JPH02186400A - Word speech recognition method - Google Patents

Abstract

PURPOSE:To allow the exact calculation of word candidates at a high speed by executing word voice recognition by the system of two stages consisting in obtaining rough broadly classification word candidates by a large classification recognition processing and further pinpointing the broadly classification word candidates by subdivided classification recognition processing to obtain subdivided classification word candidates. CONSTITUTION:A character extracting means 101 extracts character quantities from each of the plural parts of the input word voices. A broadly classification recognition means 106 computes the similarity of respective learning word voices 110 stored in a dictionary memory means 104 and unknown input word voices 111 via respective means and outputs the same as the broadly classification recognition means 112 with sequence in order of the learning word voices of the higher similarity. DP matching with the vowel systems outputted from a vowel system dictionary memory means 108 and a vowel system recognition means 107 respectively is executed and the subdivided classification word candidates 113 are outputted. The word candidates are exactly pinpointed in this way and the high-speed recognition is executed.

Description

[Detailed Description of the Invention] Summary] Regarding a word speech recognition method that extracts features from unknown input word speech and recognizes the word speech, word candidates can be accurately narrowed down and recognized at high speed. We have made a promise to realize a word speech recognition method that can perform the Classifying means for classifying a plurality of learning word sounds for each part after extracting feature quantities for each part by the quantity extraction means; representative feature quantity calculation means for computing representative feature quantities for each class; Dictionary storage means for storing representative feature values for each class and types of learning word sounds included in each class as a dictionary for each part, and a feature extraction means for extracting feature values for each part from unknown input word sounds. , similarity calculation means for calculating the degree of similarity for each part with a representative feature of each corresponding class in the dictionary storage means, and each of the learning words stored in the dictionary storage means based on the similarity. A large classification recognition means that calculates the degree of similarity between the speech and the unknown input word speech and outputs the learned word speech with a higher degree of similarity as a ranked large classification word candidate, and recognizes the vowel sequence of the unknown input word speech. vowel sequence recognition means; vowel sequence dictionary storage means for storing word-like vowel sequences of each learning word sound as a dictionary; extracting the vowel sequence of each major classification word candidate from the means;
The vowel sequence recognition means performs DP matching on each of the vowel sequences of the unknown input word speech outputted from the vowel sequence recognition means, and outputs the major classification word candidates with rankings in descending order of the degree of matching. Configure.

[Industrial application field]

The present invention is a word speech recognition method (prior art) that extracts features from unknown input word speech and recognizes the word speech.[Prior art] Features are extracted from unknown input word speech and the word speech is recognized. In this case, the characteristics of the unknown input word sound are usually
Special features of the amount and pre-memorized learning word sounds! A common method is to match the learning word sound with the highest IJ (1 degree) as the recognition result. In this case, one word sound is a time-series pattern of multiple sets of feature values. Therefore, when matching unknown input word sounds and learning word sounds, the special @ of the unknown input word sounds,
Various matching methods can be considered depending on how to match the time-series pattern of the amount with the time-series pattern of the feature amount of the learning word speech.

The first and most common conventional example is to make each time-series pattern of features of unknown input word sounds correspond to each time-series pattern of features of learned word sounds in order from the beginning of each word. Then, the similarity (distance) between each feature is calculated, and the similarity between words is calculated by summing it over the entire time series pattern. Ordinary pattern matching method that uses the high (minimum distance) learned word sound as the recognition result.
As a second conventional example, each of the time-series patterns of the features of the unknown input word speech and each of the time-series patterns of the features f & ffi of the learning word speech are expanded and contracted non-linearly in the time axis direction, and the distance between each feature is The distance calculation is repeated to calculate the cumulative distance for each word, and this is done for all the learning word sounds.
D where the learning word voice with the smallest cumulative distance is the recognition result
There is a P matching method.

C invention tries to solve the problem! However, in -111, word sounds usually have different word lengths because the utterance time is not constant.
In the conventional example, such a difference in word length is not taken into consideration, so it has a problem that it is difficult to obtain a high recognition rate.

On the other hand, the second conventional example is a method that takes into account the difference in word length, and can obtain a high recognition rate; however,
The amount of calculation is large because it is necessary to perform nonlinear expansion and contraction of the time axis for each word. In particular, if this calculation is repeated for all the learning word sounds, the total amount of calculation becomes extremely large. However, the problem is that real-time processing is difficult.

Furthermore, both the first and second conventional examples have the problem that the amount of calculation is basically large because matching with all the learning word sounds is required.

