KR20030010979A - Continuous speech recognization method utilizing meaning-word-based model and the apparatus - Google Patents

Abstract

PURPOSE: A continuous voice recognition method and system using a meaning word unit model are provided to improve voice recognition rate by using the meaning word model as well as a phoneme model. CONSTITUTION: A characteristic vector to be used for voice recognition is extracted from an input voice signal. Pattern matching for the extracted characteristic vector is performed using a phoneme unit model and a meaning word unit model, which uses a meaning word essential for constructing a sentence as a separate recognition unit(320). An optimum sentence is selected using information of meaning words among sentences obtained by performing the pattern matching.

Description

Continuous speech recognization method utilizing meaning-word-based model and the apparatus

The present invention relates to a speech recognition method and apparatus, and more particularly, to a speech recognition method and apparatus using a semantic unit model.

Conventional speech recognition methods include word-based speech recognition and phoneme-based speech recognition according to recognition units. The word-based speech recognition method sets the object to be recognized in word units so that the recognition device learns the word and finds the most similar word as the output for the input. Word-based speech recognition method can bring excellent recognition result if it is well trained, but in order to recognize word unit, it must have learning model and learning data about tens of thousands of words, especially for all word models. Equitable learning is a very difficult task and takes up a lot of memory to maintain tens of thousands of word models. That is, the word-based speech recognition method has a problem that it is difficult to sufficiently learn and costs a lot.

The phoneme-based speech recognition method constructs words by recognizing phonemes, which can perform relatively superior recognition tasks at a lower cost than the word-based speech recognition method, but it can have a fatal effect on word composition when phoneme recognition fails. .

In addition, there is a functional language model that is used to take advantage of both the word-based speech recognition method and the phoneme-based speech recognition method. The functional model is a functional word that is frequently used in sentences but is hard to hear in spoken conversation, such as 'a' in English. Recognition units such as ',' 'an', 'in', and 'and' are composed of their own word model. The functional unit unit speech recognition method is that the recognition unit model is made for the investigation of the functional words, that is, the prepositions, conjunctions and Korean of the English language, and outputs the most similar words through the pattern matching with the already learned word model as the recognition result. That is, when a word voice is input, recognition results are derived in phoneme units, and words are formed by referring to the pronunciation dictionary. In the case of functional words, recognition is performed without reference to the phonetic dictionary, instead of phoneme recognition. do. This method has high confidence in recognition results because it compares and analyzes words at once, but it is very difficult to optimally train word models.

Many papers have demonstrated that the recognition rate is improved when the functional unit model is used. In addition, there is no problem of learning that word unit recognition has. However, this method is limited to functional words and still has a problem of recognition error for words that contain important information for constructing actual sentences.

In addition, unlike the isolated word recognition method, which can recognize the desired word only with the learned phoneme model and the pronunciation dictionary, the word network is additionally used in the continuous speech recognition method to bind the recognized words to form a sentence. In this case, the size of the search space depends on how the word network is constructed, and since this affects the performance of the recognizer, the word network plays an important role in recognizing the correct sentence.

Word networks are mainly implemented with word-pair grammars, N-grams, and finite state automata (FSA). The main idea is to concatenate words that can appear after a word, either fixed by a rule or a statistical probability value. The word-pair grammar matches only words that can appear after a particular word. For example, "eat" + "want" can be connected in order, but not vice versa. An example of such a word network structure of word-pair grammar is shown in FIG. N-grams use statistical probabilities to link words to words. We use a bunch of training data to calculate the probability that a word appears after a word and then search for the higher probability. FSA combines all the configurable sentences into a network. This method improves recognition performance, but has a problem in that when a sentence other than the composed sentence comes in, it cannot be recognized.

As such, the existing word network has a limit in constructing free sentences while reducing search space. In other words, if the search space is reduced, the degree of freedom of recognizable sentences decreases (particularly in the case of Korean where the word order is free). On the contrary, the search space increases to recognize sentences of various word orders.

The present invention basically utilizes the advantages of word unit recognition using phoneme-based speech recognition techniques to solve the above problems. Unlike functional word unit recognition, the present invention plays an important role in understanding the structure of a sentence and understanding the meaning. It is to improve recognition rate by designing sentence components or key words as a separate recognition unit ("meaning word").

