JP2001083978A - Voice recognition device - Google Patents

Abstract

(57) [Summary] [Problem] An input speech has a part with high temporal continuity like a vowel and a part with low temporal continuity like a consonant. By making the analysis window length variable according to such characteristics of the input speech, a speech recognition device with higher recognition performance than before can be provided without increasing the recognition processing time. SOLUTION: A voice analysis unit 1 receives an input voice.
Numerals 03 and 104 convert the speech into feature vectors with different analysis window lengths at each time. The feature vector correlation calculation unit 105 calculates a correlation value between each feature vector at each time. An optimal feature vector is selected from feature vectors calculated with different analysis window lengths according to the correlation value between the feature vectors, and the speech is recognized by the HMM method using the optimal feature vector.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a microcomputer, an electronic computer, or the like using a time series of a feature vector of an input speech and a time series of a feature vector of a reference speech or a statistical model thereof. The present invention relates to a speech recognition device that performs automatic speech recognition by using the following.

[0002]

2. Description of the Related Art In a voice recognition apparatus, it is common to convert voice input from a microphone or the like into a time series of spectrum and recognize it. The most basic feature vector representing a spectrum is a cepstrum, which is defined by a Fourier transform of a logarithmic spectrum (hereinafter simply referred to as a spectrum). In addition, mel cepstrum that is a cepstrum made into a mel scale,
The output of FFT or BPF (bandpass filter) is also used as a feature vector for speech recognition.

Further, in order to improve the recognition rate in recent years, a delta cepstrum [S. Fu
rui: "Speaker-Independent Isolated Word Recognit
ionUsing Dynamic Features of Speech Spectrum ", IEE
E Trans., ASSP-34, No.1, pp.52-59, (1986-2)] and weighted cepstrum [Y.Tohk
ura: "A Weighted Cepstral Distance Measure for Sp
eech Recognition ", IEEE Trans., ASSP-35, No.10, p
p.1414-1422, (1987-10)].

However, when these conventional feature vectors are summarized in terms of the analysis window length, the cepstrum and the weighted cepstrum are calculated with a predetermined always constant analysis window length, and the delta cepstrum is also predetermined. It is calculated as the number of linear regression coefficients (at least two frames or more of cepstrum), and these feature vectors are used to perform speech analysis with a constant analysis window length regardless of the properties of the input speech.

A method has also been proposed in which a plurality of linear regression coefficients are combined into one to form a feature vector having a large vector dimension. However, this method requires a processing time corresponding to the plurality of linear regression coefficients. Not suitable for building real-time systems.

FIG. 7 is a block diagram of a conventional speech recognition apparatus using a cepstrum calculated from a fixed analysis window length. In FIG. 7, a sound input to a microphone 700 is cut by a low-pass filter 701 at a frequency component equal to or more than サ ン プ リ ン グ of a sampling frequency, and then a sampling frequency (for example, 8kHz) and converted to a digital signal. The digital signal is converted into a time series of feature vectors (cepstrum) by a voice analysis unit 703, and a model of each recognition target vocabulary stored in an HMM storage unit 707 and a time series of feature vectors are converted by an HMM recognition unit 704. Is calculated, and the recognition result determining unit 706 determines the speech having the highest similarity to the input voice as the recognition result.

[0007]

However, the input voice has a portion with high temporal continuity, for example, a stationary time of about 100 msec, such as a vowel, and a low temporal continuity, such as a consonant or a semi-vowel. There is a part. In the past, these were analyzed using the same analysis window length, so that the window resolution suitable for vowels resulted in lower temporal resolution of consonant parts, and conversely, the window resolution suitable for consonants resulted in lower frequency resolution in vowel parts. There is a drawback that the recognition performance deteriorates. Further, when feature vectors that have been subjected to voice analysis in advance with a plurality of analysis window lengths are directly combined, there is a disadvantage that the information becomes enormous and the processing at the time of recognition is greatly increased.

Accordingly, an object of the present invention is to make the length of the analysis window variable according to the characteristics of the input speech without significantly increasing the recognition processing time. An object of the present invention is to provide a speech recognition method having a higher recognition performance than conventional ones. The invention of claim 4 of the present application
An object of the present invention is to provide a speech recognition device with high recognition performance by emphasizing a feature vector obtained from features of input speech.