In order to solve these problems, a method can be considered that first roughly categorizes the unknown input word speech to narrow down the word candidates, and then performs detailed recognition on the word candidates narrowed down to a small number. However, in this case, there are currently few recognition methods that can narrow down word candidates quickly and accurately, and in the end, it is difficult to realize a word speech recognition method that is fast and has a high recognition rate using the method described above. The problem is that it is difficult to

An object of the present invention is to realize a word speech recognition method that can accurately narrow down word candidates and perform recognition at high speed.

[Means to solve the problem]

FIG. 1 is a block diagram of the present invention.

The feature amount extraction means 101 is a means for extracting feature amounts from each of a plurality of parts of the input word speech. For example, it calculates a feature vector such as a cepstral coefficient for each frame of the input word speech, and then extracts the feature amount from each of a plurality of parts of the input word speech. This is achieved by dividing the pattern into a plurality of parts (partial patterns), calculating the average of the feature vectors for each part, and calculating the feature amount for each part.

The classification means 102 is used when creating a dictionary, and the feature extraction means 101 is used for each of the plurality of learning word sounds 110.
After the feature values are extracted for each part, the learning word speech 1i00 is divided into a plurality of classes for each part, and is realized, for example, by means of clustering the feature values for each part.

The representative feature calculation means 103 is also used when creating a dictionary, and calculates the representative time fiTh! of each class classified into each part by the classification means 102. ! (standard partial pattern), and is realized, for example, by means of calculating the average of the feature amounts of the corresponding portions of the learning word speech 110 included in each class to obtain the representative feature (2).

The dictionary storage means 104 is a means for storing representative feature amounts for each class and the types of learning word sounds included in each class as a dictionary for each part, and is realized by, for example, a ROM (read only memory) or a disk storage means. Ru.

Next, the similarity calculation means 105 calculates the unknown input word sound 1.1.
1 P! The feature extracting means 101 extracts 1r1. n
is extracted, the representative characteristic rt1. 1 with quantity (
It is means for calculating limits, and is realized, for example, by means for calculating Euclidean distance between feature quantities.

The large classification recognition means 106 is also an unknown input word voice 111.
Calculates the degree of similarity between each learning word voice 110 stored in the dictionary storage means 104 and the unknown input word voice 111 based on the degree of similarity used in recognition and calculated by the degree of similarity calculation means 105; This is a means for outputting learning word sounds with a higher degree of similarity as a ranked major classification recognition means 112 in order. For example, for a certain learning word audio 110, the means
The class in which the learning word sound 110 is included for each part is determined by referring to the dictionary storage means 104, and the similarity calculated by the Mm degree calculating means 105 is referred to and extracted for that class. Means for calculating the sum of 8m degree and using it as similarity for the learning word speech 110, means for repeating this process for all learning word speeches 110, and similarity of each learning word speech 11O obtained in this way. This is realized by means for selecting learning word sounds in order of similarity from among them and outputting them as the major classification recognition means 112.

On the other hand, the vowel sequence recognition means 107
The unknown input word speech 11 is used during the subclassification recognition processing after the large classification recognition processing in step 6, and is used independently of the above processing.
This is a means for recognizing one vowel series, for example, by matching the feature quantities of five vowels stored in advance with the feature quantities for each frame of the unknown input word speech 111, and recognizing and extracting the vowel series for each frame. This is achieved by means of

The vowel sequence dictionary storage means 108 stores each learning word sound 110.
This is a means of storing word-like vowel sequences in a dictionary.

The sub-category recognition means 109 is used after the large-class recognition processing by the large-class recognition means 106, and is performed by the large-class recognition means 1.
The vowel series of each major classification word candidate 112 outputted from 06 is extracted from the vowel series dictionary storage means 108, and DP matching is performed on each vowel series of the unknown input word speech 111 outputted from the vowel series recognition means 107, This is a means for outputting as ranked sub-category word candidates 113 in descending order of the degree of matching.

[For production]

The operation of the above means will be explained below.

First, the dictionary creation process, which is a prerequisite for the major classification recognition process, will be explained.

■First, in the feature extracting means 101, after inputting a plurality of learning word sounds 110 in sequence, each learning word sound 110
is divided into a plurality of identical parts, and the feature amount of the entire learning word speech 110 is extracted for each part.

(2) Next, the classification means 102 sequentially determines each of the plurality of parts, and classifies the plurality of learning word sounds 110 based on the feature amount of each learning word sound 110 corresponding to the part.

(2) Thereafter, the representative feature amount calculating means I03 calculates the representative feature amount of each class classified by the above processing.

(2) Then, the representative feature amount of each class and the types of learning word sounds included in each class are stored in the dictionary storage means IO4.

The above processes (1) to (2) are executed independently for each of the plurality of parts. And when all processing is finished,
Finish the dictionary creation process.

