KR20090065102A - Method and apparatus for lexical decoding - Google Patents

Abstract

A vocabulary decoding method and an apparatus thereof are provided to shorten overall performance time of voice recognition without deteriorating voice recognition performance. Based on an inputted voice signal, length of a phoneme series which is outputted by decoding phonemes is detected. Based on the detected length of the phoneme series, recognition target words(202) similar to the length of the phoneme series are selected. Edition distance measurement with the phoneme series is carried out on the basis of the selected recognition target words. Through the edition distance measurement, at least one recognition target word having a minimum edition distance with the phoneme series is outputted.

Description

Vocabulary decoding method and apparatus {METHOD AND APPARATUS FOR LEXICAL DECODING}

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to human speech recognition (HSR) technology, and more particularly, to a method and apparatus for lexical decoding suitable for improving lexical decoding speed using word clustering and input phoneme length information. will be.

The present invention is derived from the research conducted as part of the IT growth engine technology development project of the Ministry of Information and Communication and the Ministry of Information and Telecommunications Research and Development. [Task Management Number: 2006-S-036-02, Task name: Large-capacity interactive dispersion for new growth engine industries] Development of processing voice interface technology].

In general, in order to perform HSR speech recognition, speech recognition is performed through automatic phone decoding and lexical decoding.

Hereinafter, with reference to the drawings will be described in detail.

1 is a block diagram showing the structure of a general user speech recognition system.

Referring to FIG. 1, the user speech recognition system 100 includes a phoneme decoding unit 102, a lexical decoding unit 104, and an acoustic rescoring unit 106.

The phoneme decoding unit 102 is a process of obtaining a column of a phoneme model that is output with maximum similarity with respect to an input voice signal. In the phoneme decoding process, an error may be included in the decoded phoneme string for reasons such as ambient noise and an incorrect acoustic model. In this regard, the lexical decoding unit 104 outputs N-best words, which are N target words with minimum edit distance between recognized words, while correcting such errors in the phoneme strings transmitted from the phoneme decoding unit 102. do. Thereafter, the N-best word output from the lexical decoding unit 104 is transferred to the acoustic rescoring unit 106, and the acoustic rescoring unit 106 uses the detailed acoustic model for the input N target words to perform rescoring. To reorder the recognition and output the recognition result.

In the conventional user speech recognition system operating as described above, the maximum similarity of the voice input from the user is output through the phoneme decoding, the lexical decoding, and the sound rescoring. It may take a certain amount of time for each stage to become speech recognition, and it is possible to perform speech recognition faster than the time required for existing speech recognition, but in this case, speech recognition performance is deteriorated. There was no problem.

Accordingly, the present invention provides a vocabulary decoding method and apparatus capable of improving the vocabulary decoding speed using word clustering and input phoneme length information.

In addition, the present invention, the lexical decoding method that can limit the searched words in the lexical decoding based on the length of the phoneme string output through the phoneme decoding, and to search only the search according to the difference in the number of phonemes when measuring the editing distance And an apparatus.

In addition, the present invention, by using the clustering technique based on the editing distance for all the words to be recognized, after dividing the clusters representing all words to the phoneme string input first to measure the similarity with the words representing these clusters N After selecting the -best cluster, a method and apparatus for lexical decoding that performs lexical decoding on only words constituting the cluster again.

According to an exemplary embodiment of the present invention, a method of detecting a length of a phoneme string which has been phoneme-decoded and output based on an input voice signal and selecting a recognition target word similar to the length of the phoneme string based on the detected phoneme string length Performing an editing distance measurement with the phoneme string based on the selected recognition object words, and outputting at least one recognition object word having a minimum editing distance with the phoneme string through the editing distance measurement. Process.

