JPH1185184A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the speech recognition device in which a user is not required to utter certain contents and continuous voice recognition is conducted while executing a speaker adaptive process by uttering only 'No instructor is in charge'. SOLUTION: Voice recognizers 4 and 6 refer to an acoustic model and the statistical language model in terms of word units and continuously conduct voice recognition of the word strings of an uttered voice sentence based on the voice signals of an inputted uttered voice sentence. Then, an instructor signal generating section 21 refers to the word dictionary which includes the phoneme strings corresponding to each word string and transforms the word strings outputted from the recognizer 6 into a phoneme string. An on-line speaker adaptive control section 22 uses the transformed phoneme string as the instructor signals, executes an on-line speaker adaptive process against the acoustic model to update the model. In such a case the section 22 executes an on-line speaker adaptive process based on a quasi-Bayes approximation and the statistical language model includes an N-gram of the word having a variable length N.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition apparatus provided with online speaker adaptation control means for performing speaker adaptation processing in real time.

[0002]

2. Description of the Related Art On-line speaker adaptation processing, in which speaker adaptation is performed sequentially using the speech to be recognized as it is, improves the acoustic model of an unknown speaker step by step according to the amount of speech. Therefore, it is highly expected as a practical method from the viewpoint of construction of a speech recognition system. As online speaker adaptation processing, Quasi Bayes (Qu
Asi-Bayes) online speaker adaptation methods based on approximations are described, for example, in the prior art document "Q. Huo et al.," As
tudy of on-line Quasi-Bayes adaptation for CDHMM-b
ases speech recognition ", In Proceeding of the Inte
rnational Conference on Acoustics, Speech, and Signa
l Processing, pp. 705-708, May 1996 ". ). In this method, the utterance content is configured under the condition of "supervised", that is, the teacher signal is manually applied to operate, and its effectiveness is confirmed.

[0003]

However, in the "supervised" speaker adaptation processing, it is necessary to urge the user to utter a certain amount of content, which complicates the dialog control and increases the burden on the user. It will be imitated. SUMMARY OF THE INVENTION An object of the present invention is to solve the above problems and eliminate the need to prompt the user to utter a certain content, and to continuously perform speech recognition while performing speaker adaptation processing without "supervising". It is an object of the present invention to provide a voice recognition device capable of performing the above-mentioned operations.

[0004]

According to a first aspect of the present invention, there is provided a speech recognition apparatus for inputting an uttered speech sentence with reference to a predetermined acoustic model and a predetermined word-based statistical language model. The speech recognition unit continuously recognizes the word sequence of the uttered speech sentence based on the speech signal of the above, and refers to a word dictionary including a phoneme sequence corresponding to each word sequence, and is output from the speech recognition unit. A conversion unit for converting a word sequence into a phoneme sequence, and using the phoneme sequence converted by the conversion unit as a teacher signal, performing an online speaker adaptation process on the acoustic model to convert the acoustic model Updating means for updating.

According to a second aspect of the present invention, in the speech recognition apparatus of the first aspect, the adaptation means executes an on-line speaker adaptation process based on Quasi-Bayes approximation. It is characterized by doing.

[0006] Further, the speech recognition apparatus according to claim 3 is
3. The speech recognition device according to claim 1, wherein the statistical language model includes an N-gram of a word having a variable length of N.

[0007]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a continuous speech recognition apparatus according to an embodiment of the present invention. The continuous speech recognition apparatus according to the present embodiment detects the word hypothesis of the uttered speech sentence based on the characteristic parameter of the speech signal of the input uttered speech sentence using a known one-pass Viterbi decoding method. In the continuous speech recognition apparatus provided with the word matching unit 4 which calculates and outputs the degree, (a) the same words having the same end time and different start time output from the word matching unit 4 via the buffer memory 5 For the word hypothesis, the N-
With reference to the statistical language model 13 including the variable length N-gram of each word as a gram, the total likelihood calculated from the utterance start time to the end time of the word for each head phoneme environment of the word. A word hypothesis narrowing unit 6 for narrowing down word hypotheses so as to be represented by one word hypothesis having the highest likelihood; and (b) a word string of a speech recognition result output from the word hypothesis narrowing unit 6 Is converted to a phoneme sequence with reference to the word dictionary 12, and (c) the phoneme sequence output from the teacher signal generation unit 21 is used as a teacher signal to form a well-known quasi Bayesian approximation. An online speaker adaptation control unit 22 for updating a phoneme hidden Markov model (hereinafter, referred to as a phoneme HMM) 11 as an acoustic model using an online speaker adaptation method based on the speaker model.