In addition, the present invention is to reduce the search space by using a word network centered on the semantic words to improve the performance of the recognition and to derive the optimum recognition results using the semantic information for error correction of the recognition results .

1A is a block diagram of a speech recognizer according to a conventional functional unit recognition model.

1B is a block diagram of a speech recognizer according to a conventional phoneme unit recognition model.

2 is a block diagram of a word network implemented with a conventional word-pair grammar.

3 is a schematic block diagram of a continuous speech recognition system according to the present invention;

Figure 4a is a conceptual diagram of a conventional phoneme unit HMM model.

4b is a conceptual diagram of a semantic unit HMM model according to the present invention;

5 is a conceptual diagram of a semantic centric word network according to the present invention;

Explanation of symbols on the main parts of the drawings

300: preprocessor

310: learner

320: recognizer

360: post processor

In order to solve the technical problem as described above, a sentence component, for example, a verb and a key word, which play an important role in grasping the structure of a sentence spoken by a speaker and an understanding of a meaning, is used as a separate speech recognition unit model, and a word network is used. It consists mainly of semantic words and also finds the best recognition result by using semantic information related to semantic words in post processor.

One aspect of the present invention relates to a speech recognition method, comprising extracting a feature vector to be used for recognition from an input speech signal, using a phoneme unit model, and a semantic that plays an important role in sentence construction as a separate recognition unit. And performing pattern matching on the extracted feature vector signal using the semantic unit model as a recognition unit model.

Another aspect of the present invention relates to a continuous speech recognition method, comprising: extracting a feature vector to be used for recognition from an input speech signal; and separately recognizing a phoneme unit model and a semantic that plays an important role in constructing a sentence as a recognition unit model. Performing pattern matching on the extracted feature vectors using a semantic center word network that uses a semantic unit model that is a unit and constructs a network using semantic information of the semantic word, and results of performing pattern matching Selecting the optimal sentence using the information of the semantic among the sentences.

Another aspect of the present invention relates to a speech recognizer, comprising: a preprocessor for extracting a feature vector to be used for recognition from an input speech signal; and a separate unit for recognizing a phoneme unit model and a semantic that plays an important role in sentence construction as a recognition unit model. And a recognizer for performing pattern matching on the feature vector signal output from the preprocessor using a semantic unit model.

Another aspect of the present invention relates to a continuous speech recognition system, comprising: a preprocessor for extracting feature vectors to be used for recognition from an input speech signal; A recognizer for performing pattern matching on the feature vector signal output from the preprocessor using a semantic unit word model using a semantic unit model as a recognition unit and using a semantic center word network configured with a semantic information of the semantic word. And a post-processor for selecting an optimal sentence using information possessed by the semantic word among the pattern matching result sentences of the recognizer.

Preferably, the semantic words in the speech recognition method and the speech recognizer include a verb or a key word that plays an important role in sentence construction.

In addition, preferably, in the voice recognition method and the voice recognizer, the input voice signal includes a reservation, a securities service related conversation signal.

The present invention will now be described in detail with reference to FIGS. 3 to 5.

3 shows an example of a voice recognition device according to the present invention. The speech recognition apparatus includes a preprocessor 300, a learner 310, a recognizer 320, and a postprocessor 360.

The preprocessor 300 extracts a feature vector to be used for learning and recognition from the analog voice signal input to the input. That is, the voice input signal coming into the upper end of the preprocessor is data for voice recognition, and the voice data coming into the bottom end represents data for use in learning.

The learner 310 statistically learns the basic unit of recognition, and the basic unit of recognition may be phoneme, morpheme or word. In the present invention, the learner is a recognition unit model to be used in the recognizer and a phoneme model. Create a unit model. That is, the learner 310 learns a model by constructing a main sentence component, a verb, and a key word into a separate recognition unit, that is, a semantic unit, based on a phoneme unit model. The method of learning uses the learning method of the existing speech recognition device, but adjusts the learning data so that the learning between the semantic model and the phoneme unit model is fair, and the number of the semantic models may be quite large, If it is limited to a certain part, for example, hotel reservation, securities service, etc., there may be a small number of semantic models, which may facilitate learning. Such a semantic model enables accurate semantic analysis of spoken sentences by increasing the recognition rate of words that play an important role in constructing and assigning sentences.

