JPH06118989A - Continuous speech recognizing method - Google Patents

Abstract

PURPOSE:To recognize continuous vocalization including an unregistered word by recognizing a continuous speech which includes the unregistered word by using unregistered word patterns generated by using speeches other than regis tered words as learning data. CONSTITUTION:A sentence head segmenting process part 8 divides a sentence- vocalized input equally into three and segments the 1st part as a sentence head part, an in-sentence segmenting process part 9 segments the section other than the sentence head part and sentence tail part, and a sentence tail segmenting process part 10 divides the input equally into three and segments the final part as the sentence tail part. The results of the respective sectioning processes are sent to a standard pattern generating process part 3 as well as registered words to generate three kind of unregistered word patterns, which are sent to an unregistered word standard pattern storage part 11. The unregistered word standard pattern storage part 11 stores the three unregistered word standard patterns of the sentence head part, center part, and tail part. Therefore, a difference with the position such as possible pausing at the head of a sentence can be reflected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention
The present invention relates to a method for recognizing continuously uttered voice in a machine interface.

[0002] A voice dialogue system that not only enables input and output by voice but also transmits complex information through human-like dialogues without waste is an advanced voice model method supported by speeding up of computers, for example. It is about to be put to practical use by introducing it.

In the process of developing such a system, a voice input method for extracting sufficient information from the user's utterance is required.

[0004]

2. Description of the Related Art In FIG. 1, a Markov model is used as a standard pattern, and a probability value that an input speech is output from the Markov model is used as a matching process, and a regular grammar (finite state automaton) is used as a syntax constraint. It is a block diagram in the case of implementing the method ("Speech recognition by an established model" by Seiichi Nakagawa) by hardware.

This method is realized by matching the input voice with the standard pattern for each arbitrary word string satisfying the given grammatical constraint, and advancing the voice recognition and the syntactic analysis at the same time.

For that purpose, first, the grammatical constraint storing unit 2 stores the grammatical constraint in which the position where a registered word can appear in a sentence is described in the form of a finite state automaton.

Next, in order to create a standard pattern for each registered word as learning data, the input voice corresponding to the registered word is analyzed by the acoustic analysis processing unit 1, and the result 1a (during learning) is created as a standard pattern. Send to the processing unit 3. The standard pattern 3a created by sequential processing is stored in the standard pattern storage unit 5.

In the word standard pattern connection processing unit 4, the word standard patterns stored in the word standard pattern storage unit 5 are connected according to the grammatical constraints stored in the grammar constraint storage unit 2,
A process of creating a sentence standard pattern is performed, and the sentence standard pattern is stored in the sentence standard pattern storage unit 6.

On the other hand, at the time of recognition, when a sentence utterance to be recognized by the user is input, the acoustic analysis result 1c (at the time of recognition) is sent to the pattern matching processing unit 7 and stored in the sentence standard pattern storage unit 6. All sentence standard patterns and pattern matching processing unit 7
Collated in.

Here, as the matching distance, the value of the probability that the acoustic analysis result is output from the Markov model which is the standard pattern is used. The matching distance is a numerical value indicating the similarity between the input voice and the standard pattern.

Therefore, the word string for the sentence standard pattern giving the best matching result is output as the recognition result.
This enables a system that recognizes thousands of large vocabulary words in a continuous utterance of an unspecified speaker.

[0012]

In such a conventional method, since the recognition is performed by collating the standard pattern of the registered word with the input, the user needs to speak using only the registered word.

Further, when stagnation or erroneous words occur during voice input, it is difficult to obtain a correct recognition result because there is no standard pattern that is correctly collated with the stagnation and erroneous words.

The malfunction of the continuous speech recognition apparatus for a speech input including these unregistered words and erroneous words is a problem to be solved by the present invention.

[0015]

SUMMARY OF THE INVENTION The present invention continuously recognizes continuously uttered voices by continuously collating a word standard pattern represented by a Markov model with an input voice according to grammatical constraints. In the speech recognition method, unregistered words created by defining the appearance positions of unregistered word patterns that may appear in continuous speech in grammatical constraints and using speech other than registered words as learning data. The feature is that continuous speech including unregistered words that are not registered is recognized using a pattern.

Further, in this recognition method, three types of patterns, that is, the beginning of a sentence, the middle of a sentence, and the end of a sentence, are created as unregistered word patterns, and the patterns are compared with unregistered words.