[0009]

According to a first aspect of the present invention, there is provided a voice input means for inputting voice, and at least a voice input means for converting the input voice into a feature vector with a different analysis window length at each time. Two voice analysis means, a feature vector correlation calculation means for calculating at least one of a correlation value and a similarity between feature vectors output from the voice analysis means, and a feature vector correlation value calculation means. If the correlation value or similarity is large from among the feature vectors calculated by the voice analysis means according to the correlation value or similarity, the feature vector having a large analysis window length is assumed to have a small change in the input voice. When the similarity is small, it is assumed that the input speech changes largely, and a feature vector selecting means for selecting a feature vector having a small analysis window length (hereinafter referred to as a selected feature vector); Standard pattern storage means for storing a model for the target vocabulary; recognition means for calculating the likelihood between the time series of the selected feature vector and the standard pattern stored in the standard pattern storage means; and an output of the recognition means. A recognition result determining unit that determines a recognition result based on the likelihood of each of the standard patterns.

According to the first aspect of the present invention having such a feature, when the correlation value or the similarity is large, the change of the input voice is assumed to be small, and the feature vector having a large analysis window length is converted to the correlation value or the similarity. Is small, the change in the input voice is large, and a small analysis window length can be selected. As a result, an analysis window length suitable for input speech can be selected, and recognition performance can be improved.

According to a second aspect of the present invention, there is provided a voice input means for inputting voice, a first voice analysis means for converting the input voice into a feature vector with a predetermined analysis window length, and at least one A buffer for storing at least one past time feature vector; and at least one of a correlation value and a similarity between the feature vector output from the first speech analysis means and the past feature vector stored in the buffer. Between the feature vector calculating means for calculating the correlation value or the similarity and a predetermined first threshold value. If the correlation value is equal to or larger than the threshold value, the analysis window length is longer than the predetermined analysis window length. An analysis window length setting means for comparing an analysis window length shorter than the predetermined analysis window length with a predetermined second threshold value which is smaller than the first threshold value and being smaller than the threshold value, Analysis window length Second speech analysis means for converting the input speech into a feature vector using the analysis window length set by the means, standard pattern storage means for storing a model for each vocabulary to be recognized, and second speech analysis means Recognition means for calculating the likelihood between the obtained sequence of feature vectors and the standard pattern stored in the standard pattern storage means, and a recognition result from the likelihood of each standard pattern output from the recognition means. And a recognition result judging means for judging.

According to the second aspect of the present invention having such a feature, the optimum analysis window length calculated by the analysis window length setting means is determined as a function of the correlation value or the similarity, so that the feature vector It is possible to calculate a large analysis window length as the correlation value or similarity is larger, as the change of the input voice is smaller, and a smaller analysis window length as the correlation value or similarity is smaller, as the change of the input voice is larger. As a result, an analysis window length suitable for input speech can be selected, and recognition performance can be improved.

According to a third aspect of the present invention, there is provided voice input means for inputting voice, first voice analysis means for converting the input voice into a feature vector with a predetermined analysis window length, and A buffer for storing at least two or more feature vectors obtained by the analysis means; a linear regression coefficient calculation means for calculating a linear regression coefficient from the feature vectors stored in the buffer; Comparing the analysis window length longer than a predetermined analysis window length with a predetermined second threshold value larger than the first threshold value. Analysis window length setting means for setting an analysis window length shorter than the predetermined analysis window length when the threshold value is equal to or more than the threshold value; and analyzing the input speech using the analysis window length set by the analysis window length setting means. Convert to And second sound analysis unit, and the standard pattern storage means for storing a model for each vocabulary to be recognized,
A recognition unit for calculating a likelihood between the series of feature vectors obtained by the second speech analysis unit and a standard pattern stored in the standard pattern storage unit; and a standard pattern output from the recognition unit. And a recognition result determining means for determining a recognition result from the likelihood of the recognition result.

