KR20030002197A - Speech recognition method using posterior distance - Google Patents

Abstract

PURPOSE: A method for recognizing a voice by using a distance after recognition is provided to obtain prompt word voice recognition speed and easily manage dictionaries by separating a dictionary for a viterbi search and a recognition target word dictionary. CONSTITUTION: A method for recognizing a voice by using a distance after recognition includes the steps of initializing data in state of having a word dictionary(ST-301), receiving voice data inputted from a speaker(ST-302), searching a subword sequence through a viterbi search(ST-303,ST-304), calculating distance between the subword sequence and words in a recognition target word dictionary(ST-305), obtaining a final recognition result(ST-306), and judging whether voice data exist to be continuously recognized(ST-307).

Description

Speech recognition method using posterior distance

The present invention relates to a speech recognition method using a distance after recognition, and more particularly, a sequence connected to sub-word speech units and a speech using the distance compared with words in a real word dictionary. It relates to a speech recognition method using the distance after the recognition to recognize.

As is well known, the speech recognition system is a system that derives the result of speech recognition through a recognizer by receiving a human voice, and various services using such recognition technology have been developed and commercialized.

In general, speech recognition is a process of finding a corresponding word sequence when a specific sound data sequence is input. One approach for speech recognition is pattern comparison, in which the most similar reference pattern is produced as a recognition result by comparing input speech data with a reference pattern.

Although the pattern comparison is applied to various places and shows good recognition performance, there is a problem that there is little stored information. That is, since only one or two reference patterns are stored for each word, the corresponding result value cannot be derived by sufficiently comparing one word with various pronunciations. Therefore, we need a way to hold more information.

It is a speech recognition method based on a statistical model, in which the information is not deterministically defined but statistically implied.

More specifically, it is called HMM (hidden Markov model). Assumptions based on the HMM-and various other kinds of statistical models-can well characterize the speech signal as a random process by parameters, and the parameter values can be estimated in an accurate and well-defined way. will be.

The HMM is a set of states connected by a transition, and each transition has two kinds of probability sets. One is the transition probability, which indicates the probability of a transition, and the other is the output probability density function, which indicates the probability that a particular output will be produced.

If an HMM is given to model speech, it is necessary to define a basic unit. Typical examples include words and phonemes. The word model is suitable for small vocabulary recognition, but is not practical for large vocabulary recognition.

The phoneme model, on the other hand, has a strong context dependency but is a standard, well understood and easy to train unit.

1 is an example of an HMM using phonemes. In FIG. 1, notation of probability is omitted. This model recognizes "YES" and "NO", which can be separated by a silent representation of / sil. "YES" consists of '/ YE' and '/ S', and "NO" consists of '/ N' and '/ O'.

When the speech organ generates the speech sample X and sends it out of the mouth, the speech recognition system does not know how the acoustic sequence is performed inside the speaker. The recognition system attempts to find W by reconstructing the most similar state sequence to know what word sequence W generated X. There is a formal way to find the single most desirable sequence, called the Viterbi algorithm.

2 is a flowchart illustrating a conventional speech recognition process.

Referring to this, the speech recognition system initializes data with a word dictionary including information on a word and a model of the words (ST-201), and receives voice data from the speaker ( ST-202).

Then, through the Viterbi search (ST-203), the recognizer searches the word dictionary and outputs the word sequence most similar to the input sound sequence as the recognition result (ST-204).

However, when the speech recognition is performed in the above-described manner, as the number of words to be recognized increases, the time for searching the dictionary also increases, thereby increasing the time required for the recognition.

When recognizing speech, it is necessary to first define the basic unit of recognition.

In detail, first, a word and a phoneme can be exemplified as a unit used for speech recognition. A phoneme is a kind of subword unit and a syllable and a two syllable as a subword unit. And acoustic units.

However, when the word is used as a basic unit, in order to obtain a reliable word model, there are disadvantages in that the number of word utterances required for learning must be large, and a new word model must be generated whenever there is a change in the word.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described prior art, and primarily obtains a sequence connected to subword speech units after speech input, and secondly, a word in the sequence and the actual word dictionary. It is an object of the present invention to provide a speech recognition method using a distance after recognition to improve the recognition speed by outputting a word sequence of the shortest distance as a result through a regressive recognition process that compares the results.