In the "unsupervised" speaker adaptation process, a teacher signal which is information corresponding to the utterance content in "supervised" is internally generated based on a recognition result using an acoustic model before speaker adaptation. I have. For this reason, the effect of speaker adaptation largely depends on the accuracy (recognition rate) of the generated teacher signal. "Unsupervised" line speaker adaptation based on quasi-bayesian approximation using a recognition result of a continuous speech recognizer (so-called phoneme typewriter) with only Japanese syllable rules as linguistic constraints as teacher signals In the following (hereinafter referred to as a comparative example), the recognition performance was hardly improved due to insufficient accuracy of the teacher signal. Therefore, the recognition performance of online speaker adaptation based on Quasi-Bayesian approximation based on the "unsupervised" Quasi-Bayes approximation should be used unless the recognition result of a recognizer with performance that exceeds at least the phoneme typewriter is used as a teacher signal using stronger linguistic constraints. The present inventor expects that the variable length N-gram (for example, disclosed in Japanese Patent Application Laid-Open No. 09-134192, which will be described in detail later), which is a statistical language model in units of words, is a language constraint. (Vocabulary to be recognized,
It is assumed that the domain is known. We have invented continuous speech recognition while executing online speaker adaptation processing based on "unsupervised" Quasi Bayes using the recognition result of the continuous speech recognizer as a teacher signal. This allows
As will be described later in detail, by using the variable length N-gram, a highly accurate teacher signal can be generated, and an adaptive effect can be obtained even under the condition of "no teacher".

The statistical language model 13 used in this embodiment
Are generated by the language model generation unit based on the text data for learning, and include the statistical language model 13
Is based on bigrams between part-of-speech classes (N = 2), but words that can be independently trusted are separated from part-of-speech classes, treated as a single class, and frequent words are used to improve prediction accuracy. For sequences, those words are combined and treated as a class, allowing the representation of long word chains, and thus the model generated is a statistical model that combines the features of part-of-speech bigrams and variable-length word N-grams. It is a language model, characterized by a balance between accuracy and reliability of transition probability.

First, the variable length N used in this embodiment
-The concept of gram is explained below. The N-gram is a (N-1) -fold Markov model, which is simple (1-1) to store the past (N-1) state transitions.
Heavy) Each state of the Markov model is interpreted as being separated. As an example, FIG. 3 shows a state transition diagram in which a bigram is represented as a Markov model, and FIG. 4 shows a state transition diagram in which a trigram is represented as a Markov model.

[0012] In FIG. 3, but remains in state s ₁ when output symbol a in state s _1, the state s ₁
In a transition to a state s ₂ which has output the symbol b. But it remains in the state s ₂ when outputting the symbol b in the state s _2,
Back in state s ₂ to the state s ₁ when outputting the symbol a. Figure 4 trigram, state status s ₁ bigrams s ₁₁
Vital separated into state s ₁₂ and it is believed that separation of the state s ₂ in the state s ₂₁ and the state s _22. Further, by promoting the separation of all states, a higher-order N-gram is obtained.

The variable length N-gram shown in FIG. 5 is obtained by partially separating the states of the simple Markov model. That is, in the bigram of FIG. 3, when the symbol a is output from the state s ₂ and the symbol b is continuously output (this is referred to as ab, and the symbol ab is output), the state other than b is continued. (This is expressed as a (/ b) and the symbol a (/ b) is output. Here, / represents a negative meaning and is a bar (overline)). in the former case, while transitioning from state s ₁ to the state s _12, in the latter case, the transition from state s ₂ to the state s _11. That is, in the former case,
Separated from the state s ₁ to state s _12, in which left remaining transition for outputting the symbols a to (a (/ b)) in the state s _11. Incidentally, in this model, whereas a transition to the state s ₁₂ when outputting the symbol ab in the state s _11, the state s
_{When the} symbol a (/ b) is output in step ₁₁ , the state remains s11. Further, while in the state s ₁₂ and remain in the state s ₁₂ when outputting the symbol ab, while s ₁₂ symbols a (/
b) a transition to the state s ₁₁ when the output.

