JPH05188991A - Speech recognition device - Google Patents

Abstract

PURPOSE:To provide the speech recognition device which can be improved in the ability to discriminate between similar word speeches. CONSTITUTION:A likelihood calculation part 6 calculates a series of >=1 candidates (i) from the connection model of phoneme HMMs as to an input speech and when the candidate series is entered into a previously stored similar word table 8, a back tracing part 9 finds the optimum state transition series of the candidates by the connection model of phoneme HMMs; and a feature time series is fractionized, a discrimination effective section corresponding to a feature difference part is automatically extracted, and auxiliary features in the discrimination effective section of the candidate are used by a partial similarity calculation part 10. Then a total decision part 12 outputs optimum word recognition result information I based upon the likelihood of the phoneme HMM and the partial similarity of the auxiliary features.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus which can improve the ability to identify similar words by using a probabilistic model.

[0002]

2. Description of the Related Art In recent years, research and development of voice recognition devices have been conducted. In this voice recognition device, a method based on pattern matching is generally used. For example, Document 1:
Oki Electric Research and Development, Vol. 53, No. 2, pp61
66, April 1987, an example of this type is given.

This pattern matching method calculates the degree of similarity (or distance) between a standard pattern registered in a dictionary in advance and an input pattern, and inputs the category of the standard pattern that can give the maximum degree of similarity (or minimum distance). It is identified as the recognition result of the pattern. Here, as described in Document 1 above, DP (Dynamic Programming) is used to calculate the similarity (or distance) between patterns of unequal length, such as between a standard pattern and an input pattern.
g) The matching method is generally used. However, D
The method using P matching has the property that it is extremely strong in the fluctuation and deformation of the time structure of the time series pattern, but it is weak in the fluctuation of the spectrum.

To solve such a problem, so-called Hidden Markov, in which a word is represented by a state transition, an occurrence probability in each state and a transition probability between states are given, and an input speech is applied to this model to be recognized. There is a method based on Model (Hidden Markov Model) (hereinafter, abbreviated as HMM). For example, “Voice / Hearing and Neural Network Model” by Seiichi Nakagawa, Kiyohiro Kano, and Yoichi Higashikura in Reference 2 1
August 25, 990, Ohmsha, pp. 45-pp. 5
0 describes the model.

The HMM method statistically models spectral fluctuations and temporal fluctuations of utterance by a large number of learning samples, and is considered as a powerful method of speech recognition.

The speech recognition method using the HMM employs a method of obtaining the category of the HMM model which gives the maximum likelihood as a recognition result for the identification of the input pattern.

[0007]

The conventional speech recognition method using the maximum similarity (or the minimum distance) or the maximum likelihood of the HMM is relatively unaffected by the local noise of the input pattern, and is relatively small. It has an advantage that stable voice recognition can be performed. However, since the overall similarity of the input voice pattern is mainly evaluated, there is a problem that a word including a partially similar portion is likely to be erroneously recognized.

For example, even though the input voice is the place name "Bisai (Onishi)", the voice recognition recognizes it as "Hisai (Hisui)" or the number "ichi (1)".
There is a problem of erroneously recognizing "Hichi (7)", and it has been desired to realize a voice recognition device that improves such erroneous recognition.

The present invention has been made in view of the above problems, and an object thereof is to provide a voice recognition device capable of improving the discrimination ability between similar word voices.

[0010]

In order to achieve the above object, the present invention uses a hidden Markov model method corresponding to an input speech signal and uses a plurality of candidate words corresponding to the input speech signal, and A speech recognition apparatus for performing speech recognition by obtaining likelihood is improved by including the following characteristic means.

That is, similar word storage means for preliminarily grouping and storing similar words, and standard feature information storage means for preliminarily storing at least one kind of standard feature information of spectrum envelope corresponding to recognizable words. And a feature information series extraction unit that extracts the same type of feature information of the spectrum envelope of the input speech signal as the standard feature information and outputs the feature information in a series, and the similarity from the candidate word and its likelihood. The word storage means is searched for, and a back trace process is performed according to the search result to extract an optimum discrimination effective section, and the characteristic information series of the optimum discrimination effective section and the standard characteristic information are provided. It is characterized in that the above likelihood is optimally changed to generate and output the optimal word identification result.

