) are parts for setting initial values when calculating the recurrence formula. The process when i=1 is completed in step (207), so
Steps (208) to (217) are the processes after i=2. Steps (209) to (211) are steps for determining inter-vector distances for all monosyllables in input frame i. Step (212)
~(217) is a part for calculating the cumulative distance D ^(w,Lw) (I, J ^q(w,Lw) ) for each word w. Step (213) is the part that gives the initial value at that time. Steps (214) to (217) are the part that calculates the cumulative distance for the lth monosyllable of word w, and step (215) is the part that calculates the cumulative distance for each monosyllable.
In the case of 1 and 2, steps (216) to (217) are 3
The cumulative distance is calculated for the case jJ ^q(w,l) . Step (218) corresponds to the determination section 12 and is as described above. As described above, in the first to third embodiments, the constraints shown in FIG. 2 are used as the matching path constraints, but various other paths such as those shown in FIGS. 9a to 9d can be considered. At this time, the weight for each route should be such that when the standard pattern to be compared and the input pattern are fixed, the sum of the weights along the matching route does not depend on how the route is selected. It is shown in FIG. If H ⁿ (j) = 0, the weight of each path is 1, which is the normal case. Furthermore, in the above embodiment, a case was described in which a monosyllabic speech pattern is registered as a standard pattern, but if this is used as a pattern for single erroneous speech, consecutive word speech can be recognized in exactly the same way. can. This is particularly effective when the method of continuity is predetermined. Also, if you have a VCV (vowel + consonant + vowel) pattern as a standard pattern instead of a single syllable, and combine them to form a standard word pattern, you can make the input voice sound more natural. In contrast, the recognition rate can be improved. In the third embodiment, the sum of the weights along the matching path is constant for each single syllable regardless of the path, but it is of course possible to make it constant for the entire word. It is. Naturally, it is also conceivable to introduce a weighting method as explained in the third embodiment to the free matching of the starting and ending points in the second embodiment. This can be easily achieved by setting the weight H ⁿ (j) to zero for the interval where the start and end points are free. Furthermore, although only the case with an audio signal has been described in the embodiment, it is possible to recognize a pattern configured as a series of basic patterns, and there are several predetermined ways of making the series of basic patterns. In this case, if the basic pattern is prepared as a standard pattern, the continuous pattern can be recognized in the same manner as in this embodiment. Effects of the Invention According to the present invention, the following problems are encountered in the recognition apparatus using the speaker-specific registration method for large vocabulary words: (1) The burden on the speaker during standard pattern registration is large. (2) It takes time to match the standard pattern and the input pattern, which slows down the response of the recognition device. (3) The memory for storing standard patterns becomes enormous. We were able to solve these issues all at once. In addition, it has become possible to improve the recognition rate by introducing free start and end points and weights.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a pattern comparison device in a first embodiment of the present invention, FIG. 2 is a diagram showing a DP matching path, FIG. 3 is a flowchart showing the operation in the first embodiment, and FIG. 4 is a diagram showing a DP matching path. A flowchart showing the operation in the second embodiment, FIG. 5 is a block diagram of the pattern comparison device in the third embodiment of the present invention, FIG. 6 is a diagram showing weighting in each path, and FIG. FIG. 8 is a flow chart showing the operation of the embodiment, FIG. 8 is a diagram showing examples of various weightings, and FIGS. 9 a to 9 d are diagrams showing various routes. 2... Feature extraction section, 3... Monosyllabic standard pattern storage section, 9... Word dictionary, 11... Cumulative distance calculation section, 12... Judgment section.

Claims (1)

[Claims] 1. Convert the input signal into a sequence of feature vectors a ₁ , a ₂ ,...
..., a _I , and the nth standard pattern consisting of a series of feature vectors _Ro = b ⁿ ₁ , b ⁿ ₂ , ..., b ⁿ _jo (where n is the number of types N) When n
∈{1, 2, ..., N}) and a combination pattern created by combining this standard pattern (hereinafter, the name of this combination pattern will be referred to as w∈{1, 2, ...
..., W}, and each w is called a word) is expressed as n
A combination pattern of the standard pattern corresponding to the word w and the word dictionary that stores what is expressed by the array of R ^q(w,1) R ^q(w,2) ...R ^q(w,Lw) However, q(w,k) is the word w∈{1,2,
..., W}, the name of the k∈{1, 2, ..., L _w }th standard pattern among the L _w standard patterns constituting the pattern represents a combination of patterns) and the input pattern a ₁ , a ₂ , ..., a _I is the combination of the feature vector b ^q(w,k) _j that makes up the combined pattern of the standard pattern and the feature vector a _i that makes up the input pattern. means for calculating cumulative distance (similarity) obtained by minimizing (maximizing) using dynamic programming as a function consisting of;
determination means for finding the word for which this cumulative distance (similarity) is minimum ( _maximum ) ^; The distance (similarity) d _o (i, j) to _i frame a i is n=1,...,N,
An intervector distance (similarity) calculation means that calculates j=1, ..., J _o , and when the k-th standard pattern forming the word w is (w, k), (w,
1) from the first frame to the j-th frame of (w, k) and the first partial pattern of the input pattern.
The cumulative distance (similarity) D _(w,k) (i,j) from the frame to the i-th frame with the partial pattern is calculated for each frame i, k=1, ..., L _w , j=1, ... ...,
For J _q(w,k) , using the calculated d ⁿ (i,j), calculate the intermediate cumulative distance (similarity ) calculation means, D _(w,Lw) (I,
A pattern comparison device characterized in that J _q(w,Lw) ) is a distance (similarity) between a standard pattern sequence for a word w and the input pattern.