This model is characterized in that a long chain can be represented with a simple Markov model structure by considering a plurality of consecutive symbols as new symbols. By repeating the same state separation, a longer chain can be represented locally. This is the variable length N-gram. That is, a variable-length word N-gram as a language model in which a symbol is regarded as a word is represented as a bigram between word strings (including one word).

Next, the operation of the variable length N-gram will be described. Statistical language model 22 used in this embodiment
Is a variable length N-gram of a part of speech class and a word, and is expressed as a bigram between the following three types of classes. (1) part of speech class (hereinafter referred to as first class),
(2) A class of a word separated from the part of speech class (hereinafter, referred to as a second class), and (3) a class formed by combining connected words (hereinafter, referred to as a third class).

The words belonging to the first class are mainly those having a low appearance frequency, and the reliability of the transition probability is improved as compared with the case where the words are handled alone. In addition, the words belonging to the second class are mainly those having a high frequency of appearance, and have sufficient reliability even if handled alone. Furthermore, the words connected to each other are combined and classified into the third class. , Operates as a variable-length N-gram, and the prediction accuracy of the next word is improved. However, in the present embodiment, it is not considered that the part-of-speech class and the part-of-speech class that are connected to each other and the combination of the part-of-speech class and the word are considered. Generation probability P (w of a sentence composed of a plurality of L words
₁ ^L ) is given by the following equation.

[0017]

(Equation 1)

[0018] In this case, ws _t is the sentence at the time of the classification of the above class, means the t-th word string (a stand-alone word, is also included). _{Therefore, P (ws t | c t} ) is the probability that the word string ws _t appears when found t-th _{class, P (c t | c t} -1) is the previous (t- 1) Probability that a word of the t-th class appears from the class. Further, K in the text represents the number of word strings, and K ≦ L. Therefore, Π in Equation 1 is a product from t = 1 to K.
Here, as an example, consider a sentence of an uttered voice sentence composed of the following seven words.

[0019]

[Equation 2] "I-Murayama-and-Words-I-Mas-"

The generation probability P (w ₁ ^L ) of this sentence is given by the following equation using Equation 1.

[Equation 3] P (w ₁ ^L ) = P (Watakushi | {Watakushi})
P (｛Watashi｝) ・ P (Murayama ｜ <proper noun>) ・ (<proper noun> ｜ ｛wattaku｝) ・ P (and | ｛and｝) ・ P (｛and｝ | <proper noun>) ・ P (Say | [say]) ・ P ([say] |
{When})

However, <>, ｛｝, and [] indicate that they belong to the first class, the second class, and the third class, respectively. However, each word and word string belong as follows. (1) Since “Murayama” is a noun, it belongs to the first class. (2) “I” and “to” are combinations of nouns and particles, and belong to the second class. (3) "I say" is a combination of a verb, a verb suffix, an auxiliary verb, and an auxiliary verb suffix, and belongs to the third class. Here, in the second and third classes, the appearance frequencies of the word and the class are equal, so that P (wattashi | ｛wattaku｝)
= 1, P (and | ｛and｝) = 1, P (say | [say]) = 1, and therefore, the above Equation 3 becomes the following equation.

[0022]

[Equation 4] P (w ₁ ^L ) = P (Watashi) ・ P (Murayama ｜ <proper noun>) ・ P (<proper noun> ｜ watashi) ・ P (and | <proper noun>) ・ P (say | And)

In FIG. 1, it is connected to a word collating unit 4,
For example, a phoneme HMM1 stored in a hard disk memory
1 includes each state, and each state has the following information. (A) State number (b) Acceptable context class (c) List of preceding state and succeeding state (d) Parameter of output probability density distribution (e) Self transition probability and transition probability to succeeding state Since it is necessary to specify which speaker each distribution originates from, the phoneme HMM 11 used in the embodiment is generated by converting a predetermined speaker mixed HMM. Here, the output probability density function is a Gaussian mixture distribution having a 34-dimensional diagonal covariance matrix.

The word dictionary 12 connected to the word matching unit 4 and stored in, for example, a hard disk is a phoneme HMM
For each of the 11 words, a symbol sequence or a phoneme sequence indicating a reading expressed by a phoneme symbol is stored.

In FIG. 1, a uttered voice of a speaker is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the word matching unit 4 via the buffer memory 3.