[0012]

According to the present invention, the characteristic information series extraction means allows the characteristic information of the input speech signal, for example, the energy difference between the low-frequency component and the high-frequency component of the speech spectrum, the inclination of the spectrum, and the speech signal. The likelihood of the candidate word is optimally changed by using the feature information of the spectrum envelope, which is obtained by extracting information such as the number of zero crosses (zero crossings), and the standard feature information stored in advance corresponding to the recognizable word. Since the optimum word identification result in the series of the optimum identification effective section can be output based on the difference in the degree of similarity, it is possible to accurately identify the difference in phonological features between similar words. It will be easier.

Therefore, even if a word of a similar phoneme sequence exists in the input speech to be recognized, it can be accurately identified.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT A preferred embodiment of a voice recognition apparatus according to the present invention will be described with reference to the drawings. The purpose of this embodiment is to realize a speech recognition apparatus capable of accurately outputting a recognition determination result even when a word of a similar phoneme sequence exists in a speech to be recognized, and further improving the discrimination ability of similar word speech. It is to be.

In order to achieve this purpose, this embodiment is
In, at least,
Also calculates a column of one or more word candidates (i) and
Is written in the similar word table that is stored in advance.
If it is, the phoneme HMM connection mode of the candidate
The optimal time transition sequence by
Then, the effective area for identification corresponding to the difference in features is automatically
Auxiliary features (speech space
The energy difference between the low and high frequencies of the cutout,
The tilt of the cuttle and the number of zero crossings of the audio signal
(Information such as) is used for recognition, and the likelihood of phonological HMM
And the pattern matching method are used complementarily,
To the optimum word recognition (identification) result information I
I was able to do my best.

FIG. 1 is a functional block diagram of a voice recognition device according to this embodiment.

In FIG. 1, the speech recognition apparatus according to this embodiment has a filter 1, an A / D converter 2, an acoustic feature extraction unit 3, an auxiliary feature extraction unit 4, and an optimum code vector selection unit 5. , Likelihood calculation unit 6, determination unit 7, near similarity table 8, back trace unit 9, partial similarity calculation unit 10, auxiliary feature standard pattern unit 11, comprehensive determination unit 1
2 and.

FIG. 2 shows functional blocks of the optimum code vector selection unit 5.

In FIG. 2, the optimum code vector selecting section 5 is composed of a code vector section 52 and a code vector selecting section 51.

FIG. 3 shows functional blocks of the likelihood calculating section 6.

In FIG. 3, the likelihood calculating section 6 is
It is composed of a phonological HMM part 61, a word dictionary part 62, a word concatenation model part 63, and a Viterbi comparison and verification part 64.

In FIG. 1, the filter 1 limits the high frequency component of the input audio signal S and supplies the obtained signal to the A / D converter 2. The A / D converter 2 quantizes the high-frequency-limited audio signal into a desired number of bits, and supplies the quantized signal to the acoustic feature extraction unit 3 and the auxiliary feature extraction unit 4.
The acoustic feature extraction unit 3 performs necessary speech section detection to obtain, for example, a low-order LPC cepstrum as an acoustic feature. As the order, for example, the order of 15 is assumed. This L
The calculation method of the PC cepstrum is described in, for example, Reference 3: “Digital Audio Processing”, Tokai University Press, September 25, 1985.
Sun, author: Sadahiro Furui, pp. 47-pp. 48 and the like. The time series of the LPC cepstrum obtained by this calculation is called an acoustic feature series V. The acoustic feature series V is supplied to the optimum code vector selection unit 5.