According to the third aspect of the present invention having such features, the linear regression coefficient is small because the optimal analysis window length calculated by the analysis window length setting means is determined as a function of the linear regression coefficient. A larger analysis window length can be calculated as the change in the input voice becomes smaller as the change in the input voice decreases, and a smaller analysis window length can be calculated as the change in the input voice increases as the linear regression coefficient increases. As a result, an analysis window length suitable for input speech can be selected, and recognition performance can be improved.

According to a fourth aspect of the present invention, there is provided a voice input means for inputting voice, a voice analysis means for converting the input voice into a feature vector with a predetermined analysis window length, and at least one or more voice input means. A buffer for storing a feature vector at a past time, and a feature vector for calculating at least one of a correlation value and a similarity between a feature vector output from the speech analysis unit and a past feature vector stored in the buffer. Correlation calculation means, an enhancement coefficient calculation means for comparing the correlation value or the similarity with a predetermined threshold value and calculating an enhancement coefficient for enhancing the feature vector when the correlation value or the similarity is equal to or less than a threshold value; Voice emphasizing means for replacing with a feature vector emphasized by the emphasis coefficient,
Standard pattern storage means for storing a model for each vocabulary to be recognized; recognition means for calculating the likelihood between the time series of feature vectors obtained by the voice emphasis means and the standard pattern stored in the standard pattern storage means; And a recognition result determining means for determining a recognition result from a likelihood of each standard pattern output from the recognition means.

According to the fourth aspect of the present invention having such a feature, the emphasis coefficient of the emphasis coefficient calculating means is calculated as a function of the correlation value or the similarity. The emphasis coefficient is increased on the assumption that the input voice changes largely. As a result, a feature vector suitable for input speech can be obtained, and recognition performance can be improved.

According to a fifth aspect of the present invention, there is provided a voice input means for inputting voice, and at least converting the input voice into a pitch frequency (fundamental frequency) of voice with a different analysis window length at each time. Two pitch extracting means, an inter-pitch correlation calculating means for calculating at least one of a correlation value and a similarity between pitch frequencies output from the pitch extracting means, and the pitch extracting means according to the correlation value or the similarity When the correlation value or similarity is large from among the analysis window lengths used for the means, a large analysis window length is assumed to have a small change in the input voice, and when the correlation value or similarity is small, a large change in the input voice is assumed. Analysis window length setting means for setting a small analysis window length; speech analysis means for converting the input speech into a feature vector using the analysis window length set by the analysis window length setting means; Standard pattern storage means for storing a model for a vocabulary; recognition means for calculating a likelihood between a sequence of feature vectors obtained by the voice analysis means and a standard pattern stored in the standard pattern storage means; And a recognition result determining means for determining a recognition result from the likelihood of each standard pattern output from the means.

According to the fifth aspect of the present invention having such a feature, the optimum analysis window length calculated by the analysis window length setting means is determined based on the correlation value or the similarity, whereby the pitch frequency The larger the correlation value or similarity, the larger the analysis window length as the pitch frequency of the input voice does not depend on the analysis window length, that is, the smaller the change in the input voice in the steady part of the vowel, and the smaller the correlation value or similarity. It is possible to calculate a small analysis window length assuming that there is a large change in the input voice at the transition between consonants and phonemes. As a result, an analysis window length suitable for input speech can be selected, and recognition performance can be improved.

According to a sixth aspect of the present invention, there is provided a voice input means for inputting voice, and a pitch extracting means for converting the input voice into a voice pitch frequency (basic frequency) with a predetermined analysis window length. A buffer for storing at least one or more past time pitch frequencies obtained by the pitch extracting means, and a correlation between a pitch frequency output from the pitch extracting means and a past pitch frequency stored in the buffer. A pitch-to-pitch correlation calculating means for calculating at least one of the value and the similarity, and comparing the correlation value or the similarity with a predetermined first threshold, and when the correlation is equal to or greater than the first threshold, the predetermined correlation is determined. An analysis window length longer than the analysis window length is compared with a predetermined second threshold value smaller than the first threshold value, and if the analysis window length is equal to or smaller than the second threshold value, the analysis window length shorter than the predetermined analysis window length is determined. Analysis window length setting means for setting the analysis window length, speech analysis means for converting the input speech into a feature vector using the analysis window length set by the analysis window length setting means, and a standard for storing a model for each vocabulary to be recognized. Pattern storage means; recognition means for calculating the likelihood between the sequence of feature vectors obtained by the voice analysis means and the standard pattern stored in the standard pattern storage means; and each standard output from the recognition means. And a recognition result determining means for determining the recognition result from the likelihood of the pattern.