The word collating unit 4 uses the one-pass Viterbi decoding method and the word hypothesis using the phoneme HMM 11 and the word dictionary 12 based on feature parameter data input via the buffer memory 3. Is detected, the likelihood is calculated and output. Here, the word matching unit 4 determines whether each HMM
The likelihood within a word and the likelihood from the start of utterance are calculated for each state. The likelihood is individually provided for each word identification number, word start time, and difference between preceding words. Further, in order to reduce the amount of calculation processing, the grid hypothesis of a low likelihood among the total likelihoods calculated based on the phoneme HMM 11 and the word dictionary 12 is reduced. The word collating unit 4 outputs the resulting word hypothesis and likelihood information to the word hypothesis narrowing unit 6 via the buffer memory 5 together with time information (specifically, a frame number, for example) from the utterance start time. .

The word hypothesis narrowing section 6 refers to the statistical language model 13 based on the word hypothesis output from the word collation section 4 via the buffer memory 5 and has the same end time and the same start time. Is represented by one word hypothesis having the highest likelihood among the total likelihoods calculated from the utterance start time to the end time of the word for each head phoneme environment of the word. After narrowing down the word hypotheses so as to cause them, the word string of the hypothesis having the maximum total likelihood is output as the recognition result among the word strings of all the narrowed word hypotheses. The output word string is input to the teacher signal generator 21 to generate a teacher signal.
In the present embodiment, preferably, the first phoneme environment of the word to be processed is three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. I mean a line.

[0028] For example, as shown in FIG. 2, the (i-1) th word W _i-1 of the following phoneme string a _1, a _2, ..., come i th word W _i consisting a _n Sometimes, six hypotheses Wa, Wb, Wc, Wd, We, and Wf are assumed as the word hypotheses of the word Wi _-1.
Exists. Here, the former three word hypotheses Wa, W
It is assumed that the final phonemes of b and Wc are / x /, and the final phonemes of the latter three word hypotheses Wd, We and Wf are / y /. The hypothesis with the highest total likelihood among the hypotheses in which the end time t _e is equal to the first phoneme environment (the top three word hypotheses in which the _first phoneme environment is “x / a ₁ / a ₂ ” in FIG. 2) (for example, FIG. 2
Are deleted except for the top hypothesis). Note that the fourth hypothesis from the top has a different phoneme environment, that is, since the last phoneme of the preceding word hypothesis is y instead of x,
Do not delete the fourth hypothesis from the top. That is, only one hypothesis is left for each final phoneme of the preceding word hypothesis. In the example of FIG. 2, one hypothesis is left for the final phoneme / x /, and one hypothesis is left for the final phoneme / y /.

On the other hand, the teacher signal generating section 21 converts the word string of the speech recognition result output from the word hypothesis narrowing section 6 into a phoneme string with reference to the word dictionary 12, and performs online speaker adaptation control. Output to the unit 22. In response, the online speaker adaptation control unit 22 uses the phoneme sequence output from the teacher signal generation unit 21 as a teacher signal, and uses a known online speaker adaptation method based on the well-known Quasi-Bayes approximation. For each utterance of the speaker's voice, the phoneme HMM
11 is updated to execute the online speaker adaptation processing.
Specifically, the estimation of the parameters of the target phoneme HMM 11 is performed by approximating the probability density function of each parameter. The criterion of the approximate distribution is that the mode of the distribution is the same, and the value is a known EM (Estimation-Maxim
ization) algorithm. At this time, by effectively eliminating the influence of the previous learning data, it is possible to quickly follow a change in the environment such as a change in the speaker. At this time, the influence of the learning data is stored in a parameterized form called a hyperparameter, and is updated sequentially while being used for speaker adaptation together with new utterance data.

In the above embodiment, the head phoneme environment of the word is defined as a sequence of three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. Although defined, the present invention is not limited to this. The phoneme sequence of the preceding word hypothesis including the final phoneme of the preceding word hypothesis, and at least one phoneme of the preceding word hypothesis that is continuous with the final phoneme, And a phoneme sequence that includes a phoneme sequence that includes the first phoneme of the word hypothesis.

In the above embodiment, the feature extracting unit 2
The word matching unit 4, the word hypothesis narrowing unit 6, the teacher signal generation unit 21, and the online speaker adaptation control unit 22 are constituted by, for example, a digital computer. A phoneme HMM 11, a word dictionary 12, and a statistical language model 13 are stored in a storage device such as a hard disk memory.