The auxiliary feature extraction unit 4 extracts an auxiliary feature from the quantized speech signal and outputs an auxiliary feature sequence B. The extraction type of this auxiliary feature is the auxiliary feature standard pattern section 1 described later.
It is desirable that the type is the same as the type stored in 1. This auxiliary feature series B is, for example, at a certain time interval (for example,
BPF (Band Pass) every 10 msec.
Filter) Bank analysis and FFT (Fast Fo)
high frequency components (for example, 3 kHz to 5 kHz) and low frequency components (for example, 100 Hz to 900 Hz) of the spectrum obtained by urie Transform analysis or the like.
Energy difference between and (including energy deviation)
Alternatively, the information is a series of information such as the slope of the spectrum when the spectrum is linearly approximated, the number of zero crossings of the audio signal, and the number of zero crosses (zero crossings).
For example, in general, the number of zero-crosses increases as the frequency component of the signal increases, and decreases as the frequency component of the signal decreases. Therefore, the spectrum distribution can be known by extracting the number of zero-crossings.

This auxiliary feature sequence B is used as a feature which is effective in identifying between word notation groups in the near similarity table in the partial similarity calculation unit 10 which will be described later. Moreover, since the auxiliary feature series B is used for pattern matching, it is desirable that it is a feature effective for identification. However, when it is desired to reduce the processing amount or the hardware amount, the acoustic feature sequence V
The same sequence may be used. Then, the auxiliary feature sequence B is supplied to the partial similarity calculation unit 10.

The optimal code vector selection unit 5 processes the acoustic feature sequence V, outputs an optimal code vector, and supplies it to the likelihood calculation unit 6. Then, the internal code vector portion 52 is obtained by selecting a vector that best represents these samples from the acoustic sequence of a large number of training samples in advance. The code vector can be selected by a clustering method or the like. This clustering method is described in, for example, Document 4: IEEE Transactionon co.
mmunication, Vol. com-28, N
o. 1, January, 1980, pp. 84-p
p. 95, Linde, Buzo, andGray, "
An Algorithm for Vector Q
The algorithm of Linde et al. (referred to as LBG
It is called an algorithm. ) Can be used to find out. The code vector selection portion 51 searches the acoustic feature sequence V for the code vector portion 52 by using the clustering method or the like, selects the closest code vector, and supplies it to the likelihood calculation unit 6 as the optimum code vector sequence A. To do. The likelihood calculating unit 6 calculates a candidate sequence (i) from the supplied optimal code vector sequence A and the likelihood corresponding to the candidate sequence (i), and supplies the likelihood to the determining unit 7. That is, in the likelihood calculation unit 6, the word concatenation model portion 63 calculates at least one or more candidate strings (i) using the phoneme HMM portion 61 and the dictionary portion 62. The method of connecting the word models will be clarified in the description of FIGS. 5 and 6 described later.

The Viterbi comparing and collating unit 64 obtains the likelihood of the obtained word model by the Viterbi algorithm as shown in the above-mentioned reference 2, for example. Then, the likelihood LA (i) from the highest rank
And the word number (i) that gives the likelihood is output as a candidate. This candidate sequence is represented as i ₁ , i ₂ , ... And the likelihoods corresponding to this candidate sequence are LA (i ₁ ), LA (i ₂ ) ,.
・ ・ ・The above candidate sequences i ₁ , i ₂ , ...
And ..., the likelihood corresponding to the candidate column LA _(i 1), L
And A (i ₂ ), ... Are supplied to the determination unit 7.

Next, the decision section 7 refers to the reject threshold value R0 that is input and compares it with the supplied likelihood to perform reject processing. That is, when the likelihood is less than the reject threshold R0, it is determined that the word is not similar to any word (does not correspond) and rejected. However, if it is equal to or greater than the threshold value R0, it is determined whether or not the candidate is described in the near similarity table 8. Note that the above-mentioned likelihood always takes a positive value unless there is an underflow in the calculation, so setting a reject threshold R0 = 0 is equivalent to not performing the reject process.

In the near-similarity table 8, word names having a similar similarity, that is, similar word names are grouped and described. For example, if it is a place name, the group of "Bisai (Onishi)" and "Hisai (Hisui)" or the group of "Soka (Soka)" and "Moka (Moka)",
If it is a numeral, a group such as "ichi (1)" and "shichi (7)" is registered in advance in word notation.