The recognizer 320 uses the recognition unit model 330 learned from the learner 310, the pronunciation dictionary 340, and the semantic center word network 350 to add feature vectors of the speech signal output from the preprocessor. Pattern matching is performed. When the recognizer is implemented using the HMM (Hidden Markov Model) in the present invention, the phoneme unit model generally has a three-state left and right shape as shown in FIG. 4A, and the semantic unit model is shown in FIG. 4B. As will be the left and right forms with more states than this.

In the present invention, the word network 350 used for continuous speech recognition is mainly composed of semantic terms. That is, the word network is configured by using the intention and meaning information of the semantic word. For example, suppose you have a semantic model of "reservation." The intent of this semantics is for "reservation." The information contained in the (reservation) subject, (reservation) date, and (reservation) object. , (Reservation) number, etc. Based on this information, a semantic center word network can be constructed as shown in FIG. 5. When the semantic word is detected from the semantic center word network configured as described above, it is possible to check whether the word is recognized as the correct sentence by using the category in the word network including the words generated in the entire recognition process. For example, if a semantic word for "reservation" is detected, and a category that is not related to "reservation" appears in the recognition process, the sentence is misrecognized and there is no need to perform the recognition operation any further in that direction. In this way, the search space can be reduced.

The post processor 360 finds a sentence having the least error among the N recognition result sentences outputted from the recognizer 320, and also finds an optimal sentence suitable for the meaning using information possessed by the semantic word. That is, when several semantic words appear in the recognition sentence, the priority is determined to select the most suitable word category for constructing the sentence. For example, in "Please confirm that I made a reservation yesterday", if the semantics are two "reserved" and "confirm", then the semantics that play an important role will be "confirm". Therefore, it is necessary to check in the postprocessor whether the sentence has passed the word network matching the word "confirm" or the postprocessor will find the optimal sentence among the N recognition candidate sentences.

According to the present invention as described above, the performance of speech recognition can be improved by using a semantic unit model using not only a phoneme unit model but also a semantic word such as a verb which is essential for sentence structure.

Further, according to the present invention, the search space can be reduced by constructing the word network around the semantic center.

In addition, according to the present invention, by using the semantic information on the recognition result candidates coming out through the word network, the sentence having the least error can be found.

Claims (8)

In the voice recognition method, Extracting a feature vector to be used for recognition from an input speech signal; Speech recognition comprising the step of performing pattern matching on the extracted feature vector signal using a semantic unit model having a semantic unit as a separate recognition unit as a semantic unit that plays an important role in the phoneme unit model and sentence construction. Way. In the continuous speech recognition method, Extracting a feature vector to be used for recognition from an input speech signal; As a recognition unit model, we use a semantic unit model that uses a semantic unit as a separate recognition unit as a phoneme unit model and a sentence that plays an important role in sentence structure, and uses the semantic information of the semantic to construct a network Performing pattern matching on the extracted feature vectors; The pattern matching performance continuous speech recognition method comprising the step of selecting the optimal sentence using the information that the semantic words among the sentences. The method according to claim 1 or 2, The semantic words include verbs or key words that are important for sentence construction. The method according to claim 1 or 2, The input voice signal includes a voice signal related to a reservation and securities service. In the voice recognizer, A preprocessor for extracting feature vectors to be used for recognition from an input speech signal; A recognition unit model includes a recognizer that performs pattern matching on a feature vector signal output from the preprocessor using a semantic unit model having a semantic unit as a separate recognition unit as a phoneme unit model and a sentence structure. Voice recognizer. In the continuous speech recognizer, A preprocessor for extracting feature vectors to be used for recognition from an input speech signal; As a recognition unit model, we use a semantic unit model that uses a semantic unit as a separate recognition unit as a phoneme unit model and a sentence that plays an important role in sentence structure, and uses the semantic information of the semantic to construct a network A recognizer for performing pattern matching on the feature vector signal output from the preprocessor; And a post-processor for selecting an optimal sentence using information possessed by the semantic word among the sentence matching result sentences of the recognizer. The method according to claim 5 or 6, The semantic word is a speech recognizer including a verb or a key word important to construct a sentence. The method according to claim 5 or 6, The input voice signal is a voice recognizer including a reservation, securities service-related talk signal.