According to the sixth aspect of the present invention having such a feature, the optimum analysis window length calculated by the analysis window length setting means is determined based on the correlation value or the similarity, whereby the pitch frequency The larger the correlation value or similarity is, the smaller the change in the pitch frequency of the input voice between adjacent times, that is, the larger the analysis window length as the change in the input voice is small in the steady part of the vowel, and the larger the correlation value or similarity. It is possible to calculate a small analysis window length on the assumption that the smaller the is, the more the input voice changes in the consonant or phoneme transition part. As a result, an analysis window length suitable for input speech can be selected, and recognition performance can be improved.

[0021]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to FIGS. FIG. 1 is a block diagram showing the configuration of the speech recognition device according to the first embodiment of the present invention. In the present embodiment, a case where there are two voice analysis units will be described. In FIG. 1, a microphone 100 is for inputting voice, and the input voice (hereinafter, referred to as input voice) is input to a low-pass filter 101. The low-pass filter 101 cuts a frequency component equal to or more than の of the sampling frequency, and its output is provided to an A / D converter 102. A /
The D conversion unit 102 samples the input signal at a predetermined sampling frequency (for example, 8 kHz) and converts it into a digital signal. Microphone 100,
The low-pass filter 101 and the A / D converter 102 constitute an audio input unit for inputting audio. The first speech analysis unit 103 converts the digital signal into a first feature vector (cepstrum) using a predetermined analysis window length 1. The second speech analysis unit 104 converts the digital signal into a second feature vector (cepstrum) using a predetermined analysis window length 2 different from the analysis window length 1. The feature vector correlation calculator 105 calculates a correlation value or a similarity between the first and second feature vectors. The feature vector selection unit 106 selects a feature vector having an optimum analysis window length from two feature vectors based on the correlation value or similarity as a selected feature vector. The HMM storage unit 108 is a standard pattern storage unit that stores a model for a vocabulary to be recognized.
The recognition unit 107 is a recognition unit that calculates the likelihood between the model of each vocabulary to be recognized stored in the HMM storage unit 108 and the selected feature vector sequence. The recognition result determination unit 109 is a recognition result determination unit that determines the one with the highest likelihood as a recognition result.

In the speech recognition apparatus of the present invention configured as described above, the case where the correlation value or similarity is large is considered as an example when the determination of the optimal analysis window length selection is considered as the magnitude of the correlation value or similarity. The operation will be described separately for the case of the small case. However, the analysis window lengths 1 and 2 of the voice analysis units 103 and 104
Is described here as W and twice as much as 2W. (Step 1-1) Digital signal U = ｛u ₁ , u ₂ , u ₃ ,..., U sampled by A / D conversion section 102
_T } is converted into a first feature vector sequence A = {a ₁ , a ₂ ,..., A _i,.
.., A _I }. In addition, the analysis window length 2W (2 of W)
Feature vector series of the second voice analysis unit 104 of the second fold) B = convert _{{b 1, b 2, ···} , b j, ··· b J} in.

(Step 1-2) The correlation value r (a _i , b _j ) between the feature vectors of the first feature vector sequence A and the second feature vector sequence B at the same time is calculated according to, for example, Equation 1. .

(Equation 1) Here, s (a, b) indicates the variance of a and b. Further, in the case of the similarity h (a _i , b _j ), the similarity is calculated according to Equation 2.

(Equation 2) The magnitude of the correlation value or the similarity is considered to have the following meaning. The fact that the correlation value or similarity is large means that the same feature vector is calculated even if the voice at the same time is analyzed using the long analysis window length 2W and the short analysis window length W. (Vowel part). Conversely, when the correlation value or similarity is small, it can be said that the consonant part has a large frequency change.