FIG. 4 shows a specific table configuration of an example of the near similarity table 8. In this FIG.
The near-similarity table 8 is stored in, for example, a memory or the like, and its configuration is composed of a first group to a GN0th group, and the total number GN0 of groups is set at the head of the table. Then, each group has, for example, the total number GM (i) in the i-th group, the first word notation HYO (i, 1), and the second word notation HYO.
(I, 2), ..., GM (i) th word notation H
YO (i, GM (i)) and identification start ratio WSTART
(I) and the identification end ratio WEND (i).

Then, the judging section 7 judges whether or not the candidates up to the m-th place are listed in the near similarity table 8. Here, if the near similarity table 8 describes, for example, m = 2, GM (i) = 2, HYO (i, 1) = hisai, HYO (i, 2) = Visai, the likelihood calculation If the pair of "Hisai" and "Visai" is supplied as the 1st and 2nd place of the calculation result of the unit 6, it is determined that there is a description in the table and is output.

Moreover, in the processing in the judging section 7,
If there is no description in the near similarity table 8, the first candidate i ₁ that gives the maximum likelihood LA (i ₁ ) is output as the word recognition determination result information I. That is, I = i ₁ is output. On the other hand, if there is a description in the near similarity table 8, this is supplied to the back trace unit 9.

The back trace unit 9 supplies the candidate sequence i ₁ supplied from the Viterbi comparing and collating unit 64 of the likelihood calculating unit 6,
The optimal state transition sequence of the word concatenation model part 63 concatenated with the model of the phoneme HMM part 61 is obtained by backtrace processing, assuming that i ₂ , ... Perform segmentation. A specific example of this backtrace processing will be described with reference to FIGS. 7 and 8 described later.

Specifically, since each phoneme HMM uses a model with the same number of states, if the optimum state sequence is obtained by backtrace processing, a "time change point" (corresponding to the connected portion of the phoneme HMM) (corresponding to Frame number) to be used. It is considered that there are usually a plurality of LL time change points, and time (1) and time (1) respectively.
(2), ...

Identification start ratio W in the near similarity table 8
From START (i), the identification end ratio WEND (i), and the number of input frames T, an effective section [S
T, ED] is obtained as follows.

In the section [WSTART (i) * T (identification start frame), WEND (i) * T (identification end frame)], ST = minimum time (L) forL = 1
~ LL ED = maximum time (L) for1 = 1
~ LL By obtaining this [ST, ED] section, it is possible to automatically extract the different portion of the features when the candidate sequences i ₁ , i ₂ , ... Are assumed. Therefore, ST and E above
D is [WSTART (i) * T (identification start frame),
WEND (i) * T (identification end frame)]
It represents the minimum value and the maximum value of me (L). Above S
A specific example of the calculation processing method of T and ED is shown in FIG. 9 described later.

The back trace result of the back trace unit 9 and the auxiliary feature series B are obtained by the partial similarity calculating unit 10.
Is supplied to.

The partial similarity calculation unit 10 uses the backtrace result of the backtrace unit 9 and the auxiliary feature sequence B to obtain the section [ST, ED] effective for identifying the candidate.
Of the auxiliary feature series B (hereinafter, referred to as partial similarity) LB (i ₁ ), LB (i ₂ ), ...
For this calculation, the standard pattern REF (i, of the auxiliary feature pattern previously stored in the auxiliary feature standard pattern section 11 is used.
j, k) are used. The number RNO of stored standard patterns corresponds to the number of recognized words.

Various methods can be considered for the definition of the partial similarity, but here, the angle between the vectors is used as an example of the partial similarity.

That is, the input frame length T is set to the time length VLEN.
The auxiliary feature sequence B time-normalized to SPEC (i,
j), (i = 1 to DNO, j = 1 to VLEN, DNO:
With the number of auxiliary feature dimensions (the number of auxiliary feature types), VLEN: fixed number of frames, the standard pattern of the auxiliary feature standard pattern unit 11 is REF (i, j, k), (i = 1 to DNO,
j = 1 to VLEN, k = 1 to RNO), the angle r (j, j of the vector of the auxiliary feature series of the jth frame and the auxiliary feature standard pattern of the jth frame, kth category
k) is expressed by the following equation 1.

