JPS62209497A - Continuous voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a technique for improving the degree of recognition in a continuous speech recognition device.

(Prior Art) Speech recognition is effective as a means of inputting data and commands to a computer, and many recognition devices have been on the market. Among them, continuous speech recognition, which can recognize continuously spoken words, is more effective than discrete speech recognition.
It has the advantage of reducing the burden on the user and allowing high-speed input. Many recognition devices employ a recognition method based on a pattern matching method. In the pattern matching method, one or more tlA quasi-speech patterns of the word to be recognized are prepared, and the most similar standard is selected by comparing unknown utterances, that is, input patterns, and those standard patterns. The recognition results are word names belonging to the pattern category. At that time, the patterns are compared to measure the degree of orchid similarity (hereinafter referred to as distance).
Conventionally, a nonlinear matching method using dynamic programming has been used to find
Publication No. 8278).

The principle of this method will be briefly explained below. A speech pattern is generally expressed as a time series of feature vectors, A”al t al v a3 t ””* a
l , ...aI Here, al ” (aliy alit ”” t aLl) is a vector indicating the characteristics of the audio at time 1, and for example, one obtained by frequency analysis of the L channel can be used. This pattern and the following B"'bl# J* b3ν""t l)j, b
Consider the case of comparing with J. Set the time axis of pattern A to 1
, the time axis of pattern B is j, then time normalization of both patterns is equivalent to finding a distortion function j=,1(t) that maps i to j. That is, finding a distortion function that minimizes the distance between patterns A and B that have undergone mapping is to perform optimal time normalized matching of both patterns. The distance 1i111D at that time can be defined as follows.

From (A, B) = min (Σd(1, j)] j town (
i) i=1 a(tej) is the distance between the vectors l'ial and bJ, and the distance D(A, B) can be obtained by the following recurrence formula.

Here, this recurrence formula can be interpreted as providing flexibility in matching. That is, in the case of this recursion formula, as shown in FIG. 2, the input pattern can be expanded or contracted by % to twice as much as the standard pattern.

Continuous speech recognition using this method can be considered as follows. First, say the recognition word search in advance,
The pattern is the standard pattern ElseB1. ...,B
N, and the input voice pattern is held as input pattern A. Next, %Bt-klta is arranged to create a standard pattern sequence of digits n(1) n(2) n(3)
n(k)B ■B ＜+) B Cf
)...(Create 9B and find the IA similarity with A. Select the single chain sequence (n (IL n(''),
~Hn (k)) is the result to be obtained. As an algorithm for such continuous speech recognition, Japanese Patent Application Laid-open No. 58-
No. 98796 and Acoustical Society of Japan Research Group Material 881-6
5 and can be used.

(Problems to be Solved by the Invention) In the conventional continuous speech recognition device as explained above, the standard pattern is a pattern that retains the audio information of discrete words uttered by the user or other speakers as is. was used. In other words, the characteristics of the utterance length were maintained as they were.

Generally speaking, it is known that the average utterance length per word decreases when words are uttered in consecutive words, but in reality, all words do not become shorter at the same rate, but when uttered in discrete utterances, the utterance length decreases. The longer the length, the greater the rate of shortening. Furthermore, there is a tendency that the faster the speaking speed becomes, the more the average length of time per word becomes constant. In other words, in continuous vocalization,
Because each word is influenced by the words before and after it, the length of each word becomes equal. This phenomenon is seen when the combination of words such as numbers has no meaning, and in such cases, humans
This is thought to be due to the fact that in order to make vocalization easier, they begin to vocalize at a rhythmic interval.

Therefore, if continuous word recognition is performed using utterance length information as a standard pattern as in the conventional method, long standard patterns will always be compressed and matched. This means that only part of the expandable range is being used.

FIG. 3 shows an example of this matching. Continuous vocalization @381' and standard pattern @31°fi8#
, @l# and the matching pattern. The acclimatization formula described above is used. As shown in the shaded area in the figure, the range of expansion and contraction allowed by the above-mentioned recurrence formula for Mi "8" and "1" is longer than that for 'a3' due to the long duration of discrete utterances. It becomes smaller.

Therefore, equation (3) in the acclimatization equation is placed in the S-string, and equation (1) is hardly used. As described above, in the conventional method, the flexibility of matching is not fully utilized, and the accuracy of recognition may deteriorate. Furthermore, if the difference between the utterance length of a discrete single chain and the utterance length of a continuous word is large, the expansion/contraction range allowed during matching may be exceeded, and in such cases, correct matching may not be performed. Certification shellas occur frequently.

To solve this problem, a method was proposed in a patent application filed in 1983 that increases the flexibility of matching and performs acclimatization formula calculations for parts with large expansion and contraction when there is a difference in utterance length between discrete utterances and continuous utterances. -245328. This method changes the flexibility of matching by calculating acclimatization equations for a plurality of different degrees of expansion and contraction, as shown in FIGS. 4(a) to (C). As shown in the figure, when the flexibility of matching is increased, the sum gt of the recurrence formula increases proportionally. That is, this method has the disadvantage that the amount of calculation and memory required during recognition increases significantly.

SUMMARY OF THE INVENTION An object of the present invention is to eliminate the above-mentioned drawbacks and provide a continuous speech recognition device with good performance without making unreasonable time axis correspondences to all words in a recognition vocabulary collection.

(Means for solving problems) The continuous speech recognition device according to the present invention consists of several parts.