In the above embodiment, speech recognition is performed using the word collating unit 4 and the word hypothesis narrowing unit 6.
The present invention is not limited to this, and includes, for example, a phoneme matching unit that refers to the phoneme HMM 11 and a speech recognition unit that performs speech recognition of words by referring to the statistical language model 13 using, for example, the One Pass DP algorithm. You may.

In the above embodiment, the online speaker adaptation control unit 22 uses the phoneme sequence output from the teacher signal generation unit 21 as a teacher signal to perform online speaker adaptation based on the well-known Quasi-Bayes approximation. Although the phoneme HMM 11 that is an acoustic model is updated using the method, the present invention is not limited to this, and another online speaker adaptation method may be used.

[0034]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present inventor conducted an experiment on the continuous speech recognition apparatus shown in FIG. 1 as follows. The speaker adaptation is performed for each utterance, and the acoustic model to which the adaptation has been performed is used as a model of the adaptation source of the speaker adaptation using the generation of the teacher signal in the next utterance. Originally, the online speaker adaptation process
The speech to be recognized is used as adaptive data as it is,
In the experiment of the present invention, a speech to be recognized (test data) was prepared separately from the adaptation data in order to compare the effect of the adaptation result for each utterance under the same conditions. Here, the generation of the teacher signal was performed for the five speakers under the following conditions.

[0035]

[Table 1] Acoustic analysis 標本 Sampling: 12 kHz Frame interval: 10 msec Hamming window: 20 msec Feature value: log power + 16th order LPC cepstrum + Δlog power + 16th order LPC cepstrum ─────────────────────────────

[Table 2] Acoustic model ─────────────────────────────────── Unspecified speaker HMNet 600 state 5 mixture +1 state 10 mixed silence model ───────────────────────────────────

[Table 3] Language model {6,370 words, variable length N-gram (number of separated classes: 500)} ──────────────────────────

Table 4 shows the adaptation data for each speaker. Table 5 shows the phoneme error rate when the speaker-independent model, which is the initial model, is used, that is, the accuracy of the teacher signal.

[0037]

[Table 4] Data used for adaptation ──────────────────────────── Speaker name Number of conversations Number of utterances Speech time ──── ──────────────────────── FYOMA 11 282 415.48 seconds FUYO 9 245 383.31 seconds FHITA 15 508 699.47 seconds FTOAS 6 215 317. 92 seconds MATA 246 78.20 seconds ────────────────────────────

[Table 5]

It can be seen that the accuracy of the teacher signal is greatly improved in the speech recognition apparatus using the method of the present embodiment, as compared with the comparative example using the phoneme typewriter as the language constraint. Since the actual teacher signal is generated using an acoustic model that has been successively advanced, the accuracy of the teacher signal is further improved. For example, the speaker FYOMA has increased from 10.6% to 8.6%.

A teacher signal was generated by the speech recognition apparatus shown in FIG. 1 to perform speaker adaptation. Table 6 shows the phoneme error rate (in the comparative example, a phoneme typewriter was used) for the test data when the online speaker adaptation for each utterance was completed for all utterances. FIG. 6 shows the test data TAS
The change of the phoneme error rate when the adaptive data is increased by one utterance for 12008 is shown.

[0040]

[Table 6] Phoneme misrecognition rate after speaker adaptation ─────────────────────────────────── Test data Speaker Phoneme misrecognition rate (%) 形態 Embodiment With teacher No adaptation ───────────────── ────────────────── TAS12008. A FYOMA 19.33 17.09 24.23 TAS12010. A FYOMA 14.55 13.09 22.18 TAS 13005. B FYUYO 15.49 16.20 21.83 TAS13009. B FYUYO 16.11 14.99 26.02 TAS22001. B FHITA 17.99 15.35 27.01 TAS23002. A FHITA 16.19 14.52 32.59 TAS32002. A FYOAS 16.75 13.66 21.65 TAS 33001. B FYOAS 12.95 9.76 19.53 TAS33011. B MDATA 17.83 17.04 29.86 ─────────────────────────────────── Average 17.46 15.74 25.08───────────────────────────────────

In Table 6, "supervised" is a case where a teacher signal is manually given, and "no adaptation" is a case where a phoneme HMM without speaker adaptation is used. As is clear from Table 6 and FIG. 6, the “unsupervised” online adaptation of this embodiment is not as good as “supervised”, and the recognition performance is clearly improved and the utterance required for adaptation is also “supervised”. It was confirmed that there was no big difference from the case of "."

