KR100608644B1 - Method For Recognizing Variable-Length Connected Digit Speech - Google Patents

Abstract

The present invention relates to a number sound recognition method.

In the method of recognizing variable-length linked numbers according to the present invention, a sound model and a language model for subdivided number sounds are generated by distinguishing the numbers sounded differently depending on whether the numbers sound adjacent to the left and right and the soft sound in a predetermined manner. and a) receiving a variable length connection digit by using a variable length and recognizing the digit by a predetermined method using the acoustic model and the language model.

According to the present invention, it is possible to have a high recognition rate for the digits that are pronounced differently according to the digits and softphones adjacent to the left and right.

Speech Recognition, Concatenated Digits, Granular Digits, Transition Penalties

Description

Method for Recognizing Variable-Length Connected Digit Speech}

1 is a view showing an HBT model according to a conventional technique.

2 is a view for explaining a connection speech recognition method using the HBT model according to the prior art.

3 is a diagram illustrating a segmented numeric model according to an embodiment of the present invention.

4 is a functional block diagram of a digit recognition device connected with a variable length according to an embodiment of the present invention.

5 is a functional block diagram of an apparatus for generating and updating an acoustic model and a language model according to an embodiment of the present invention.

6 is a flowchart illustrating a process of recognizing a number sound and generating and updating a sound model and a language model according to an embodiment of the present invention.

The present invention relates to speech digit sound recognition, and more particularly, to connected digits of a variable length based on a segmented digit sound model by distinguishing digit sounds that are pronounced differently depending on whether or not the digit sounds are adjacent to the left and right digits. It relates to a recognition method.

Human desire to pursue convenient life brings about technological development in various fields. Speech recognition is one of them. In the past, speech recognition technology has been developed from isolated word recognition to continuous speech recognition, and the size of the recognized words is gradually increasing. In addition, the voice-dependent voice recognition is gradually developing from speaker-independent speech recognition. The input voice is being developed from the intentional pronunciation of the reading to the technology of recognizing the speech in everyday conversation. In addition, the recognition in the low noise is developing into the recognition in the noisy real environment.

Speech recognition technology has different accuracy requirements depending on its purpose. In everyday conversation, for example, you can convey meaning without correctly recognizing all the syllables, while for commerce and numbers such as banking, you have to have a very high accuracy. The existing methods in terms of the recognition of numeric sounds have various problems. In other words, when recognizing a variable length concatenated phoneme using a simple word model, the performance of digital phone speech recognition is poor because it cannot reflect the presence or absence of the sound that may occur between the numbers and the effects of the articulation combination by the sound. not. In particular, this problem is more prominent in the case of Korean concatenated digits consisting of single syllables, unlike English. For example, in the case of "Oh" and "Oh", "Ie" and "I", or "Il" and "Il", there is a high possibility of insertion or deletion errors. Therefore, various methods have been attempted to increase the recognition rate of numeric sounds.

One of the technologies that has a relatively good numerical recognition rate compared to the previous numeric recognition, there is a method of recognizing a connected numeral sound using a head-body-tail (HBT) structured word model. 1 is a diagram showing an HBT model. In the HBT model, one digit has a head, a torso, and a tail. If the beginning or end of a word is the same pronunciation, the head or tail is shared. The body is composed of a context-independent model, and the head and tail are composed of a context-dependent model considering words that may be adjacent to left and right. Speech recognition technology based on HBT model applies the minimum classification error learning method in addition to the maximum likelihood learning method. The HBT model speech recognition method is mainly applied to connected speech recognition related to numbers, dates, days of the week, and times, and improves the overall false recognition rate of 17 to 24% compared with the whole word model.

2 is a view for explaining a connection speech recognition method using the HBT model according to the prior art.

Every word in the HBT model has a head, torso, and tail structure. As shown in FIG. 2, when two numbers "one" are connected, "one" and "this" each have one body, a plurality of heads, and tails. The body of "work" is context independent, but the tail is the number of the number of notes that can follow. On the other hand, in the case of "this", the body part is context independent, but the head part exists as many as the number that can be preceded. In Korean, the number of heads and tails of one number is 10 to 11, respectively. The relationship between the tail of each word model and the head of the next word applies the grammar or n-gram language model.

The HBT model has better performance than the previous number recognition because it considers left and right adjacent numbers. However, like the conventional numerical recognition method, the HBT model also has a limitation in that it does not reflect the variability of the pronunciation generated by the tone. In particular, unlike in the case of English, in the case of Korean, "i" or "oh", such as the number may not be recognized when not considering whether the tone. Accordingly, there is a need for a numerical sound recognition technology capable of having a better numerical sound recognition rate in consideration of the effects of adjacent numeric sounds and the effects of softening.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described needs, and it is an object of the present invention to provide a method for recognizing a variable length connected number sound having an excellent number sound recognition rate in consideration of effects due to adjacent number sounds and soft or not. .