Through the above process, the learning word speech is first divided into multiple parts, and each part is further divided into classes, and only the representative features of each class are stored in the dictionary, so the total storage capacity is Compared to the case where the feature amount of the learning word speech 110 of is memorized, it decreases according to the number of classification divisions. Note that since the storage capacity required to store the types of learning patterns included in each class is smaller than the storage capacity of feature quantities, the storage capacity of the dictionary can be reduced overall.

Next, the major classification recognition process will be explained.

■First, the feature extracting means 101 inputs the unknown input word speech 111, divides it into a plurality of parts that are the same as the learning word speech 110, and separates each part into special 1tl [ff
Extract i.

0 Next, the similarity calculating means 105 sequentially determines each of the plurality of parts, and calculates the unknown input word sound 111 corresponding to the part.
The degree of similarity between the feature amount and the representative feature amount of each class corresponding to the above portion in the dictionary storage means 104 is calculated.

■Repeat the above process (■) for all parts.

■Then, the Oita R recognition means 106 reads the dictionary storage means 10.
Regarding a certain learning word sound 110 stored in 4,
The degree of similarity of the learning word sounds is calculated from the similarity of the classes for each part in which the learning word sounds are included.

■Repeat the process in ■ above for all the learning word sounds.

[Phase] As a result of the process (2) above, the learning word sounds are ranked in order of similarity and output as major classification word candidates 112.

With the Oita ML recognition process described above, the similarity between the unknown input word speech 111 and the learning word speech 110 only needs to be calculated using the representative feature values of each class for each part. Compared to calculating the similarity between
The amount of calculation can be significantly reduced.

Next, detailed classification recognition processing will be explained.

(2) First, independent of the above-mentioned major classification recognition process, the vowel sequence of the unknown input word speech 111 is recognized by the vowel sequence recognition means 107.

0 Next, the major classification recognition means 10 is processed by the major classification recognition process.
The major classification word candidates 112 outputted from 6 are input to the subclass recognition means 109, and the vowel series of each major classification word candidate 112 is extracted with reference to the vowel series dictionary storage means 108.

■Among the extracted vowel sequences, each major classification word candidate 112
DP by the vowel sequence corresponding to
Perform matching.

(2) The above process (2) is repeated for all major classification word candidates 112.

(2) As a result of the process (2) above, the learned word sounds are ranked in descending order of matching degree and output as subclassified word candidates 113.

By the above-mentioned sub-classification recognition processing, L2
The ranking of the recognized major classification word candidates 112 is further narrowed down, and finally, the detailed classification word candidates 1 are accurately ranked.
13 can be obtained. In this case, major classification word candidate 1
12, the number of candidates was originally narrowed down to about 10th place, and the number of vowel sequences per - word candidate is not so large (several), so the vowel sequence of each major classification word candidate 112 and the unknown input word sound DP matching with the 111 vowel series can be performed at high speed.

As described above, in the present invention, when recognizing word sounds based on word candidates, first, a rough classification word candidate 112 is obtained by a major classification recognition process, and then a rough classification is performed by a subclassification recognition process. The feature is that word candidates 112 are narrowed down to obtain subclassified word candidates 113. Such a two-step method makes it possible to calculate word candidates quickly and accurately. Then, for the subclassified word candidates 113 whose number of candidates has been narrowed down in this way, the D
By performing final recognition using P matching etc.
Compared to recognition methods that perform DP matching etc. from the beginning,
It is possible to realize a word speech recognition system that requires significantly less calculation and has an equivalent or higher recognition rate.

The vowel sequence of the unknown input word speech 111 outputted from the vowel sequence recognition means 107 is configured so that a plurality of vowel sequence candidates are outputted, and is combined with the vowel sequence of each major classification word candidate 112 in the subclassification v2 recognition means 109. By performing DP matching on the plurality of vowel sequence candidates mentioned above,
The influence of vowel recognition errors in the vowel sequence recognition means 107 can be minimized.

Furthermore, in this case, by adding ranks to the plurality of vowel sequence candidates of the unknown input word speech 111 and adding weights according to the ranks during DP matching, more appropriate matching can be performed.

〔Example] Hereinafter, embodiments of the present invention will be described in detail.

(Configuration of this embodiment) FIG. 2 is an overall configuration diagram of an embodiment of a word speech recognition device according to the present invention.

The input speech 209 is first input to the preprocessing unit 201, where it is subjected to various processes (described later), such as word segment extraction, linear normalization, and speech analysis, to produce a linearly normalized feature vector time series. Output pattern 21O.

This feature vector time series pattern 210 is input to a partial pattern extraction unit 202, where a partial pattern 211 is obtained.