(Step 1-3) (When the correlation value or similarity is larger than a predetermined threshold)
If the correlation value r (a _i , b _j ) or the similarity h (a _i , b _j ) is larger than a predetermined threshold value, the feature vector selection unit 106 may select the longer analysis window length 2W. The feature vector having a length of 2 W is selected as the optimal feature vector at the current time. (When the correlation value or similarity is smaller than a predetermined threshold)
On the other hand, the correlation value r (a _i , b _j ) or the similarity h (a _i , b
_{If j} ) is smaller than the predetermined threshold value, a short analysis window length W may be selected, and the feature vector selection unit 106 selects the feature vector of the analysis window length W as the optimal feature vector at the current time. As a result, the optimum analysis window length for the vowel / consonant is automatically selected.

(Step 1-4) The Hidden Markov Model λ of the vocabulary to be recognized stored in the HMM storage unit 108 and the optimal feature vector sequence C obtained as described above
The HMM recognition unit 107 calculates the likelihood L (w) with (w) according to Equation 3.
Is calculated by

(Equation 3)

(Step 1-5) When the likelihood L (w) of each vocabulary to be recognized is calculated, the recognition result judgment unit 109 calculates
Is determined according to the following.

(Equation 4) From the above (Step 1-1) to (Step 1-
According to 5), it is possible to realize speech recognition with high recognition accuracy without making the analysis window length variable and significantly increasing the recognition processing time.

Although this embodiment has been described in connection with the case where there are two voice analysis units, the feature vector sequence from the voice analysis unit is used by associating the correlation value or similarity with a plurality of voice analysis units. May be selected.

Next, FIG. 2 is a block diagram showing a configuration of a speech recognition apparatus according to a second embodiment of the present invention. In the present embodiment, a case will be described in which only one past feature vector is stored in the buffer one time earlier. In FIG. 2, a microphone 200 inputs voice, and the input voice (hereinafter, referred to as input voice) is input to a low-pass filter 201. Low-pass filter 20
Reference numeral 1 denotes a component for cutting a frequency component equal to or more than の of the sampling frequency, and its output is supplied to the A / D converter 202. The A / D converter 202 samples an input signal at a predetermined sampling frequency (for example, 8 kHz) and converts it into a digital signal. The microphone 200, the low-pass filter 201, and the A / D conversion unit 202 constitute a voice input unit for inputting voice. The first speech analysis unit 203 converts this digital signal into a feature vector (cepstrum) using a predetermined analysis window length. The buffer 204 stores the feature vector one time ago. A feature vector correlation calculator 205 calculates a correlation value or similarity between the two vectors from the feature vector calculated by the speech analyzer 203 and the feature vector one time ago stored in the buffer 204. It is. Analysis window length setting unit 2
06 is set to an analysis window length longer than the predetermined analysis window length of the voice analysis unit 203 when the correlation value or the similarity is larger than the predetermined first threshold value, and is set to be larger than the first threshold value. If it is smaller than the second predetermined small threshold value, the analysis window length is set to be shorter than the predetermined analysis window length of the voice analysis unit 203. Then, the second speech analysis unit 207 uses the analysis window length (hereinafter, referred to as an optimum analysis window length).
The voice analysis is performed again from the digital signal output from the A / D conversion unit 202, and its characteristic vector (hereinafter, referred to as an optimum characteristic vector) is output. The likelihood with each recognition target vocabulary model stored in the HMM storage unit 209 is calculated by the HMM recognizing unit 208 with respect to this optimal feature vector sequence, and the recognition result determining unit 210 calculates the maximum likelihood. Are determined as recognition results.

The operation of the speech recognition apparatus of the present invention configured as described above will be described separately for cases where the correlation value or similarity is large and cases where it is small. However, description will be made assuming that the analysis window length of the voice analysis unit 203 is W. (Step 2-1) Digital signal U = ｛u ₁ , u ₂ , u ₃ ,..., U sampled by A / D conversion section 202
_T } is converted into a feature vector a _i by the speech analysis unit 203 having the analysis window length W. The buffer 204 stores a feature vector sequence a _i-1 one time before.

