JPH09237097A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To avoid badness of the sharing property of leaning data by providing the standard pattern expressing the center part of a word and the standard pattern expressing the connection part of a word and connecting them alternately. SOLUTION: An inputted voice is converted into an electric signal in a voice input means 1. The voice converted into the electric signal is further analyzed in a voice analyzing means 2 and then time series of feature vectors are outputted. The time series of the feature vectors of the input voice are collated with standard patterns preliminarily stored in a standard pattern storage means 5 in a collating means 3 and then scores are calculated every word of recognition objects. In this case, the standard pattern expressing the center part of a word and the standard pattern expressing the connection part of a word are provided as standard patterns and the standard patter of a word string is constituted by connecting them alternately. Moreover, collations are performed under the control of a finite state automaton 6. Then, recognized results are outputted based on the scores of respective words in a decision means 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous number voice recognition apparatus for recognizing numbers spoken by voice.

[0002]

2. Description of the Related Art Conventionally, as a unit of a standard pattern in a continuous numeral voice recognition device, a standard pattern of word unit is generally used. Recognize arbitrary sentences in Japanese and recognize large vocabulary words (for example, all names of Japanese people)
In a voice recognition device for which the target pattern is, it is practically impossible to prepare standard patterns for each word because the number of standard patterns becomes too large. However, 1
In a small vocabulary number voice recognition device that recognizes only 0 numbers, the number of standard patterns is only about 10 even if the standard pattern of word units (single digit numbers) is used.

When a standard pattern is prepared for each word, the articulatory combination within a word is incorporated into the standard pattern, and the articulatory combination within a word can be dealt with. However, it cannot deal with articulatory coupling that occurs between words.

As a method for coordinating articulatory coupling between words, there is a Japanese Society for Acoustics, Speech Study Group material, Material No. S84-64, "Continuous digit speech recognition using half-word pair standard pattern". There is an example as described in. In this example, the range from the central portion of the preceding numeral to the central portion of the succeeding numeral is set as the unit of the standard pattern. In this way, since the part between the numbers is set as the standard pattern, the articulatory coupling generated between the words is taken into the standard pattern, and the articulatory coupling between the words can be dealt with. In this method, since a standard pattern is prepared for each word pair, the number of standard patterns is 100, that is, the number of combinations of two-digit numbers.

As another conventional example corresponding to articulatory coupling generated between words, as a technical research report of the Institute of Electronics, Information and Communication Engineers,
There is an example as described in SP95-23 “Examination of acoustic model learning method in continuous numeral speech recognition”. In this example,
A word is divided into three parts, head, body, and tail, and the head part prepares a different model for each preceding word. As well as tail
A different model is prepared for each word that follows the part. head
The articulatory coupling that occurs between the word preceding the portion and the word that follows the tail portion is captured, and the articulatory coupling between the words can be accommodated. In the case of numbers, there are 10 types of body parts, 10 types of head parts and 10 types of tail parts for each body, so the total number of standard patterns is 210.

[0006]

According to the above-mentioned two conventional techniques, a standard pattern corresponding to articulatory coupling between words can be prepared, and the recognition accuracy for continuous numbers is improved. However, the first conventional example has a problem that the standard pattern can be learned only in the combination of two-digit numbers, and the sharing of learning data is poor. Also, in the second conventional example, there is a problem that the sharing of the learning data is good in the body part, but the sharing of the learning data is poor in the head part and the tail part as in the first conventional example. Further, in the second conventional example, there is a problem that learning data that satisfies all combinations of the head part, the body part, and the tail part is required, and 1000 kinds of learning data are required.

It is an object of the present invention to provide means for avoiding the poor sharing of learning data, which is a problem in the prior art, while making use of the ability to cope with articulatory coupling between words in the prior art.

[0008]

The object of the present invention is to provide a standard pattern representing a central portion of a word and a standard pattern representing a connecting portion of words as standard patterns, and connecting these alternately to form a word. This is achieved by making a standard pattern of rows.

In the case of numbers, there are only a total of 10 standard patterns that represent the central portion of a word, and the sharing of learning data can be improved. There are a total of 100 standard patterns that represent connected parts of words in the case of numbers, and although the sharing of learning data is a little poor, it is possible to deal with articulatory coupling between words. As described above, it is possible to create a standard pattern that can support articulatory coupling between words while maintaining high sharing of learning data, and realize highly accurate voice recognition.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the continuous numeral voice recognition apparatus of the present invention. The input voice is converted into an electric signal by the voice input unit 1. The voice converted into the electric signal is further analyzed by the voice analysis means 2, and a time series of feature vectors is output. The standard pattern stored in advance in the standard pattern storage unit 5 and the feature vector time series of the input voice are collated by the collation unit 3, and a score is obtained for each word to be recognized. The matching is performed under the control of the finite state automaton 6. The determining means 4 outputs a recognition result based on the score of each word.