[0040]

[Equation 1] Further, the above-mentioned partial similarity can be expressed by the following mathematical formula 2 by the mathematical formula 1.

[0041]

[Equation 2] The partial similarity thus obtained is supplied to the comprehensive determination unit 12.

The comprehensive judgment unit 12 calculates the calculation result in the likelihood calculation unit 6 (likelihood in Viterbi calculation) and the partial similarity calculation unit 1.
A word recognition determination result I that is comprehensively determined is output using the calculation result of 0 (partial similarity of auxiliary features).

In order to output this word recognition determination result I, first, for the calculated candidate strings i ₁ , i ₂ , ...
Likelihood LA (i ₁ ), LA (i ₂ ) ... in the Viterbi operation,
Auxiliary feature partial similarity LB (i ₁ ), LB (i ₂ ), ...
And are used to obtain the overall determination similarity L (i). This comprehensive determination similarity L (i) is defined by the following equation (3).

[Equation 3] The w1 and w2 are positive constant weighting coefficients.

Finally, the word recognition (identification) determination result I is expressed by the following equation (4).

[0045]

[Equation 4] The word recognition determination result I is output as i which gives the maximum value of the above L (i _k ) obtained by the equation (4).
Then, a desired word name can be output by supplying the word recognition determination result I to, for example, a table. For example, in this table, a plurality of word names (or category names, etc.) are stored in advance corresponding to the value of I, and by supplying any of the above I,
An arbitrary word name can be output according to the value of I.

Here, KK in the above equation 4 is the number of word notation candidates in the group in the near similarity table, and KK =
It is practical to set the value to 2 or 3. In addition, it is desirable that the weighting coefficients of w1 and w2 be set to the relationship of w2> w1 from the viewpoint of placing more importance on the partial similarity.

By the comprehensive judgment processing in the comprehensive judgment unit 12, different methods of the similarity between the phoneme HMM and the pattern matching are used complementarily, so that the discrimination improving effect for the similar word recognition is obtained. Is brought.

Next, the phonological HMM part 6 in FIG. 3 described above.
An example of a word connection method using 1 and the word dictionary portion 62 and the word connection model portion 63 will be described with reference to FIGS. 5 and 6.

FIG. 5 shows the phoneme HMM of the phoneme HMM part 61.
It is an example of a configuration of. FIG. 5A shows a phoneme HMM model with four states and three loops. FIG. 5B shows a phonological HMM model with two states and one loop. They are,
The state transition probability and the output probability are obtained from the learning sample in advance by, for example, the “Baum-Welch parameter estimation method” shown in the above-mentioned document (2), and the model is estimated and stored.

FIG. 6 shows a method of constructing an HMM of a recognized word. The phoneme HMM is connected using the phoneme HMM part 61 and the word dictionary part 62. Here, as an example of the phoneme HMM, the four states 3 in FIG.
The loop model is used as an example. In the word dictionary portion 62, the recognition target word name is described in the orthography (phonetic symbol notation). The phoneme HMM corresponding to the phoneme symbol in the word dictionary part is selected, and the final state of the immediately preceding phoneme HMM (see FIG. 6).
(B)) and the initial state of the phonological HMM immediately after (FIG. 6).
(A)) can be simply connected to obtain a concatenation model of words. The connection method using the two-state one-loop model in FIG. 5 can be realized by the same method.

FIG. 7 is an explanatory diagram of a word model using a left-to-right (Left-to-right) HMM when each phoneme can be represented by one state.

In FIG. 7, FIG. 7A shows "bis
ai ”(Onishi: Bisai), state transition probabilities (a ₁₁ , a ₁₂ , ...) Between states S1 to S5 and code output probabilities (b _11k , b _12k , ...). .. Further, FIG. 7B shows a state transition probability (a) between states S1 ^{* to} S5 ^* for “hisai” (Hirai: Hisai).
_{1 1} ^* , a ₁₂ ^* , ...) And the code output probability (b
_11k ^* , b _12k ^* , ...). The above figures (a) and (b) are (a _{i, i} ) +
An example of the case of (a _{i, i + 1} ) = 1 is shown.