Subsequently, the partial pattern 211 is input to the inter-partial pattern distance table calculating section 203, where the distances from each standard partial pattern stored in advance in the standard partial pattern dictionary section 204 are calculated, and the partial pattern distance table 2 is calculated. ]
2 is created.

The partial pattern distance table 212 is generated by the major classification recognition unit 2.
05, the distance for each word is calculated based on the partial pattern distance table 212 while referring to the word dictionary section 206, and word recognition is performed based on the calculation result. Candidate 213 is output.

The major classification word candidates 213 are further input to the subclassification recognition unit 207 . Subclassification P! In the Oribe 207, the input audio 209
Enter again to perform subdivision I11?2 recognition.

Then, subclassification recognition is performed while referring to the intra-word vowel sequence dictionary section 208, and as a result, a final subclassification word candidate 214 is obtained.

(Operation of the present embodiment) The operation of the present embodiment having the above configuration will be explained below.

First, the operation of the preprocessing section 201 shown in FIG. 2 will be explained. FIG. 3 is a diagram showing an operation flowchart of the preprocessing section 201. The explanation will be given below.

First, the input audio 209 is 9kHz, 60dB.
Passed through a 10ct low pass filter (LPF) (5
301).

Next, it is A/D converted at 24 kHz with a 12-bit sign and converted into a digital audio signal (3302).
.

Next, the temporal change in power is calculated for this signal, and a power pattern 40 is obtained as shown in FIG.
1, by setting a constant threshold value 402, a word section 403 as shown in the figure is automatically cut out (S303).

Next, the word section thus extracted is linearly normalized into 48 frames in the time direction (S304). Here, a frame is a basic time unit of speech analysis, which will be explained next. , one frame is 21.3 m5ec.

Next, audio analysis is performed for each frame (S305-53
09-S305 loop).

That is, from the beginning of the digital audio signal linearly normalized as described above, the time length is 21.3 m5ec.
(512 points) Hama+ing IJJin
The signal is multiplied by do-, and the digital audio signal of each frame is cut out (S305).

A 512-point FFT (Fast Fourier Transform) analysis is performed on the digital audio signal of each frame extracted in this way and converted into a power spectrum (3
306).

For the obtained 512-point power spectrum,
50-channel bandpass filter (BP
F) performs division and integration, and converts the frequency axis into a Mel scale (5307).

Then, by cepstral analysis, cepstral coefficients (xi', x2'', . . . up to the 10th order) are determined.

xlo') is determined (330B), however, the determined cepstrum coefficients are normalized as shown in the following equation.

X1=X('/ΣX+”...(
1) j = 1 The above process is repeated for all 48 frames of the engineering word section 5309-3305), and the cepstrum coefficients normalized by the above equation (1) for 48 frames obtained by this are It is output as a vector from the preprocessing unit 201 in FIG.

° Pattern/Preservation In this embodiment, word recognition is performed using the feature vectors obtained as described above. Here, when performing word recognition, which will be described later, it is necessary to register standard partial patterns and word dictionaries in advance in the standard partial pattern dictionary section 204 and word dictionary section 206 shown in FIG. 2, respectively. Therefore, first, before the word recognition process, the standard partial pattern and word dictionary creation process will be explained.

First, the preprocessing unit 201 processes the learning pattern of the input audio 209, and the resulting n frames (
In the above explanation, the time series pattern of the feature vector of the word speech Wk linearly normalized to n=48 frames is expressed as Wk= (Xk+, Xhz, . . . , Xk−)
...(2) k=1.2. ..., N (see Figure 4), where N is the genus of the word, Xht
(m=1, 2, . . . 1 mouth) is a feature vector (consisting of 1st to 10th order cepstral coefficients) of the Lth frame in the temporal direction of the word Wk.

Next, the time-series pattern of the feature vectors in equation (2) above is input to the partial pattern extraction unit 202 in FIG. 2, where the feature vector time-series pattern of the word speech Wk is A mask ψ, t(li''1 to ~M) is applied to extract the partial pattern W(''l,).

That is, w'''h=XP-(Wb)・
...Perform the calculation (3). The above equation (3) is calculated by N types of word sounds W, , W2. ..., the set of partial patterns obtained for each feature vector time series pattern of WN is (w ''' , , w '''')2... , (
s), )...(4) m=1.2. ...Represented in 1M.

Now, focusing on the partial pattern w'' corresponding to one mask ψ out of the set of partial patterns in equation (4) above, N kinds of word sounds W,, W2゜..., WN are calculated. Divide into T classes C"' t (m=1 to T). This classification is performed using the most distant clustering algorithm in the following steps.

However, N>T, C"' 1 represents the first class of the partial pattern corresponding to in the mask.