According to another exemplary embodiment of the present invention, there is provided a process of generating clusters by clustering words having high similarity with respect to edit distances for all recognized words, selecting a word representing the generated clusters, and Measuring the first similarity of the selected words and the phoneme sequence decoded and output based on the speech signal, selecting a corresponding cluster of words having an optimal similarity, and targeting only the words constituting the selected cluster And measuring a second similarity with the phoneme string and outputting an optimal word through the similarity measurement.

According to an embodiment of the present invention, an apparatus includes: a detector configured to detect a length of a phoneme string that is decoded and output based on an input voice signal, a word selector that selects a word to be recognized similar to the length of the detected phoneme string, and the selection; And an editing distance measuring unit configured to measure an editing distance with the phoneme string based on the recognized recognition words, and output at least one recognition object word having a minimum editing distance with the phoneme string through the editing distance measurement.

According to another exemplary embodiment of the present invention, a clustering unit generates a cluster by clustering words having high similarity with respect to editing distances for all recognized words, and a representative word line for selecting a word representing the generated cluster. A first similarity measurer configured to measure similarity between the selected words and the phoneme strings which are phoneme decoded and output based on the voice signal, and select a corresponding cluster of words having an optimal similarity, and the selected cluster And a second similarity measurer that measures similarity with the phoneme string based on only words to be output and outputs an optimal word through the similarity measure.

In the present invention, the effects obtained by the representative ones of the disclosed inventions will be briefly described as follows.

When the lexical decoding process is performed in a user speech recognition system, the present invention can improve the speed of the lexical decoding process by performing the lexical decoding process using the word clustering method or the input phoneme length information. There is an effect that can shorten the overall speech recognition performance time without degradation.

Hereinafter, the operating principle of the present invention will be described in detail with reference to the accompanying drawings. In the following description of the present invention, when it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. In addition, terms to be described below are terms defined in consideration of functions in the present invention, which may vary according to a user's or operator's intention or custom. Therefore, the definition should be made based on the contents throughout the specification.

The present invention is to improve the lexical decoding speed by using the word clustering method and the input phoneme length information, and limits the words searched at the time of lexical decoding based on the length of the phoneme string output through the phoneme decoding, and editing Limit the search to only as much as the phoneme difference in distance measurement, or divide the clusters representing all words into clusters representing all words using the editing distance-based clustering technique for all recognized words. After selecting the N-best cluster by measuring the similarity with words representing the clusters, lexical decoding is performed on only the words constituting the cluster.

As described above with reference to FIG. 1, user speech recognition through the user speech recognition system 100 is performed through the phoneme decoding unit 102, the lexical decoding unit 104, and the acoustic rescoring unit 106.

The phoneme decoding unit 102 converts a voice signal input through a phoneme decoding process into a phoneme sequence, and the vocabulary decoding unit 104 recognizes a recognition target word from the phoneme sequence converted through the vocabulary decoding process.

에 대해 최대 사후 확률을 가지는 음소열

을 구하는 과정이다.The phoneme decoding process is a feature vector sequence of a given speech signal as shown in Equation 1.

Phoneme series with maximum posterior probability for

The process of obtaining

과

이 필요하다.

는 음소로부터 관측된 음향 특징 벡터에 대한 조건부 확률로 일반적으로는 은닉 마르코프 모델(HMM : Hidden Markov Model)을 사용하여 모델링한다.

은 단어를 구성하는 음소들 간의 연결관계를 확률 모델로 정의한 것으로 언어모델이라고 하며 <수학식 2>와 같다.In order to find the solution of Equation 1, two probabilistic models

and

This is necessary.

Is a conditional probability for the acoustic feature vector observed from the phoneme, and is generally modeled using the Hidden Markov Model (HMM).

Is a linguistic model that defines the relationship between the phonemes that make up a word. It is called a language model.

However, most real speech recognition systems use n-grams like Equation 3 as an approximation to Equation 2, assuming that the current phoneme is affected only by the previous (n-1) phonemes. . Generally, you will use 2-gram or 3-gram.