1 is a view showing an example of a HMM (hidden Markov model) of a simple grammar according to a conventional embodiment,

2 is a flowchart illustrating a voice recognition process according to a conventional embodiment;

3 is a flowchart illustrating a speech recognition process using a distance after recognition according to an embodiment of the present invention.

* Description of the symbols for the main parts of the drawings *

ST-304: Sub-Sequence Determination Step ST-305: Recognition Distance Measurement Step

ST-306: Word Sequence Determination Step

In order to achieve the above object, according to a preferred embodiment of the present invention, in the speech recognition system for recognizing the input speech data, the sub-unit speech does not change even if the words are inserted or removed as a word dictionary for speech recognition Unit dictionary; Provided is a speech recognition method using a distance after recognition, wherein the recognition target word dictionary can be changed at the same time according to a change of a word.

Preferably, using the subword phonetic unit dictionary to find a subword sequence; The speech recognition method using the post-recognition distance is provided by measuring a distance after the recognition between the subword sequence and the words in the word dictionary and obtaining a word sequence having the shortest distance as a recognition result.

In more detail, when the sub-unit speech unit is inserted or deleted from the recognized sub-word unit sequence, the distance after the recognition may be obtained by calculating a distance from the target word of the sub-word unit sequence to obtain a final recognition result. Provided is a speech recognition method.

In addition, there is provided a speech recognition method using a distance after recognition, wherein a result sequence is obtained by using phonetic information and information imposed by a grammar in generating the subword sequence.

On the other hand, the distance measurement process after the recognition is provided by the speech recognition method using a distance after recognition, characterized in that by using the stored information to calculate the distance for each phoneme using a phonetic model.

Preferably, the distance measurement process after the recognition is provided using a speech recognition distance using a distance after recognition, characterized in that by using the information calculated by pre-determining the distance value for the phoneme of the additional word.

EMBODIMENT OF THE INVENTION Hereinafter, this invention is demonstrated in detail with reference to drawings.

First, the present invention is to easily handle the word dictionary and to increase the recognition speed in the large vocabulary speech recognition. In the present invention, a plurality of dictionaries are used for the above-described purposes. The first dictionary is composed of subword phonetic units used for Viterbi search, and the second dictionary has information on words to be recognized. The advantage of using a separate dictionary is that it is easy to add or delete words to be recognized regardless of Viterbi search.

Therefore, the recognition operation according to the present invention goes through the following two steps. First, a process of outputting an intermediate recognition result of a subword speech unit sequence through a Viterbi search based on an input speech.

Secondly, the distance between the intermediate result and the words in the recognition word dictionary is calculated and the word sequence in the shortest distance is output as the final recognition result.

3 shows a speech recognition process according to the present invention.

First, some of the processes 301, 302, 303 and 304 in the speech recognition process according to the present invention coincide with the parts 201, 202, 203 and 204 in FIG. However, the first dictionary used in the present invention and the dictionary in FIG. 2 are different from each other.

Hereinafter, a speech recognition process according to the present invention will be described.

For example, suppose that a recognition engine for recognizing "father" and "mother" is made. Here, sub-unit speech units are assumed to be phonemes. When following the conventional method, the word dictionary can be constructed as follows.

Father: ㅏ ㅂ ㅓ ㅣ

Mother: ㅓ ㅁ ㅓ ㄴ ㅣ

Using a predefined model, the recognition engine calculates the probability that each word sequence generated an acoustic sequence, and outputs the word with the highest probability as the recognition result.

Two dictionaries are used in the present invention. The first dictionary is a subword unit dictionary, which has phoneme information, and is composed as follows.

ㅏ

ㅂ

ㅓ

ㅈ

ㅣ

M

N

This dictionary should include a variety of information about each model, but is omitted because it is not the gist of the present invention.