As described above, according to the present embodiment, a teacher signal is obtained by a continuous speech recognizer using a statistical language model in units of words as a linguistic constraint, and "unsupervised" Quasi-Bayes approximation is performed. The teacher signal can be automatically generated by executing the online speaker adaptation process used. A recognition rate can be obtained. In this embodiment, the more the device is used, the more the acoustic model can be adapted, so that the speech recognition performance can be significantly improved. Also, by using an acoustic model including a variable length N-gram, a highly accurate teacher signal can be generated, and a large speaker adaptation effect can be obtained even under the condition of "no teacher".

[0043]

As described above in detail, according to the present invention, a predetermined acoustic model and a predetermined statistical language model in units of words are referred to based on a speech signal of an input uttered speech sentence. With reference to a speech recognition means for continuously recognizing the word string of the uttered speech sentence and a word dictionary including a phoneme string corresponding to each word string, the word string output from the speech recognition means is converted into a phoneme string. A conversion unit for converting, and an adaptation unit for updating the acoustic model by executing an online speaker adaptation process on the acoustic model using the phoneme sequence converted by the conversion unit as a teacher signal. Is provided. Here, the adaptation means is preferably quasi
Perform an online speaker adaptation process based on the Quasi-Bayes approximation, wherein the statistical language model preferably includes N-grams of words of variable length N.

Therefore, the teacher signal can be automatically generated by obtaining the teacher signal by the speech recognition means using the statistical language model of each word as a linguistic constraint and executing the online speaker adaptation processing. In addition, it is possible to obtain an extremely high speech recognition rate as compared with the related art without prior preparation for the uttered voice of an unspecified speaker. In the present invention, the more the device is used, the more the acoustic model can be adapted, so that the speech recognition performance can be greatly improved. Also, by using an acoustic model including a variable length N-gram of a word, a highly accurate teacher signal can be generated, and a large speaker adaptation effect can be obtained even under the condition of "no teacher".

[Brief description of the drawings]

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a timing chart showing a process of a word hypothesis narrowing section 6 in the continuous speech recognition device of FIG.

FIG. 3 is a state transition diagram showing a bigram statistical language model.

FIG. 4 is a state transition diagram showing a statistical language model of a trigram.

FIG. 5 is a state transition diagram showing a model under a variable length N-gram used in the continuous speech recognition device of FIG. 1;

FIG. 6 is a graph showing experimental results of the continuous speech recognition device of FIG. 1, showing a phoneme recognition error rate with respect to an adaptive data amount.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3, 5 ... Buffer memory, 4 ... Word collation part, 6 ... Word hypothesis narrowing part, 11 ... Phoneme HMM, 12 ... Word dictionary, 13 ... Language model generation part, 21 Teacher signal generator, 22 ... Online speaker adaptation controller.

Continued on the front page (72) Inventor Harald Singer 5 Shiraya, Inaya, Koika-cho, Soraku-cho, Soraku-gun, Kyoto ATI Spoken Language Translation and Communication Research Laboratories, Inc. (72) Inventor Atsushi Nakamura Seika, Soraku-gun, Kyoto 5, Aoyama-cho, Sanraya, Kota-cho, A / T ARL Speech Translation Research Institute, Inc. (72) Inventor Zhang Huo 5-san, Hiratani, Seiya-cho, Seika-cho, Soraku-gun, Kyoto A-5 Aoyama, Sanriya Translation Communication Laboratory

Claims (3)

[Claims] 1. A word sequence of said uttered voice sentence is continuously spoken based on a voice signal of an input uttered voice sentence with reference to a predetermined acoustic model and a predetermined word-based statistical language model. With reference to a speech recognition means to recognize and a word dictionary including phoneme strings corresponding to the respective word strings,
A conversion unit that converts a word sequence output from the speech recognition unit into a phoneme sequence; and performs an online speaker adaptation process on the acoustic model using the phoneme sequence converted by the conversion unit as a teacher signal. And an adaptation means for updating the acoustic model. 2. The speech recognition apparatus according to claim 1, wherein said adapting means includes a Quasi-Basic device.
yes) An online speaker adaptation process based on approximation. 3. The speech recognition apparatus according to claim 1, wherein said statistical language model includes an N-gram of a word having a variable length of N.