In the lexical decoding process of decoding a word string from a recognized phoneme sequence, as shown in FIG. 2, a word having a minimum editing distance between a result of phoneme decoding including a misidentified phoneme string and a correct phoneme string for a recognized word is found. Based on a dynamic programming algorithm such as 4>.

는 누적 거리이고,

는 기준 음소 c_x가 t_y로 치환(substitution) 오류가 발생한 경우에 대한 cost 함수이고,

는 기준 음소 c_x에 대해 손실(deletion) 오류가 발생한 경우의 cost이고

는 기준 음소

에 대해 t_y의 삽입(insertion) 오류가 발생한 경우의 cost가 된다. 이들 cost들은 <수학식 5>와 같은 음소의 혼동 확률(confusion probability)에 대한 네거티브 로그(negative logarithm)로 표현된다.here,

Is the cumulative distance,

Is the cost function for the case where the reference phoneme c _x has a substitution error with t _y ,

Is the cost of a loss error for the reference phoneme _x

Phoneme reference

This is the cost when t _y insertion error occurs. These costs are expressed as a negative logarithm of the phoneme's confusion probability, as shown in Equation 5.

Hereinafter, as a method for improving the lexical decoding speed, there are a method using input phoneme length information and a word clustering method. First, a method of improving the lexical decoding speed using a method using phoneme length information will be described. do.

2 is a block diagram illustrating a lexical decoding structure of a search space reduction method using length information of phoneme strings according to an exemplary embodiment of the present invention.

Referring to FIG. 2, for example, if a phoneme string consisting of all N is output as a result of phoneme decoding, the least editing distance from the word to be recognized is most likely that the word is composed of N phonemes. The phoneme string number detector 200 in the lexical decoding unit 104 detects the phoneme string number from the input phoneme string. Thereafter, the detected phoneme string number is transferred to the recognition target word selector 204 based on the phoneme sequence number, and the phoneme sequence similar to the phoneme sequence number transferred among the recognition target words 202 in the recognition target word selector 204. The recognition target word 202 is selected and transmitted to the edit distance measurer 206 based on the phoneme length constraint.

Accordingly, the edit distance measurer 206 measures the edit distance between the pronunciation strings of the words constituting the word 202 to be recognized for the phoneme string, and outputs N words whose edit distance is the minimum as a result of lexical decoding. do. In this case, in order to know the phoneme recognition result and the editing distance between each word, a dynamic program such as <Equation 4> should be applied to all words, but the word to be searched by the number of phonemes resulting from phoneme decoding Even in the case of applying Equation 4 for limiting through the selection in (204) and obtaining the editing distance for these words, it is possible to reduce the overall search time by limiting the search space.

4 is a diagram illustrating a vocabulary decoding process according to phoneme length according to a preferred embodiment of the present invention.

Referring to FIG. 4, if a phoneme decoding result of a user pronounced voice is output as “gor jv g E ba xl” 400 consisting of nine phonemes, the phoneme string and two words “Koryeo Logistics Center” (404) ), It is likely that the editing distance of "Korea Construction" (402) is small when measuring the editing distance between each of "Korea Construction" (402). Compared to the phoneme strings 19 constituting the distribution center 404, the number of phoneme strings to be input is similar, and thus the editing distance for insertion is relatively small. Therefore, if the number of phoneme strings is given as a result of the phoneme decoding, if all the words that are to be recognized are searched for only words having a difference of a predetermined number (delta) compared to the number of phoneme strings to be input, Thus, N words having the minimum editing distance can be obtained without obtaining the editing distance.

As described above, the word selected by the phoneme length is used to obtain an editing distance between two phoneme strings using a dynamic program such as Equation 4. The process is to apply a global constraint. When two edit distances having the same phoneme length are obtained, an optimal edit distance is obtained through a dynamic program for the entire search space composed of two given phoneme sequences as shown in FIG. 5.