In order to achieve the above technical problem, according to the present invention, a variable length connection number recognition method according to the present invention distinguishes a number sound which is pronounced differently depending on whether the number sound is adjacent to the left and right by a predetermined method and the sound for the subdivided number sound. (A) generating a model and a language model, and (b) receiving a variable length connection digit and receiving the digit by a predetermined method using the acoustic model and the language model. In the step (a), the acoustic model is a subdivided digit sound that is divided into digits that are pronounced differently depending on whether the digits are adjacent to the left and right, and the language model is a connection relationship between the subdivided digit vocabulary words. It is preferable that it is a Bygram language model that defines. On the other hand, comprising a step of learning a variable word transition penalty according to the pair of digits, and whether or not consecutively, it is preferable to recognize the digits by applying the learned word transition penalty, wherein, the variable bullet The method of learning the penalty may be a minimum classification error (MCE) learning.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

3 is a diagram illustrating a segmented numeric model according to an embodiment of the present invention.

The number "day" can be pronounced differently depending on the number following it. For example, as shown in FIG. 3, when the phrase "gu" is followed, it is pronounced as "day", but when the word "yuk" follows, it is often pronounced as "il". If "oh" follows, it is most likely pronounced "yi". On the other hand, the same string of numbers is not always pronounced the same. In other words, when pronounced "one phrase", the "one" and "nine" is read separately from the case of reading and continuous reading. In the present invention, in order to consider the variation of pronunciation according to the ringtone, the case of broken reading and the case of reading in connection are treated as different. In case of broken reading, it is followed by #, and in case of concatenating reading, it is separated by $. In the present invention, the numerical sounds are subdivided in consideration of the effects of the adjacent numerical sounds on the left and right, and the effects of softening. That is, in the present invention, "one #", one $, two, one "are all treated as different numeric sounds. The description of the granular numeral sounds will be described later with reference to FIG.

4 is a functional block diagram of a digit recognition device connected with a variable length according to an embodiment of the present invention.

The variable length connection sound recognition device includes a voice endpoint detection unit 10, a feature extraction unit 20, a Viterbi search unit 30, a re-estimation unit 40, an acoustic model DB 50, and a pronunciation dictionary DB. 60, and language model DB 70.

The voice endpoint detecting unit 10 determines a start point and an end point of the input voice. In other words, if you read the number by cutting off the number, the voice is not input even when all the connected numbers are not input. Therefore, it is judged that the voice has not been input continuously for a certain amount of time. Judgment should be made. When the voice endpoint detection unit 10 determines whether the input of the numeric sound is terminated, when the human inputs the voice, the user presses the input moment on the button and the end of the connected numeric sound based on the input voice while the button is pressed. It may be determined, but may be automatically determined in consideration of the degree of energy level of the input numeric sound and the time interval in which the voice is not input. Either way, accurately determining the end of the input numeric digits affects the performance of the concatenated digits.

The feature extractor 20 extracts a feature of the input concatenated sound. Feature extraction refers to the extraction of components useful for speech recognition from speech signals and is generally associated with the compression and dimensional reduction of information. The ideal method for feature extraction is currently unknown, but it is a feature expression that reflects human perceptually meaningful characteristics, robust to various noise environments / speakers / channel variations, and feature extraction that expresses temporal changes well. It is mainly studied in the feature extraction field. The main features used for speech recognition include linear predictive coding (LPC) cepstrum, perceptual linear prediction (PLP) cepstrum, mel frequency cepstral coefficient (MFCC), differential cepstrum, filter bank energy, and differential energy. In the embodiment of the present invention, a feature vector of 25th order per frame is used for feature extraction. The 12th order is MFCC, the 12th order is differential MFCC, and the remaining 1st order is differential energy.

The Viterbi search unit 30 searches forward the word string W that best matches the feature vector sequence based on the λ of the acoustic model, the pronunciation dictionary, and the language model, and the observable feature vector sequence of the connected numerals. Search). This process can be expressed as an equation (1).

Equation applied to searching for the most likely word string

In Equation 1, Pr (X | W, λ) represents the probability that the column of the acoustic model determined by the pronunciation dictionary is matched with the feature vector sequence, and Pr (W | λ) represents the language model probability value of the word sequence W. . In the forward search, all word grids are detected using the Viterbi search algorithm, and in the backward search, the final recognized number string is determined through the re-estimation 40. The Viterbi search algorithm is well known in the speech recognition field and will not be described in detail. In the omnidirectional search, in order to reduce the search cost, the triphone model is applied only inside the word, and the recalculation unit 40 applies the interword triphone model to reestimate the probability values to be matched to obtain a final recognition result.