Next, the standard pattern used in the present invention will be described. In the present invention, a probabilistic model is adopted as the standard pattern. FIG. 2 is a diagram showing a stochastic model (Hidden Markov Model, hereinafter abbreviated as HMM) used in the present invention. In the figure, each circle represents a state, and arrows represent transitions between the states. The symbol aij attached to the arrow represents the probability of transition from the state i to the state j, and the symbol bij (k)
Represents the probability that a feature vector belonging to the kth classification will be output when a transition from state i to state j occurs.
Given the feature vector time series of the input voice, the probability that the feature vector time series of the input voice is output from this probability model (HMM) can be calculated using the state transition probability and the output probability. The matching means 3 in FIG. 1 performs this probability calculation process. For details of the probability calculation process, see Kluwer Academic Publishers, Norwe
l, MA, 1989 “Automatic Speech Recognition”, 95
Known methods described on page-97 may be used.

Further, referring to FIG. 3, a standard pattern for continuous numeral voice recognition, which is the main feature of the present invention, will be described. In FIG. 3, the two-digit number "12" is formed by alternately connecting the standard pattern of the word center part and the standard pattern of the word connection part. In the figure, * indicates silence at the beginning and end of a word. In the example of FIG. 3, the standard pattern in the center of the word has an HMM of 5 states.
Is assigned to the standard pattern of the word connection part,
M is assigned. In the figure, 31 is a standard pattern of a connecting portion between the silence and the numeral "1", 32 is a standard pattern of a central portion of the numeral "1", 33 is a numeral "1" and a numeral "2".
, 34 is a standard pattern of the central portion of the numeral "2", 34 is a standard pattern of the central portion of the numeral "2", and 35 is a standard pattern of the joint between the numeral "2" and the ending silence. There are a total of 10 standard patterns in the center of the word, and 120 standard patterns in the word connecting part (more than 100 due to the connection with silence at the beginning of the word). It goes without saying that the following rules are followed in the connection of the standard pattern of the word center part and the standard pattern of the word connection part. That is, in constructing the standard pattern of consecutive numbers XY, the standard pattern XY of the word connecting portion is connected between the standard pattern of the X word central portion and the standard pattern of the Y word central portion.

Next, the finite state automaton used in the control of the matching unit will be described. FIG. 4 is a finite state automaton that expresses an arbitrary number sequence between 1 and 10 digits. In the figure, the circles indicate the respective states, and the states are connected by arrows (arcs). When the recognition is started, the state 1 is entered first. Each arc between states represents a one-digit number and recognizes a one-digit number each time a state transition occurs. The finite state automaton in FIG. 4 is composed of 11 states in total. It starts from state 1 and eventually reaches state 0 through any of the other states and ends. In the case of a one-digit number, the state 1 is changed to the state 0, and the recognition is ended. In the case of a two-digit number, the state 1 transits to the state 2 and the state 0 to end the recognition. State 1 for N-digit numbers
To state 2, state 3, ..., State N, state 0, and the recognition is ended. Actually, since the number of digits of the number to be input is not known in advance, all of these possibilities are evaluated and the one giving the highest probability is the recognition result. FIG.
Each arc in the table corresponds to a one-digit number, but when the standard pattern of the present invention is used, a standard pattern for the central part of the word is assigned to the part of this arc, and for each word connection part in each state. Using the standard pattern of, the arcs (words) that come into this state and the arcs (words) that leave this state are connected. It goes without saying that this connection follows the rules described above. FIG.
Then, each arc represented a one-digit number, but the sign (+,
By constructing a finite state automaton using an arc representing −) or a decimal point (.), It is possible to easily represent a real number including a part below the decimal point or a negative number.

Next, a normal learning method of the HMM which is a standard pattern used in the speech recognition apparatus of the present invention will be described. The HMM is implemented by parameter estimation using a large number of training speech samples. FIG. 5 is a flowchart showing the outline of the learning flow. First
Create an initial model of HMM by some method (10
1) and then the parameter re-estimation process (102) using the learning voice sample is repeated (103) until the convergence condition is satisfied. This learning method is originally an iterative estimation algorithm, and the accuracy of the model improves as the number of iterations increases. Therefore, it is not always necessary to create the initial model with high accuracy. There are several methods for creating the initial model, but a method such as giving a random number may be used. The method of parameter re-estimation will be described later. There are several possible methods for determining the convergence condition, but for example, the number of repetitions is fixed and a fixed number of times (for example, 5 times).
There is no problem in practice by the method of ending after repeating.