5 is a view showing whether or not to limit the search space when measuring the editing distance according to the phoneme length in accordance with a preferred embodiment of the present invention.

Referring to FIG. 5, for example, assuming that two phoneme strings are all composed of four, it is necessary to perform a search for all nodes of all 4x4 search spaces. However, if the length of the phoneme is defined in this way, the search space is limited by the difference in length between the two phoneme strings.

For example, if two phoneme strings have the same length, the search space may be sufficient to search only the diagonal line 500 indicated by the solid line of FIG. 5. However, if the number of two phoneme strings differs, the search space is diagonally divided by the difference. It should increase proportionally. A section 502 indicated by a dotted line is a case where a difference between two phonemes is generated by one, and a section 504 indicated by a thick dotted line is a case where a difference between two phoneme strings is generated by two. Therefore, by limiting the space to be searched in proportion to the difference between the lengths of the two phoneme strings for which the editing distance is obtained, the search time can be significantly reduced compared to finding all existing search sections.

Meanwhile, another method for improving the lexical decoding speed is to perform the lexical decoding process in two steps.

3 is a block diagram illustrating a two-stage lexical decoding structure using a word clustering method according to a preferred embodiment of the present invention.

Referring to FIG. 3, in the word clustering unit 306 based on the editing distance in the lexical decoding unit 104, a group having high similarity with respect to all recognized words 304 is bundled into a cluster, and a word cluster representative word selector In 308, a word representing this cluster is selected. Subsequently, the similarity measurer 300 between the phoneme string and the cluster representative word selects N optimal clusters by comparing the phoneme string and similarity first input to the word representing the cluster, and measures the similarity between the phoneme string and all words in the cluster. The unit 302 improves the overall lexical decoding speed by performing lexical decoding in two stages of lexical decoding on all words included in the cluster.

In this method, first, N clusters representing all the recognized vocabulary words should be generated and the words representing them are selected. This process is to obtain N clusters from all the words, which is obtained through the cluster partitioning process as shown in FIG.

6 is a flowchart illustrating a word clustering algorithm based on an edit distance between words according to phoneme lengths according to an exemplary embodiment of the present invention.

FIG. 6 illustrates an operation procedure in the word clustering unit 306 and the representative word selecting unit 308. In step 600, all recognition target words are input, and in step 602, the editing distance is measured between all recognition target words. In operation 604, the words having the minimum editing distance among the words measured in the step are incorporated into each other's cluster. Thereafter, in step 606, the entire editing distance is measured to determine whether the threshold is higher than the threshold. Return to step to measure the editing distance from the cluster. Incorporation is again performed between the clusters, and the procedure is repeated to end clustering only when the measured total edit distance is lower than the threshold.

The process of obtaining N clusters representing all words is called a quantization process of words to be recognized. In the lexical decoding process for the quantized words, first, the similarity between the phoneme string input from the similarity measurer 300 and the phoneme string input from the cluster representative word is measured between the words representing the cluster and the input phoneme string, so that the N most similarity values are small. Select clusters. The similarity measurer 302 between the phoneme string and all the words in the cluster for the selected cluster performs lexical decoding on all the words constituting the cluster and outputs the result of the final lexical decoding.

Since the two-step lexical decoding process finds a cluster which is a set of words in the first step and finds a word in the cluster based on the first step, it can be expected to improve much speed compared to a method of searching all the existing words at once.

As described above, the present invention is to improve the lexical decoding speed by using a word clustering method and input phoneme length information, and limits words searched in lexical decoding based on the length of a phoneme string output through phoneme decoding. When limiting the number of phonemes when editing the editing distance, limit the search to only the number, or use the editing distance based clustering technique for all recognized words to divide them into clusters representing all the words. First, the N-best cluster is selected by measuring similarity with words representing the clusters, and then lexical decoding is performed on only the words constituting the cluster.

Meanwhile, in the detailed description of the present invention, specific embodiments have been described, but various modifications are possible without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the scope of the following claims, but also by those equivalent to the scope of the claims.