FIG. 8 shows an example of a concrete state of the signal of each component in the phoneme of the example of FIG.

In FIG. 8, each phoneme can be represented by one state, and as shown in FIGS. 7 (a) and 7 (b), "bisai" and "h" are displayed.
It is assumed that the word model “isai” is given in advance as an element of the word concatenation model part 63 (illustrated in FIG. 3). Then, at this time, it is assumed that the input voice is "Xisai" as shown in FIG. This "X" means indeterminacy. The input voice energy at this time is shown in FIG. 8B, and this voice section is shown in FIG.
Shown in. Then, the first and second candidate words of the calculation result of the likelihood calculating unit 6 are output as "Hirai" and "Onishi", and these candidate words are shown in the near similarity table 8 (illustrated in FIG. 1). ). At this time, the Viterbi comparison and collation part 64 (shown in FIG. 3) with respect to the word model of “Onishi” (shown in FIG. 7A) is processed by the state S and the input frame as shown in FIG. 8F. Correlation with T can be performed. By processing this state by the back trace unit 9 (shown in FIG. 1),
It is possible to obtain an optimum state transition sequence (however, the frame direction is thinned out) as shown in FIG. 8 (i). This Figure 8
The time change points time (1) to time (5) for "Onishi" in FIG. 8 (g) can be obtained from the optimum state series of (i).

In the same manner, backtrace processing is performed on the word model of "Hisui", and the optimum state transition sequence can be obtained as shown in FIG. 8 (j). This Figure 8
The time change points time (6) to time (10) for "Hisui" in FIG. 8 (h) can be obtained from the optimum state transition sequence in (j).

The above time change points time (L), L = 1
-10 and the value of the near similarity table 8 (shown in FIG. 1), ST (identification start frame) and ED (identification end frame) can be determined as described above. ..

FIG. 9 shows a specific calculation process flowchart of an example of ST (valid identification start frame) and ED (valid identification end frame).

In FIG. 9, first, WSTART
(I) * T = s (identification start frame) is calculated, and WEND
(I) * T = e (identification end frame) is calculated (S9)
1). WSTART (i) is the identification start rate, and T is the number of input frames or the input frame length. Next, let T be S
^*, And e ^* is set to e ^* = 1 (S92). Then L
= 1 is set (S93). Next, the above s ≦ time
It is determined whether (L) is satisfied (S94). If this condition is satisfied, then it is determined whether or not the above e satisfies e ≧ time (L) (S95). If this condition is satisfied then determines whether the ^{s *} satisfies ^{s * ≧ time (L) (} S96). If this condition is satisfied, then s ^{* is set} to s ^* = time (L) (S97). Next, it is determined whether or not the above e ^* satisfies e ^* ≤time (L) (S98). If this condition is satisfied, then e ^{* is changed} to e
^* = Time (L) is set (S99). Then L to L + 1
(S100). Next, this updated L is L>
It is determined whether LL is satisfied (S101). If this condition is satisfied, the value of s ^* is output as ST (valid identification start frame), and the value of e ^* is output as ED (valid identification end frame) (S102).

If it is determined in S94, S95, and S98 that the conditions are not satisfied, the process of S100 is performed. If it is determined that the condition is not satisfied in S96, the process of S98 is performed next. And the above S
The processes of 94 to S100 are continuously processed while the condition of S101 is not satisfied.

According to the above embodiment, at least one or more word candidates (i) are selected by the phoneme HMM concatenation model.
Is calculated, and if the candidate column is described in the close similarity table 8, an optimal state transition sequence of the candidate by the concatenation model of phonological HMM is obtained, and the feature time series is segmented, Since the identification effective section [ST, ED] corresponding to the different part of the feature is automatically extracted and the auxiliary feature of the identification effective section of the candidate is used for the recognition, the phoneme HMM and the pattern matching method are complemented. It is possible to output the accurate recognition judgment result even when there is a similar phoneme sequence word in the speech to be recognized by changing to the optimal likelihood, and to improve the recognition ability of the similar word speech. A recognition device can be realized.