The second dictionary is a recognized word dictionary composed of phonemes from the first dictionary. In this example, the conventional word dictionary described above with reference to FIG. 2 is used as it is.

In the first step, the recognizer generates a sequence of phonemes as a result of the Viterbi search (ST-304). For example, if a user says "father", the phoneme sequence is <'ㅏ', 'ㅂ', 'ㅓ', 'ㅈ', 'ㅣ'> or <'ㅏ', 'ㅁ', ' ㅓ ',' ㅈ ',' ㅣ '> or something else.

Of course, the correct sequence is <'ㅏ', 'ㅂ', 'ㅓ', 'ㅈ', 'ㅣ'> and this will produce this result if the recognizer performed correctly. During the Viterbi search, you can reduce the recognition time if you use grammar relationships or grammar. Such information may be included in a subword dictionary or may be kept separately.

In the second step, the distance between the phoneme sequence and the words in the second dictionary is calculated (ST-305). Since this distance is obtained using the results already recognized by the Viterbi search, we will call it 'distance after recognition'.

If the phoneme sequence is <'ㅏ', 'ㅂ', 'ㅓ', 'ㅈ', 'ㅣ'> the shortest distance is '0' and the word is 'father'. As another example, if the phoneme sequence is <'ㅏ', 'ㅁ', 'ㅓ', 'ㅈ', 'ㅣ'>, the distance from the 'father' is not '0'. But perhaps this distance is the shortest compared to other words, so this time the recognition will be "father."

If the distance from other words is shorter, this is a misunderstanding.

There are many ways to measure distance. An example is shown below.

First, the distance from other phonemes is calculated in advance for each phoneme. Phonemes that are phonetically similar will come out shortly.

Second, even when phonemes are inserted or deleted in a word, distance values are defined and calculated. In the case of insertion, since the effect of the same phoneme being inserted at which position of which word is different, the distance must be calculated in consideration of the context dependency. This is also the case when using other subunit speech units.

If the distance value is already calculated, the overall distance is calculated using these values in the second step of recognition (ST-305), and thus the final recognition result is obtained (ST-306).

Then, it is determined whether or not the voice data to be continuously recognized exists (ST-307), and if the voice data exists, the process returns to the voice input process, and if the voice data does not exist, the recognizer is terminated.

As described above, the present invention is a speech recognition method over two stages and uses a distance after recognition in the second stage. The present invention provides more flexibility in dealing with word dictionaries and can improve the recognition speed of a large vocabulary recognition system.

On the other hand, the speech recognition method using the distance after the recognition according to an embodiment of the present invention is not limited only to the above-described embodiment, various modifications are possible within the scope not departing from the technical gist.

As described above, the speech recognition method using the distance after recognition according to the present invention can obtain a faster recognition speed in the large vocabulary speech recognition by using the sub-sequence determination and the distance measurement method after the recognition by the result value. In addition, there is an effect of facilitating the management of the dictionary by separating the dictionary for the Viterbi search and the word dictionary to be recognized.

Claims (6)

In the speech recognition system for recognizing input voice data, A subunit speech unit dictionary which does not change even when the words are inserted or removed as a dictionary for speech recognition; Speech recognition method using a distance after recognition, characterized in that for using the dictionary of recognition target words that can change according to the change of the word. In the speech recognition system for recognizing input voice data, Finding a subword sequence using a subword speech unit dictionary; And measuring the distance after recognition between the subword sequence and the words in the word dictionary to obtain the shortest word sequence as a recognition result. The post-recognition method of claim 2, wherein the final recognition result is obtained by calculating a distance from the target word of the subword unit sequence when the subunit unit is inserted or deleted from the recognized subword unit sequence. Speech recognition method using distance. The speech recognition method of claim 2, wherein the generating of the subword sequence comprises obtaining a result sequence using phonetic information and information imposed by a grammar. The speech recognition method of claim 2, wherein the distance measurement process after the recognition is performed by using a stored information by calculating a distance for each phoneme using a phonetic model. The speech recognition method of claim 2, wherein the distance measurement process after the recognition is performed by using information calculated based on a predetermined distance value of a phoneme of an additional word.