■c"'t = (w++mJ),
(m=1.2+ ” ' IN) count=
N That is, here, it is divided into N classes as an initial value, and each class has one partial pattern W3''.
) I. Note that count is a variable representing the number of class divisions, and as described later, clustering ends when count=T.

■D((: <112', , C(Ill', )
=D (C"'+, C"'J) =
m a X (d (W ”'p, W ”'a)
)w (-1pE C(ml it W
Ts), E C(llll jd (w ”' p+
W ””)Q)-Σ (w ”' pt −w ”'
. )2m・・・(5) Here, w l+″)pl, w″″)□ are the feature vectors of the L-th frame in the temporal direction in the partial patterns w fl+1 p and w (m)9, respectively, d (w ++wl, W(ffi)q) defined by the above formula indicates the Euclidean distance between the partial patterns w (+'')p and w is'''9.

(2) C(m+' and C(Ml') are integrated into one class.

■ eount=eount −1 ■ If count is not equal to T, repeat the steps from ■ to ■; if count is equal to T, end the clustering process.

By using the above clustering algorithm, N kinds of word sounds W,, W, corresponding to one mask ψ, are obtained.

..., each partial pattern w of WN (+'''k(k・1
~N) are classified into T classes C(aゝ+ (t=X~T).

Next, standard partial patterns w (t″l) of T classes corresponding to the mask ψ found as above.

Calculate (m=1-T). This process is performed for each class C(I
For each class (m=1-T), it is calculated as the average value of the partial patterns included in that class, and the partial pattern corresponding to the mask ψyo of the word sounds included in each class is calculated as the average value of the partial patterns included in that class. is represented by the standard partial pattern.

The state of the clustering process and standard partial pattern creation process is shown in FIG. 7 as an example for the mask ψ.

The above processing is performed for N types of word sounds W, , W2. ...WH
The same process is performed for each of the masks ψ1 to ψN.

This shows that the N types of word sounds W, , W, . . . , WN are classified into T classes independently for each mask. Then, the standard partial pattern w''>1 (m=1-T) is calculated together with this.

(m=1-M) is the standard partial pattern dictionary section 2 in FIG.
04.

Next, by the above clustering process, it is determined to which class among the T classes each of the N types of word sounds W,,W,,...,WN are classified for each of 1 to ψ in each mask. is stored in the word dictionary section 206 in FIG. An example of a word dictionary is shown in FIG.

That is, N types of word sounds Wl, W2 . ...,WN
The class number for each mask ψ1 to ψH is stored for each mask. With this word dictionary, each word sound W, , W2 .
..., WN can be expressed by a collection of standard partial patterns of classes included in each mask ψ, ~ψ, over all masks.

Person mΔ1 Large-separation recognition of input speech 209 is performed based on the standard partial pattern and word dictionary created as described above.

First, an unknown input word is processed into a partial pattern by a partial pattern extracting unit 202 after a feature vector time series pattern 210 is calculated by the preprocessing unit 201 in FIG. (5th
(see Figure 6).

Now, with the unknown input word as Z, the distance between W and the dictionary word Wk is 1
9tD (W, Wb) is defined by the following equation.

D (V/, Wk) = ΣdL′″'e(k+lI)
(116) However, d'''')c(k+m)-d(w '''C(11+l
Il+ ” ”’) ” (7) 曾1)=ψ, (kai)
...(8) Here, c (k,
m) and “ ”’ e (k, sl are masks ψ, respectively)
These are the number of the class to which Wk belongs and the standard partial pattern representing that class. Therefore, d(ml.

(k+ml is the standard partial pattern W(' c lk,
m) and the unknown input word-only partial pattern g (ml) (according to the definition of equation (5) above), the unknown input word W and the dictionary word W defined by equation (6) above, The distance is given by summing the distances between the partial pattern of the unknown input word and the standard partial pattern of the dictionary word in each mask ψ, ~ψH over the entire mask. An example of calculation processing of the distance between and dictionary word W is shown.

In the major classification recognition in this embodiment, the above distance calculation is performed in the second
This is done for all dictionary words W, ~WN in the figure, but in this case, it is not necessary to calculate the distance for each dictionary word from the standard partial pattern of the class to which that dictionary word belongs.
This is because one class includes multiple dictionary words.

Therefore, first, before calculating the distance to each dictionary word, each mask ψ is calculated in advance in the inter-partial pattern distance table calculation unit 203 in FIG. 2. (11 to M), the partial pattern 1 (ml) of the unknown input word W and the standard partial pattern g('I) of each class C (1 ml) stored in the standard partial pattern dictionary section 204 are Distance d ++''h (
t=t~T) is determined and output as a partial pattern distance table 212 as shown in FIG. 1O.