In the acoustic feature extraction unit 3 of FIG. 1 of the above embodiment, the LPC cepstrum shown in the above-mentioned reference 3 is taken as an example, but the present invention is not limited to this.

Further, in the above embodiment, the optimal code vector selecting section 5 in FIG. 2 has been described by taking the clustering method shown in the above-mentioned document 4 as an example, but the present invention is not limited to this method.

Further, in the above embodiment, the configuration of the near similarity table 8 is configured as shown in FIG. 4, but it is not limited to this configuration.

Although the example of the phoneme HMM connection model has been described with reference to FIG. 6 of the above embodiment, the connection method is not limited to the above example.

Further, in FIG. 1 of the above-mentioned embodiment, the comprehensive judgment unit 12 outputs the word recognition judgment result number I as the judgment result, but instead of this, the I-th category name is the final one. You may output as a recognition determination result.

Further, although the calculation processing method of ST (effective identification start frame) and ED (effective identification end frame) has been described with reference to FIG. 9 of the above embodiment, the present invention is not limited to this processing procedure.

[0067]

As described in detail above, according to the present invention, the feature information sequence extracting means, the similar word storing means, the standard feature information storing means, and the extracting means are provided, and The characteristic information series of the optimal identification effective section,
Using the standard feature information, the likelihood is optimally changed and the optimum word identification result is generated and output. Therefore, even if a word of a similar phonological sequence exists in the recognition target speech signal, the word is accurately output. It is possible to realize a voice recognition device which can be recognized and whose recognition ability of similar word voice is further improved.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a voice recognition device according to an embodiment.

FIG. 2 is a functional block diagram of an optimum code vector selection unit of the voice recognition device according to the embodiment.

FIG. 3 is a functional block diagram of a likelihood calculation unit of the voice recognition device according to the embodiment.

FIG. 4 is a configuration diagram of a near similarity table of the voice recognition device according to the embodiment.

FIG. 5 is a phoneme HMM of the voice recognition device according to the embodiment.
FIG.

FIG. 6 is a phoneme HMM of the voice recognition device according to the embodiment.
It is a connection model figure of.

FIG. 7 is an explanatory diagram of a word model using a left-to-right HMM of the voice recognition device according to the embodiment.

FIG. 8 is an example of a signal state of each unit of the voice recognition device according to the embodiment.

FIG. 9: ST and ED of the voice recognition device according to this embodiment
It is a calculation processing flowchart.

[Explanation of symbols]

3 ... Acoustic feature extraction unit, 4 ... Auxiliary feature extraction unit, 5 ... Optimal code vector selection unit, 6 ... Likelihood calculation unit, 7 ... Judgment unit, 8
... near similarity table, 9 ... back trace part, 10 ... partial similarity calculation part, 11 ... auxiliary feature standard pattern part, 12
... Comprehensive judgment part, 51 ... Code vector selection parts 51,
2 ... Code vector part, 61 ... Phonological HMM part, 62
... word dictionary part, 63 ... word connection model part, 64 ... Viterbi comparison and collation part.

Claims (1)

[Claims] 1. A speech recognition apparatus which performs a speech recognition by obtaining a likelihood of a plurality of candidate words corresponding to an input speech signal by using a hidden Markov model method corresponding to an input speech signal, and a similar word is previously prepared. A similar word storage means for storing the grouped groups, and a standard feature information storage means for storing at least one type of standard feature information of the spectrum envelope corresponding to the recognizable words in advance, and a spectrum envelope of the input speech signal. The same type of feature information as the standard feature information is extracted, the feature information sequence extraction means for serializing and outputting the feature information, and the similar word storage means is searched from the candidate word and its likelihood, and An extraction unit that performs backtrace processing according to the search result to extract the optimum identification effective section, and the feature information series of the optimum identification effective section and the standard feature information. DOO using speech recognition apparatus characterized by generating and outputting an estimated word identification result optimally changes the likelihood.