[0017]

FIG. 2 is a diagram showing a case in which a voice other than a registered word is recognized as an unregistered word by using an unregistered word pattern created by using a voice other than a registered word as learning data, and a matching process with an input voice is performed. An example of the present invention will be shown.

A grammar constraint storage unit 2, a word standard pattern processing unit 4, a standard pattern creation processing unit 3, a word standard pattern storage unit 5, a sentence standard pattern storage unit 6, a pattern matching processing unit 7 are provided.
Performs the same processing as the conventional method.

The sentence head cutout processing unit 8, the sentence cutout processing unit 9, the sentence end cutout processing unit 10, and the unregistered word standard pattern storage unit 11 are newly added.

For the analysis result 1a of the voice input for creating the word standard pattern, a word standard pattern is created for each registered word and stored in the word standard pattern storage unit, as in the conventional method.

Next, at the time of learning the unregistered word pattern, the analysis result 1b of the voice input is divided into three sections, that is, the beginning of a sentence, the middle of a sentence, and the end of a sentence by the sentence head cutout processing unit 8, the sentence cutout processing unit 9, and the sentence end cutout processing unit 10. Be transformed.

That is, the sentence head cut-out processing unit 8 divides a sentence-uttered input into three equal parts, and cuts out the first one-third portion as the sentence head. The in-sentence cutout processing unit 9 performs a process of cutting out a section other than the sentence head and the sentence end from the sentence utterance. The sentence end cutout processing unit 10 divides the input from the sentence utterance into three equal parts, and cuts out the last one third as a sentence end part.

The results of each segmentation process are sent to the standard pattern creation processing unit 3 as in the case of registered words, three types of unregistered word patterns are created, and sent to the unregistered word standard pattern storage unit 11. .

The unregistered word standard pattern storage unit 11 stores three unregistered word standard patterns of the beginning of a sentence, the middle of a sentence, and the end of a sentence.

As described above, the learning voice of the unregistered word pattern is divided into the three types of the beginning, the sentence, and the end of the sentence, and the standard pattern is created for each of them. For example, at the beginning of the sentence, "Ano" and "Eh" It is possible to reflect the difference depending on the position such as is likely to appear.

Although the use of the grammar constraint for connecting the word standard patterns is similar to the conventional example, in the practice of the present invention, the appearance position of the unregistered word is shown in the grammar constraint as shown in FIG.

Next, as the input voice, "Well, on the right, 90
Turn around. The difference between the present embodiment and the conventional method will be described by taking the case where the utterance "" is input as an example.

In the conventional method, since the standard patterns corresponding to "um" and "(turn)" are not prepared, there is a high possibility that an incorrect recognition result will be output.

In the present embodiment, in the state 0, the "um" portion is matched with the unregistered word pattern at the beginning of the sentence with high probability.

In the state 9, "(turn) yo" is matched with the unregistered word pattern at the end of the sentence with high probability.

As shown in this embodiment, the present invention makes it possible to correctly recognize a continuous voice in which an unregistered word such as stagnation is inserted during utterance.

[0032]

According to the present invention, it is possible to realize a voice recognition device which recognizes not only voice input using registered words but also continuous speech including unregistered words similar to human beings.

Further, it is possible to improve the recognition rate, limit the utterances to the user, and receive free utterances, and it is possible to expand the range of use.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing a case where a conventional method is implemented by hardware.

FIG. 2 is a configuration diagram showing a case where an embodiment of the present invention is implemented by hardware.

FIG. 3 is a diagram showing grammatical constraints given by a finite state automaton used in an embodiment of the present invention.

[Explanation of symbols]

 1 Acoustic analysis processing section 2 Grammar constraint storage section 3 Standard pattern creation processing section 4 Word standard pattern concatenation processing section 5 Word standard pattern storage section 6 Sentence standard pattern storage section 7 Pattern matching processing section 8 Sentence extraction processing section 9 Sentence extraction processing section 10 Sentence cutout processing unit 11 Non-registered word standard pattern storage unit

Claims (2)

[Claims] 1. A continuous speech recognition method for recognizing a continuously uttered speech by continuously matching a word standard pattern represented by a Markov model with an input speech according to grammatical constraints. The feature is that the position of the unregistered word pattern that may appear in the inside is defined in the grammatical constraint, and continuous speech including unregistered words that are not registered is recognized using the unregistered word pattern. And continuous speech recognition method. 2. The continuous speech recognition method according to claim 1, wherein three types of patterns, that is, the beginning of a sentence, the middle of a sentence, and the end of a sentence, are created as unregistered word patterns and the unregistered words are compared with each other. .