When the convergence condition is satisfied, the iteration is terminated and the parameters of each HMM obtained by the parameter estimation are stored (104).

Next, the parameter re-estimation processing of the HMM will be described. As shown in the flowchart of FIG. 5, the HMM parameter re-estimation process is repeatedly performed in the learning flow. Here, the one-time processing will be described with reference to the flowchart of FIG. The HMM parameter re-estimation process is performed using speech samples for learning. Assuming that the number of training voice samples is N, the parameter estimation calculation processing similar to N times is performed, and after this processing is completed, the parameters of each HMM are updated to new values. In the parameter estimation process using each voice sample, first, the HMMs of the basic recognition units are concatenated according to the utterance content of the voice sample (203), and a method called the Forward-Backward algorithm is used for this concatenated HMM. Parameter estimation is performed (204).
By decomposing the concatenated HMMs into the original recognition basic units, the parameter estimation value of the HMM of each recognition basic unit is obtained (205). However, at this point, the HM of each recognition basic unit
The parameters of M are not updated, and after the parameter estimates are obtained for all speech samples, the parameter estimates of all recognition parameters obtained so far are combined to update the parameters of the HMM of each recognition basic unit (207). . For the specific calculation procedure of parameter estimation (Forward-Backward algorithm), see Kluwer Academic Publishers, Norwel,
MA, 1989 “Automatic Speech Recognition”, page 95-9
The known method described on page 7 may be used.

[0018]

As described above, according to the present invention, it is possible to create a standard pattern capable of coordinating articulations between words while maintaining high sharing of learning data, and to realize highly accurate continuous digit speech recognition.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a continuous numeral voice recognition device of the present invention.

FIG. 2 is a diagram illustrating a hidden Markov model used in the continuous numeral voice recognition device of the present invention.

FIG. 3 is a diagram for explaining a method of connecting a standard pattern of a word center portion and a standard pattern of a word connection portion used in the continuous numeral voice recognition device of the present invention.

FIG. 4 is a finite state automaton that expresses an arbitrary digit string of 1 to 10 digits.

FIG. 5 is a flowchart illustrating a standard pattern learning method of the present invention.

FIG. 6 is a flowchart illustrating a parameter estimation process in the standard pattern learning method of the present invention.

[Explanation of symbols]

1 ... voice input means, 2 ... voice analysis means, 3 ...
Collation means, 4 determination means, 5 standard pattern storage means, 6 finite state automaton 101, initial model creation processing, 102, parameter re-estimation processing, 204 forward-backward algorithm.

Claims (7)

[Claims] 1. A voice input means for inputting continuous numeric voice,
A voice analysis unit that analyzes the input voice and outputs a time series of feature vectors, a standard pattern storage unit that stores a standard pattern that serves as a reference for recognizing continuous numeric voice, and a time when the feature vector is used. In a voice recognition device comprising a collating means for collating a series and the standard pattern, and recognizing based on a collating result in the collating means, the standard pattern storing means includes a standard pattern corresponding to a central portion of each numeral and A standard pattern corresponding to a connecting portion of two consecutive numbers is stored, and the collating means is a standard in which the standard pattern of the central portion of the numeral and the standard pattern of the connecting portion of the numeral and the numeral are alternately connected. A voice recognition device, which outputs a recognition result by comparing a pattern with a time series of the feature vector of the input voice. 2. The continuous numeral voice recognition apparatus according to claim 1, wherein the standard pattern is constituted by an established model based on the stochastic model learned based on a learning voice sample. 3. The continuous digit speech recognition apparatus according to claim 2, wherein the probabilistic model is a hidden Markov model. 4. The continuous numeral voice recognition apparatus according to claim 1, wherein the matching means is controlled by a finite state automaton. 5. The continuous numeral voice recognition apparatus according to claim 4, wherein the finite state automaton is configured to limit the number of digits of an input numeral. 6. The continuous numeral voice recognition device according to claim 4, wherein the finite state automaton is configured to limit a range of numerical values of inputted numerals. 7. The standard pattern includes "+ (plus)" and "-".
A sign of "(minus)" and a standard pattern of decimal point ". (Ten)" are provided, and the finite state automaton is configured to limit an input number as a real number including a part after the decimal point. 5. The method according to claim 4, wherein
The described continuous numeral voice recognition device.