1 is a block diagram showing the structure of a general user speech recognition system;

2 is a block diagram illustrating a lexical decoding structure of a search space reduction method using length information of a phoneme string according to a preferred embodiment of the present invention;

3 is a block diagram illustrating a two-step lexical decoding structure using a word clustering method according to an embodiment of the present invention;

4 is a diagram illustrating a vocabulary decoding process based on a phoneme length according to a preferred embodiment of the present invention;

5 is a view showing whether or not to limit the search space when measuring the editing distance according to the phoneme length in accordance with a preferred embodiment of the present invention,

6 is a flowchart illustrating a word clustering algorithm based on an edit distance between words according to a phoneme length according to a preferred embodiment of the present invention.

<Explanation of Signs of Major Parts of Drawings>

100: user speech recognition system 102: phoneme decoding unit

104: vocabulary decoding section 106: acoustic recoring section

200: phoneme number detector 202, 304: recognition target word

204: Recognition target word selector based on phoneme number

206: Editing distance measurer based on phoneme length constraint

300: a similarity measuring unit between phoneme strings and cluster representative words

302: a similarity measurer between phoneme strings and all words in a cluster

306: word clustering based on editing distance

308: word cluster representative word selection unit

Claims (10)

Detecting the length of the phoneme string output by phoneme decoding based on the input voice signal; Selecting words to be recognized similar to the length of the phoneme string based on the detected length of the phoneme string; Performing an editing distance measurement with the phoneme string based on the selected recognition target words; Outputting at least one word to be recognized that the editing distance with the phoneme string is minimum through the editing distance measurement; Vocabulary decoding method comprising a. The method of claim 1, The phoneme decoding is, And a phoneme string of a phoneme model is obtained by outputting a maximum similarity with respect to the input voice signal. The method of claim 1, The method, Determining a space to be searched according to the number of phoneme strings when the editing distance measurement is performed Vocabulary decoding method further comprising. Generating clusters by clustering words having similarities for each editing distance for all recognized words; Selecting a word representing the generated cluster; Selecting a corresponding cluster of words having an optimal similarity by measuring a first similarity of the selected words and the phoneme strings which are phoneme decoded and output based on the speech signal; Measuring a second similarity degree with the phoneme string based on only words constituting the selected cluster; The process of outputting the optimum word by measuring the similarity Vocabulary decoding method comprising a. The method of claim 4, wherein The phoneme decoding is, And outputting a maximum similarity degree to the input voice signal, thereby obtaining a column of a phoneme model. A detector which detects the length of the phoneme strings which are decoded and output based on the input voice signal; A word selecting unit which selects a word to be recognized similar to the detected phoneme length; An editing distance measuring unit configured to measure an editing distance with the phoneme string based on the selected recognition target words, and output at least one recognition object word having a minimum editing distance with the phoneme string through the editing distance measurement Vocabulary decoding apparatus comprising a. The method of claim 6, The phoneme decoding is, And a phoneme string of a phoneme model is obtained by outputting a maximum similarity with respect to the input voice signal. The method of claim 6, The editing distance measuring unit, Determining a space to be searched according to the number of phoneme strings when the editing distance measurement is performed Vocabulary decoding device further comprising. A clustering unit generating clusters by clustering words having high similarity with respect to each editing distance for all recognized words; A representative word selection unit for selecting a word representing the generated cluster; A first similarity measurer for measuring similarity between the selected words and the phoneme strings which are phoneme decoded and output based on a voice signal, and selecting a corresponding cluster of words having an optimal similarity; A second similarity measurer configured to measure similarity with the phoneme string based on only words constituting the selected cluster, and output an optimal word through the similarity measure Vocabulary decoding apparatus comprising a. The method of claim 9, The phoneme decoding is, And a string of phoneme models is obtained by outputting the maximum similarity with respect to the input voice signal.

Images