The acoustic model DB 50 stores an acoustic model for recognizing the connected numeral sound. The acoustic model used in the present invention is based on the Hidden Markov Model (hereinafter referred to as HMM). The unit of the acoustic model for speech recognition may be a phoneme, a diphone, a triphone, a quinphone, a syllable, a word, or the like. Since the present invention considers the effects of adjacent digits (coarticaulation), it is possible to use a diphone, a triphone, a quinone, etc., but it is necessary to consider both the effects of the front and back of the digits and the number of vocabularies since only the digits are recognized. Since it is small, it is desirable to make an acoustic model with a triphone.

The pronunciation dictionary DB 60 stores a pronunciation model for modeling the pronunciation of a word that is a recognition unit. The pronunciation model is a simple model with one pronunciation per word using a representative phoneme obtained from a standard phonetic dictionary, a multiple phonetic model using multiple headings in a recognized vocabulary dictionary to take into account acceptable pronunciations, dialects and accents. There may be a variety of statistical pronunciation models that consider the probability of each pronunciation, phoneme-based lexical pronunciation model. In an embodiment of the present invention, a phoneme-based pronunciation dictionary is generated using a dictionary phonetic model, and then it is extended to a triphone pronunciation dictionary.

The language model DB 70 stores grammar. The language model is mainly used for continuous speech recognition. The speech recognizer reduces the search space of the recognizer by using the language model in the search process, and improves the recognition rate because the language model plays a role of increasing the probability of sentences that match the grammar. Grammar types include syntaxes for formal languages such as Finite State Network (FSN) and Context-Free Grammar (CFG), and statistical grammars such as n-gram. The n-gram is a grammar that defines the probability of the next word from the past n-1 words. Types include bigograms, trigrams, and four grams. In the embodiment of the present invention, a bigram is used, which treats a differently pronounced number sound as another word according to the variation according to the adjacent number and whether or not the melody is used, and increases the recognition rate by using the grammar of the language model for the linkability of each word. For example, in the case of a pronunciation in which "yl" and "yuk" are connected, it may be determined that "yl + six #" or "yl # | yuk #" is correct for grammar, but "yi + yng #". Can be judged to be incorrect in grammar. If you allow this, it makes it difficult to distinguish between "this flesh" and "one flesh". If it is determined that the grammar is wrong, a small probability value is given to reduce the possibility of selecting the phoneme string.

5 is a functional block diagram of an apparatus for generating and updating an acoustic model and a language model according to an embodiment of the present invention.

Learning process is required to create (or update) acoustic model and language model. First of all, there should be voice samples pronounced by various people. The voice DB 110 collects and stores data on various pronunciations of various people. Meanwhile, the word DB 120 stores word level transcriptions for the voice stored in the voice DB 110. Speech samples and word-level scripts thereof find the most optimal state sequence through the decoder 130. In the embodiment of the present invention, the decoding unit 130 uses a Viterbi algorithm, and since the Viterbi algorithm is well known in the art, detailed description thereof will be omitted. Through the decoding unit 130, the voice samples and the word level transcripts are changed into phoneme level transcripts and word level transcripts subdivided according to the present invention, respectively, which are phoneme level transcription DB 150 and word level transcription DB 140, respectively. Are stored in. Phoneme level transcriptions stored in the phoneme level transcription DB 150 are used to generate (or update) an acoustic model through the acoustic model learning tool 160, and the generated sound model is stored in the acoustic model DB 180. Meanwhile, word level transcriptions stored in the word level transcription DB 140 generate (or update) a language model through the language model learning tool 170, and the generated language models are stored in the language model DB 190. In addition, word level transcriptions generate a pronunciation dictionary 210 by setting a word transition penalty by the language model recalculation 200 after the language model learning tool 170, at which time MCE (Minimum Classification Error) learning method. This is used.

When recognizing the concatenated digits through the Viterbi search unit 30 and the re-estimation unit 40 based on the segmented digit model, the recognition error rate is not known when the length of the concatenated digits is not known. high. This is prominent in relation to the recognition of a single vowel sound. Recognition errors are discussed in Table 1 and Table 2.

TABLE 1

Comparison table for each digit

As can be seen from Table 1, in the case of one, two, and five, the addition error and the deletion error are more prominent than the other numbers.