That is, an input pattern storage section that holds the time series of feature vectors of the input speech as an input pattern, a -11 training pattern storage section that holds the time series of the feature vectors of the W11 training speech as a training pattern, and the training pattern storage section. an equalization processing unit that performs time-axis compression/expansion processing on the stored training patterns; a standard pattern storage unit that stores the patterns obtained by the equalization processing unit as quasi-patterns; and the input pattern storage unit. These are each part of the recognition processing section that performs recognition processing based on the input pattern read out from the standard pattern storage section and the ta quasi-pattern read out from the standard pattern storage section.

(Operation) The present invention increases the degree of freedom in matching each word and improves the accuracy of time length normalization by expanding or contracting the time length of a standard pattern in advance. This can be realized by providing the equalization processing section in the present invention. That is, the equalization processing section performs a time length expansion/contraction process on the uttered discrete training patterns so that the time length becomes the same as when the uttered discrete training patterns are consecutively uttered. By using these as standard patterns and performing recognition by pattern matching, highly accurate recognition can be achieved.

(Example) Hereinafter, implementation of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the continuous speech recognition device of the present invention.

Here, a case will be described in which the numbers "zero" to "kyu" are recognized when they are consecutively uttered.

The -11 training pattern storage unit 1 stores in advance feature patterns of numeric sounds uttered by the user or other speakers. Here, each i and l11 practice pattern will be expressed as follows.

Efl"bl y bt v bB t...t b,
i,,-...,bJn, where n= it 2r ・”
*N bt is the characteristic parameter at time t of each -11 practice pattern, and Jn is -11 practice pattern B
The time length is n. The patterns stored in the training pattern storage section are those that are retained as they were uttered, and the time length j is relatively small for patterns such as % @ni 't@go #, and "1chi -, -roku'". ,
-hachi” has a long time length.

In particular, the equalization processing section 2 reads the training pattern from the training pattern storage section 1, and performs an elongation process on the pattern so that the time length is equalized. This equalization process is performed, for example, as follows.

First, the average time length of one word in continuous utterance is calculated using the following formula.

y is a coefficient indicating the expansion/contraction ratio of continuous utterance length to one discrete utterance length, and is set to, for example, 0.8 here.

111# on the arrow! Each word of the pattern is expanded or contracted so that it becomes the determined time length J. That is, standard pattern Cn = C1t C1t...t C1t
”'t Characteristics of CJX at time 1) /L/
Ct can be determined by the following equation.

C1=br*r=(tXJn/, yx+ 1) However, [2] is the maximum integer value not exceeding 2.

Pattern Cnn obtained by the stretching process described above=
1*2w=·e N t'fA is stored in the standard pattern storage unit 3 as a quasi-pattern.

At the time of recognition, the input pattern is stored in the input pattern storage section 4. Subsequently, the recognition processing section 5 reads the input pattern from the input pattern storage section 40 and the standard pattern from the standard pattern storage section 3, performs recognition processing using a pattern matching method, and obtains the results. The processing in the recognition processing section 5 is based on the above-mentioned
796 can be used.

The operation of the continuous speech recognition device according to the present invention has been described above, but the expansion/contraction process is not limited to the above process, and the following can also be used.

et,'=(t-α)×br+α+1r=Ct X
Jn/JX) α=ixJn/JX-r In addition to the above linear expansion/contraction, dynamic programming and
>(11 Non-linear expansion and contraction can also be performed based on changes in the vector.Furthermore, in the example, a word-based standard pattern was used, and the standard pattern length was set to TX for each category, but the standard pattern As, phoneme unit,
Patterns such as syllable units, VCV units, half word units, etc. may also be used, and the standard pattern length may be determined for each category.

(Effects of the Invention) In the continuous speech recognition device of the present invention described above, even if the -11 training pattern is a discrete utterance, a standard pattern having a time length in continuous utterance is prepared.

Therefore, as in the past, the difference in time length between the continuous utterance pattern and the quasi-a pattern is not large, and this solves the drawback that a large time length difference prevents correct matching and causes an error. Effortless matching of patterns becomes possible. As a result, a more accurate degree of similarity can be determined, and recognition accuracy is improved. Also,
Conventionally, the only way to widen the matching tolerance was to use an acclimation formula, which had the drawback of significantly increasing both the amount of calculation and the amount of memory; however, according to the present invention, the total amount of n The matching tolerance can be expanded without increasing the amount of data and memory. Furthermore, in order to prevent an increase in the amount of calculation and memory, a method has been proposed in which the recursion formula used is changed depending on the degree of expansion/contraction, but in the method according to the present invention, only one type of recursion formula is used. . Therefore, since the processing at the time of recognition is simplified, there is an advantage that the recognition circuit is simplified.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing the range of matching flexibility in the conventional method, FIG. 5g3 is a diagram showing the state of matching in the conventional method, and FIG. 4(a) to 4(C) are diagrams for explaining a method of changing the flexibility of matching in the conventional method. In the figure, 1...-11 training pattern storage section, 2... uniformity processing section, 3... standard pattern storage section, 4... input pattern storage section, 5... recognition processing section . W j input/\turn '381'' Figure 3 ('+-1.j-2) Figure 4 (1)) Figure 4 (c)

Claims (1)

[Claims] an input pattern storage section that holds a time series of feature vectors of input speech as an input pattern; a training pattern storage section that holds a time series of feature vectors of a training speech as a training pattern; a standard pattern storage section that stores the pattern obtained by the uniformization processing section as a standard pattern; and a standard pattern storage section that stores the pattern obtained by the uniformization processing section as a standard pattern. 1. A continuous speech recognition device comprising: a recognition processing unit that performs recognition processing based on an input pattern read from the standard pattern storage unit and a standard pattern read from the standard pattern storage unit.