(Step 2-2) The feature vectors a _i and 1
The correlation value r (a _i , a _i-1 ) between the feature vectors of the feature vector a _i-1 before the time is calculated, for example, according to Equation 5.

(Equation 5) Here, s (a, a) indicates the variance of a and a. In the case of the similarity h (a _i , a _i−1 ), the similarity is calculated in accordance with Equation 6.

(Equation 6) The magnitude of the correlation value or the similarity is considered to have the following meaning. A large correlation value or similarity means that there is a correlation between adjacent feature vectors and a stationary part (vowel part) with little change in frequency. Conversely, a small correlation value or similarity means that the correlation value or similarity is small. It can be said that the consonant part has a lot of dynamic changes.

(Step 2-3) Here, the analysis window length W is converted into W 'by the following equation. r (a _i , a _i-1 ) or h
_Assuming that (a _i , a _i-1 ) is x, the function f (x) used here is a function for determining a multiple to W depending on x.

(Equation 7) For example, f (x) = (x−Th1) +1 (x ≧ Th1) f (x) = exp (x−Th2) (x ≦ Th2) f (x) = 1 (Th2 <x <Th1) If the value or similarity is greater than the first threshold Th1, a function that returns a value that makes the analysis window length W larger, and if the value or the similarity is smaller than the second threshold Th2, a function that returns a value that makes the analysis window length smaller, I do. When a value between the threshold value Th1 and the threshold value Th2 is taken, 1 is returned. (When the correlation value or similarity is larger than the first threshold Th1) The correlation value r ( _ai , _ai-1 ) or the similarity h ( _ai , a)
_{If i-1} ) is larger than Th1, the analysis window length W 'larger than W is calculated and set according to equation (7). (When the correlation value or similarity is smaller than the second threshold value Th2) On the other hand, the correlation value r (a _i , a _i−1 ) or the similarity h (a _i
, a _i-1 ) is smaller than the second threshold value Th2, a short analysis window length W ′ is calculated and set using Equation 7.

(Step 2-4) Using the second analysis window length W 'obtained as described above, the second speech analysis unit 207 calculates a feature vector again, and the sequence D of the feature vector and the HMM Likelihood L ′ (w) of each vocabulary to be recognized stored in storage unit 209 with hidden Markov model λ (w)
Is calculated by the HMM recognizing unit 208 according to Equation 8.

(Equation 8)

(Step 2-5) After the likelihood L '(w) of each vocabulary to be recognized has been calculated, the recognition result determination unit 210 determines the recognition result y' according to equation (9).

(Equation 9) As described above, from (Step 2-1) to (Step 2-
By making the analysis window length variable according to 5), a speech recognition method with high recognition accuracy can be realized without significantly increasing the recognition processing time. The function f (x) is not limited to the function shown here, and any function having such a tendency can be used.

FIG. 3 is a block diagram showing a configuration of a speech recognition apparatus according to a third embodiment of the present invention. In FIG. 3, a microphone 300 is for inputting voice, and the input voice (hereinafter, referred to as input voice) is input to a low-pass filter 301. Low-pass filter 3
01 is for cutting frequency components equal to or more than の of the sampling frequency.
Given to. The A / D converter 302 samples an input signal at a predetermined sampling frequency (for example, 8 kHz) and converts it into a digital signal. The microphone 300, the low-pass filter 301, and the A / D converter 302 constitute a voice input unit for inputting voice. The first speech analysis unit 303 converts this digital signal into a feature vector (cepstrum) using a predetermined analysis window length. The buffer 304 stores at least two or more feature vectors for the linear regression coefficient calculation unit 305 to calculate the linear regression coefficient. The analysis window length setting unit 306 includes a linear regression coefficient calculation unit 305.
If the linear regression coefficient of the feature vector calculated in step (1) is smaller than a predetermined first threshold Th3, the speech analysis unit 30
3 is longer than the predetermined analysis window length W.
And a second predetermined threshold value Th4 (> Th
If it is larger than 3), the analysis window length W "is set to be shorter than the predetermined analysis window length of the speech analysis unit 303. The second speech analysis unit 307 sets the analysis window length W".
The voice analysis is performed again from the digital signal output from the A / D conversion unit 302 at. Then, the HMM recognizing unit 308 applies the HM
The likelihood of each recognition target vocabulary stored in the M storage unit 309 with the model is calculated, and the recognition result determination unit 310 determines the input speech having the highest likelihood as the recognition result.