On the other hand, when the length of the concatenated number is known and the length is not known, the error degree is different. The recognition error rate according to the presence or absence of information on the length can be seen in Table 2.

TABLE 2

Error rate with length information

If you do not know the length If you know the length Length determination error String error rate String error rate HBT model 11.00 18.16 11.81 The present invention 11.97 17.18 7.75

As can be seen from Table 2, when the length is unknown, the error rate is almost similar to that of the present invention or the HBT. However, if the length information is known, it can be seen that the error rate of the present invention is much smaller. That is, in the present invention, it can be seen that the error rate due to the negative number length error is higher than that of the conventional method. Misprints of numeric length are closely related to insertion or deletion errors, and various methods have been attempted to reduce such insertion or deletion errors. One of them is a method of giving a word insertion penalty according to the word length, and Equation 2 shows a formula for finding a recognition result by applying a word insertion penalty used in the conventional speech recognition field.

Speech recognition based on word insertion penalty

However, this method has difficulty in controlling word addition and deletion errors when a word insertion penalty value of the same value is given because there is no consideration of whether a consecutive sound is included in Korean number recognition consisting of a single syllable vocabulary.

Accordingly, the present invention introduces a method called word transition penalty. In other words, instead of using fixed word insertion penalty in the conventional speech recognition problem, the variable penalties are applied at the transition between the subdivided numbers to reflect the effect on the effect of the sound. The formula for the word transition penalty is shown in equation (3).

Speech recognition by word transition penalty

The word transition penalty was repeatedly estimated for the word transition penalty values using a minimum classification error (MCE) learning method to minimize the error of the training data.

Table 3 shows an example of calculating word transition penalty values for each pair of digits (word pairs).

TABLE 3

Example of the word transition penalty

6 is a flowchart summarizing a process of recognizing a number sound and generating and updating a sound model and a language model according to an embodiment of the present invention.

First, voice number files and word level transcriptions thereof are input (S10). The input voice number files and the Dalbell warriors generate (or update) an acoustic model and a language model (S30) through a decoding process (S20). Then, when the voice digit is input (S40), the feature extraction process (S50) and the number string determination process (S60) are performed.

7 is a block diagram illustrating an apparatus for recognizing a numeric sound and generating and updating an acoustic model and a language model according to an embodiment of the present invention.

First, in the learning process, the phonetic number DB and word level transcriptions are received as inputs, and the string and phoneme level transcription of the segmented numeric model is obtained through the Viterbi decoding process. The trained language model is used to generate (or update) a language model using language model learning tools. The language model is reestimated through discreet learning methods such as MCE. Using all of the learned knowledge models, the input voice recognizes the input voice and the number string with the highest matching score through the voice endpoint detection unit, feature extraction process, Viterbi search process, and reestimation process.

Those skilled in the art will appreciate that the present invention can be embodied in other specific forms without changing the technical spirit or essential features of the present invention. Accordingly, the embodiments described above are to be understood in all respects as illustrative and not restrictive. The scope of the present invention is indicated by the scope of the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and the equivalent concept are included in the scope of the present invention. Should be interpreted.

According to the present invention, the variation of left and right numbers and the influence of soft noise are considered in the recognition of variable length concatenated numbers, and the high recognition rate in the recognition of variable length concatenated numbers by reducing the number of deletion and addition errors. Has In the present invention, the error rate of the linked numeral recognition and the comparison result by the conventional method are shown in Table 4.                     

TABLE 4

As can be seen from Table 4, it can be seen that the numerical length of the known length has a better recognition rate than the conventional methods in the case of using the granular number vocabulary according to the present invention. Introduced to improve performance.

Claims (4)

delete (A) generating a sound model and a language model for subdivided digit sounds by distinguishing digit sounds that are differently pronounced according to whether the digit sounds adjacent to the left and right are predetermined by a predetermined method; And (B) receiving a variable length connection number and recognizing the number in a predetermined manner using the acoustic model and the language model, The initial consonants of the digits are pronounced differently depending on the finality of the adjacent digits, In the step (a), the acoustic model is a subdivided digit sound which is divided into digits which are pronounced differently depending on whether the digit is adjacent to the left and right, and the language model is a connection between the subdivided digit vocabulary. Variable length concatenated speech recognition method characterized in that it defines a bigram language model The method of claim 2, further comprising: learning a word transition penalty variable according to whether a pair of digits and a consecutive tone are determined by a predetermined method, and recognizing the digits by applying the learned word transition penalty. Recognition method of concatenated number sound of variable length 4. The method of claim 3, wherein the method of learning the variable word transition penalty is a minimum classification error (MCE) learning method.

Images