Based on the inter-partial pattern distance table 212 obtained through the above processing, the major classification recognition unit 205 in FIG.
Calculate the distance to W.

That is, when calculating the distance between the unknown input word Z and the specific dictionary word Wk based on the above equations (6) to (8), the standard partial pattern g7 ( Distance d ”' c (k+ sl
Since it always exists in the inter-partial pattern distance table 212, it is only necessary to select the value from the distance table and calculate the above equation (6).

The above processing is performed for all dictionary words w and -wn, and as a result, dictionary words with small inter-word distances up to a predetermined rank are output as major classification word candidates 213 in FIG.

In the process up to Oita RL'2+ knowledge explained above for the first time,
For the feature vector time series pattern 210 linearly normalized to 48 frames output from the preprocessing unit 201 in FIG. frames), the number of classifications T is about 25, and the types of words to be recognized N are about 212, the recognition rate is 10 (major classification word candidates up to 10th place 213
(The percentage of correct results in Figure 2) is 99.6%, and a high classification recognition rate can be obtained. This P! Recognition rate, 10
As far as the position recognition rate is concerned, the recognition rate is almost the same as the first conventional pattern matching method or the second conventional DP matching method shown in the "Prior Art" section.

In the case of the large classification recognition method in this embodiment, the number of times of matching for each mask is not the number of words in each dictionary, N.
Since the number of classes for each class is only T, the amount of calculation can be significantly reduced, and is 1/1 compared to the first conventional pattern matching method.
The amount of calculation can be reduced to about 0, and the amount of calculation can be reduced to about 1/200 compared to the second conventional DP matching method. It can be seen that this is a very effective means as a broad classification recognition method for obtaining candidates 213.

Through the above-mentioned major classification recognition processing, it is possible to ensure that the major classification word candidates 213 (FIG. 2) up to about the 10th rank include correct recognition results with a high probability. Therefore, using this major classification word candidate 213, the subclassification recognition unit 2 in FIG.
In step 07, the major classification word candidates 213 up to about the 10th rank are re-recognized and further narrowed down to the top, thereby finally obtaining the subclassification word candidates 214 with an even higher recognition rate. below,
The detailed classification recognition process will be explained.

First, the overall processing operation will be briefly explained. In the detailed classification L2 recognition process according to this embodiment, the word-like vowel series dictionary section 208 in FIG.
A series of vowels within each dictionary word of the species is registered in advance as a vowel label.

First, the unknown input word input as the input speech 209 is recognized by an appropriate vowel recognition process to obtain a vowel sequence corresponding to the unknown input word. The simplest vowel recognition method is to use each feature of the five vowels (a, i, tsu, ku, o) (such as the cepstral coefficients mentioned above).
There is a method of obtaining a vowel sequence for each frame by keeping it as a dictionary and matching it with the feature amount obtained for each frame of an unknown input word. There are various other conventional methods, but in this example, the first recognition rate is 92%.
Any method may be used as long as it is of a certain degree.

On the other hand, each vowel sequence of the major classification word candidate 213 output from the major classification recognition unit 205 in FIG. 2 is obtained by referring to the word-like vowel sequence dictionary unit 208.

Regarding the vowel series of the unknown input word and the vowel series of the major classification word candidate 213 obtained as above, the DP described below
Perform DP matching using a matching algorithm,
- Delete undefeated words from the major classification word candidates 213 and output the final paper classification word candidates 214.

The specific operation of the above processing will be explained next.

First, as mentioned above, in vowel recognition processing for unknown input words, extraction errors may occur, so -In is the number of vowel sequences of the unknown input word, and the number of vowel sequences of the major classification word candidates 213 is -In. It may not. Therefore, in this embodiment, the number of vowels is not limited, and the matching is performed using the DP matching method, and in consideration of the possibility that vowel recognition processing for unknown input words may be incorrect, the Matching is performed on up to the second-ranked candidates in the vowel series of the input word. The above processing is a feature of this embodiment.

Specifically, let the vowel sequence recognized by the vowel recognition process for the unknown input word W be ν, and </= (V+, Vz
, ---, vt) ---(9) v, = (v
,',v,')...0O)(vt
' , Q l") = (/a/, /i/, /u/
, /e/, 10/ )i=1.2. ...,■
...(11). Here, ■ is the number of vowels of the unknown input word W. In addition, kai represents the i-th vowel in the vowel sequence recognized by vowel recognition processing for the unknown input word W, and 91' and 91'' are the first and second candidates of the i-th vowel, respectively. It is.