The operation of the speech recognition apparatus of the present invention configured as described above will be described. However, the description is made assuming that the analysis window length of the voice analysis unit 303 is W. (Step 3-1) Digital signal U = {u ₁ , u ₂ , u ₃ ,..., X sampled by A / D conversion section 302
_T } is converted into a feature vector a _i by the speech analysis unit 303 having the analysis window length W. The buffer 304 stores at least two or more feature vectors.

(Step 3-2) The linear regression coefficient Δa of the feature vector is calculated by, for example, Expression 10.

(Equation 10) The magnitude of this linear regression coefficient is considered to have the following meaning. The fact that the linear regression coefficient is smaller than the predetermined first threshold Th3 is a steady part (vowel part) in which the variation of the neighboring feature vector is small and the frequency is not changed. Determined second threshold value Th4
If it is larger, it can be said that the consonant part in which the neighboring feature vector fluctuates greatly and the frequency changes frequently.

(Step 3-3) The analysis window length W is converted to W ″ using the following equation (11).

[Equation 11] g (Δa) is a function that determines a multiple to W depending on Δa. For example, g (Δa) = (Th3-Δa) +1 (Δa ≦ Th3) g (Δa) = exp (Th4-Δa) (Δa ≧ Th4) g (Δa) = 1 (Th3 <Δa <Th4) If the regression coefficient Δa is smaller than the threshold Th3, a value that makes the analysis window length W larger is returned.
If h4 is larger than h4, the function returns a value that reduces the analysis window length. When a value between the threshold value Th3 and the threshold value Th4 is taken, 1 is returned. (When the linear regression coefficient is smaller than a predetermined first threshold value Th3) When the linear regression coefficient Δa is small, the analysis window length W ″ larger than W is calculated and set according to Expression 11. (Linear regression) On the other hand, when the coefficient is larger than the second threshold Th4, the small analysis window length W ″ is calculated by the following equation (1).
Calculate and set using 1.

(Step 3-4) Using the second analysis window length W ″ obtained as described above, the second speech analysis unit 307 calculates a feature vector again, and the feature vector sequence E and the HMM Likelihood L ″ (w) of each vocabulary to be recognized stored in storage unit 309 with hidden Markov model λ (w)
Is calculated by the HMM recognizing unit 308 according to Equation 12.

(Equation 12)

(Step 3-5) When the likelihood L ″ (w) of each vocabulary to be recognized is calculated, the recognition result determination unit 310 determines the recognition result y ″ according to Expression 13.

(Equation 13) From the above (Step 3-1) to (Step 3-
As shown in 5), a speech recognition method with high recognition accuracy can be realized without greatly increasing the recognition processing time by making the analysis window length variable. Note that the function g (Δa) is not limited to the function shown here, and any function having such a tendency can be used.

FIG. 4 is a block diagram showing a configuration of a voice recognition device according to a fourth embodiment of the present invention. In the present embodiment, the past feature vector stored in the buffer is 1
The case of only one before the time will be described. In FIG. 4, a microphone 400 is for inputting voice, and the input voice (hereinafter, referred to as input voice) is input to a low-pass filter 401. The low-pass filter 401 cuts a frequency component equal to or more than の of the sampling frequency, and its output is provided to the A / D converter 402. The A / D converter 402 samples the input signal at a predetermined sampling frequency (for example, 8 kHz),
It is to be converted into a digital signal. Microphone 400, low-pass filter 401, A / D converter 402
Constitutes voice input means for inputting voice. First
The voice analysis unit 403 converts the digital signal into a feature vector (cepstrum) using a predetermined analysis window length. The buffer 404 stores the feature vector one time ago, and the inter-feature-vector correlation calculation unit 405 calculates the feature vector calculated by the speech analysis unit 403 and the feature vector one time ago stored in the buffer 404. The correlation value or similarity between these two vectors is calculated. The voice emphasis coefficient calculation unit 406 sets an emphasis coefficient to emphasize the feature vector only when the correlation value or the similarity is smaller than a predetermined threshold. The speech enhancement unit 407 emphasizes the feature vector, and the HMM recognition unit 4
At 08, the likelihood of each vocabulary to be recognized stored in the HMM storage unit 409 with the model is calculated, and the recognition result determination unit 2
In step 10, the input speech having the highest likelihood is determined as the recognition result.