On the other hand, the word candidate W obtained as the major classification word candidate 213
Let V be the vowel series of k (up to the 10th place = 1.2..., 10), and Vh = (Vh+, VkZ+ H+ +,
VhJ) 1″Q7J(vht) = (/a/,/
i/, /u/, /e/, 10/)j・1.2. ...
, J... side. Here, J is the number of vowels of the word candidate Wk. Also, VkJ is word candidate Wk
represents the j-th vowel in the vowel series.

The vowel sequence V of the unknown input word W and the vowel sequence V of the word candidate Wk are matched by determining the cumulative path FiID (1, J) using the DP matching algorithm and comparing and determining their magnitude. I do.

That is, the vowel sequence vh word candidate Wk exclusively for the unknown input word
The cumulative distance D(1, J) of the vowel series Vk is calculated by the DP matching algorithm given by the following recurrence formula.

■Initial value setting D(0,0)=O D(+,0)=D(i,-1) =D(0,j)=D(
-1, j) = (1) (in, 2...., I j
=1.2. ..., J) ■ Recurrence formula% formula%) Now, as an example, each vowel series ♀, ■ of the unknown input word W and the word candidate Wi is ν= (vl, vz, %rl) That is, r =3%
'l = (vl', ?+'') = I/a/, 10/
) ν2 ”'(vz ′, 'l>2°') = (1
0/, /a/ )vl = (V3', ?3°') = (
/l/, /e/)Vh = (VHH5Vhz), that is, assume that the vowel sequence is J=2Vk1=/a/Vmz=/i/. 9 and Vk in this case
The cumulative distance D(1, J) is calculated according to the DP munching algorithm from (1) to (2) above, and is as shown in FIG. 1I. As a result, the cumulative distance D(1,
J)=O. That is, the vowel sequence V of the unknown input word W and the vowel sequence V of the word candidate W.

It can be seen that the cumulative distance of W becomes the minimum, and in the end, the word candidate W obtained as the major classification word candidate 213 in FIG. 2 remains as the small molecule n word candidate 214 in FIG. 2 for the unknown input word Z.

! The recognition rate for the 10th place in the major classification word candidate 213 was 99.6%, but the recognition rate for the 10th place in the subclassification recognition candidate 214 was 99.6%.
The position recognition rate was improved to 99.8%, and the recognition rate for the same <15th position was increased to 99.86%. In addition, the word candidates narrowed down to 10th place in the major classification word candidates 213 can be narrowed down to 8th place in the subdivision 8B2 classification) city 214, and the word candidates ranked 30th can be narrowed down to 15th place.

In this case, the number of major classification word candidates 213 is originally narrowed down to a small number, and the number of vowel sequences for each word is also small (several), so the vowel sequences of the word candidate wk and the unknown input word zeng are DP matching can be performed quickly.

In this embodiment, input speech 209 (second
This is an explanation of a method for obtaining word candidates for (Fig.) as subclassified word candidates 214. After word candidates are obtained in this way, further detailed recognition may be performed from the subclassified word candidates 214 to obtain the final recognition result.
Various methods can be considered for this purpose, so we will not discuss them in detail here, but for example, by further performing DP matching on the subclassified word candidates 214 and the input speech 209, a method with the highest degree of matching can be selected. It is conceivable to use this as the final recognition result.

Even if P matching is performed as described above after the two-step recognition processing of major classification recognition processing and subdivision tQ recognition processing, the computational complexity will be lower than when performing DP matching directly from the beginning. can be significantly reduced, and the recognition rate can be the same or higher.

In the subclassification recognition process in the above two embodiments, each vowel v1 of the vowel series 9 of the unknown input word zeng is used up to the second candidate as shown in the above formula 00, and Distance calculation in DP matching is performed using the same weight for both the first candidate and the second candidate, as shown in equation 04) above. That is, the distance dis(i, j
), the distance is set to 0 when the vowel Vkj of the vowel sequence Vk of the word candidate W, is equal to the first candidate νi or the second candidate 91'°.

On the other hand, in the second embodiment, when calculating the distance #dis(i,j) of the above-mentioned formula 04, if the vowel VkJ of the vowel sequence Vk of the word candidate Wk is equal to the first candidate v1'', , and the second candidate 9.'' are given different distances. That is, if it is equal to the first candidate ``5'', the distance is set to O, while if it is equal to the second candidate vi'', the distance is given as a value between 0 and 1.

This gives greater weight to the first candidate, making it possible to perform more appropriate subclassification recognition processing.

〔Effect of the invention〕

According to the present invention, it is only necessary to calculate the degree of similarity between unknown input word sounds and learning word sounds in the large classification recognition process using representative features of each class for each part. Compared to calculating the similarity between the two, the amount of calculation can be significantly reduced.

In addition, the subclassification recognition process further narrows down the ranking of the large classification word candidates recognized in the large molecule 14 recognition process, and finally it is possible to obtain subclassification word candidates that are accurately ranked. In this case, the number of major classification word candidates is narrowed down, and the number of vowel sequences per word candidate is not so large, so the vowel series of each major classification word candidate and the vowel sequence of the unknown input word sound are DP matching can be performed at high speed.

Particularly in this case, the vowel sequence of the unknown input word speech output from the vowel sequence am means is outputted as a plurality of vowel sequence candidates, and the subclassification recognition means performs DP matching with the vowel sequence of each major classification word candidate. By performing this on the plurality of vowel sequence candidates, the influence of sound recognition errors in the vowel sequence recognition means can be minimized.

Furthermore, in this case, by adding ranks to the plurality of vowel sequence candidates of the unknown input word speech and adding weights corresponding to the ranks during DP matching, more appropriate matching can be performed.

As described above, in the present invention, when performing word speech recognition based on word candidates, first, rough classification word candidates are obtained through major classification recognition processing, and then major classification word candidates are further obtained through detailed classification recognition processing. The two-step method of narrowing down and obtaining subclassified word candidates makes it possible to calculate word candidates quickly and accurately. Then, by performing final recognition on the subclassified word candidates whose number of candidates has been narrowed down in this way, the amount of calculation is significantly reduced compared to a recognition method that performs DP matching etc. from the beginning ( ,and,
It is possible to realize a word speech recognition system with an equal or higher recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the present invention. FIG. 2 is an overall configuration diagram of an embodiment of the present invention. FIG. 3 is a diagram showing an operation flowchart of the preprocessing section. An explanatory diagram of the automatic extraction process. Figure 5 is a diagram showing the time-series pattern of feature vectors of word sounds. Figure 6 is an explanatory diagram of the partial pattern extraction process. Figure 7 is a standard partial pattern creation process. FIG. 8 is a diagram showing a word dictionary. FIG. 9 is an explanatory diagram of an example of calculating the distance between an unknown input word and a dictionary word. FIG. 10 is an explanation of the operation of creating a distance table between partial patterns. Figure 11 is an explanatory diagram. 101... 102... 103... 104... 105... 106... 107... 108... 109... 110... 111... 112... 113. ... Feature extracting means for calculating the cumulative distance between vowel sequences, classification means, representative feature calculation means, dictionary storage means, similarity calculation means, major classification recognition means, vowel sequence recognition means, vowel sequence dictionary storage means, subclassification recognition means, learning word speech, unknown input word speech, major classification word candidates, and subclassification word candidates.

Claims (1)

[Scope of Claims] 1) Feature extracting means (101) for extracting features from each of a plurality of parts of input word speech; After extracting the feature amount for each part in step 101), a classification means (102) classifies the plurality of learning word sounds (110) for each part, and calculates a representative feature amount for each class. Representative feature calculation means (10
3); dictionary storage means (104) for storing representative feature amounts for each class and types of the learning word sounds (110) included in each class as a dictionary for each part; and unknown input word sounds (104); 111), the feature extracting means (101)
), after extracting the feature amount for each of the portions, similarity calculating means () for calculating the similarity for each portion with the representative feature amount of the corresponding class in the dictionary storage means (104);
105), and calculates the similarity between each of the learning word sounds (110) stored in the dictionary storage means (104) and the unknown input word sounds (111) based on the similarity, and calculates the similarity. Ranked major classification word candidates (11
2); a vowel sequence recognition means (107) that recognizes the vowel sequence of the unknown input word speech (111); and an intra-word vowel sequence of each learning word speech (110). a vowel sequence dictionary storage means (108) for storing as a dictionary; and a vowel sequence of each major classification word candidate (112) is extracted from the means, and the unknown input word is outputted from the vowel sequence recognition means (107). It is characterized by having a subclassification recognition means (109) that performs DP matching with the vowel series of the speech (111) and outputs them as ranked subclassification word candidates (113) in descending order of the degree of matching. Word speech recognition method. 2) The vowel sequence of the unknown input word speech (111) output from the vowel sequence recognition means (107) is a plurality of vowel sequence candidates, and the vowel sequence is outputted from the subclassification recognition means (109).
2. The word speech recognition method according to claim 1, wherein the DP matching with the vowel series of each major classification word candidate (112) is performed on the plurality of vowel series candidates. 3) The vowel sequence of the unknown input word speech (111) output from the vowel sequence recognition means (107) is a plurality of vowel sequence candidates that are ranked and output, and the subclassification recognition means (109) Word speech recognition according to claim 1, characterized in that the DP matching with the vowel series of the major classification word candidates (112) is performed by adding weights to the plurality of vowel series candidates according to their ranks. method.