The operation of the speech recognition apparatus of the present invention configured as described above will be described. (Step 4-1) Digital signal U = ｛u ₁ , u ₂ , u ₃ ,..., U sampled by A / D conversion section 402
_T } is converted into a feature vector a _i by the speech analysis unit 403 of the analysis window length W. The buffer 404 stores a feature vector sequence a _i-1 one time before.

(Step 4-2) The feature vectors a _i and 1
The correlation value r (a _i , a _i-1 ) between the feature vectors of the feature vector a _i-1 before the time is calculated according to, for example, Expression 14.

[Equation 14] Here, s (a, a) indicates the variance of a, a. In the case of the similarity h (a _i , a _i-1 ), the similarity is calculated according to Equation 15.

(Equation 15) The magnitude of the correlation value or the similarity is considered to have the following meaning. If the correlation is large, it is considered to be a vowel part with little frequency change and need not be emphasized. On the other hand, a small correlation value or similarity is considered to be a consonant part with many frequency changes, so it is emphasized. Thereby, the recognition performance can be improved.

(Step 4-3) Correlation value r (a _i , a _i-1
) Or only when the similarity h (a _i , a _i−1 ) is small, the speech emphasis coefficient Z is calculated and set by using Expression (16).

(Equation 16) Assuming that r (a _i , a _i-1 ) or h (a _i , a _i-1 ) is x, e (x) is a function for calculating a speech enhancement coefficient depending on x. For example, when the correlation value or the similarity is smaller than the threshold value Th such that e (x) = (x−Th) +1 (x ≦ Th) e (x) = 1 (Th <x), the feature vector of each frame is set as A threshold value for returning a value to be emphasized. If the correlation value or similarity is equal to or greater than the threshold Th, 1 is returned. Speech emphasis is performed using the calculated speech emphasis coefficient Z as follows. performing emphasis a _i as _{a i = a i + (a} i -a i-1) * Z is corresponding to the magnitude of the change direction changed compared to a _i-1.

(Step 4-4) The feature vector output from the speech analysis unit 403 is enhanced using the speech enhancement coefficient obtained as described above, and the sequence F of the feature vector is stored in the HMM storage unit 409. The likelihood L ′ ″ (w) of each vocabulary to be recognized with the hidden Markov model λ (w) is calculated by the HMM recognition unit 408 according to Expression 17.

[Equation 17]

(Step 4-5) After the likelihood L ′ ″ (w) of each vocabulary to be recognized has been calculated, the recognition result determining unit 410 determines the recognition result y ′ ″ according to equation (18).

(Equation 18) As described above, from (Step 4-1) to (Step 4-
According to 5), a speech recognition method with high recognition accuracy can be realized without making the analysis window length variable and significantly increasing the recognition processing time. Note that the function e (x) is not limited to the function shown here, and an arbitrary function having such a tendency can be used.

Next, FIG. 5 is a block diagram showing a configuration of a speech recognition apparatus according to a fifth embodiment of the present invention. In the present embodiment, a case where there are two pitch extraction units will be described. In FIG. 5, a microphone 500 is for inputting voice, and the input voice (hereinafter, referred to as input voice) is input to a low-pass filter 501. The low-pass filter 501 cuts a frequency component equal to or more than １ ／ of the sampling frequency, and its output is A / D
This is provided to conversion section 502. The A / D converter 502 converts the input signal to a predetermined sampling frequency (for example, 8 k
Hz) and converts it into a digital signal. The microphone 500, the low-pass filter 501, and the A / D converter 502 constitute a voice input unit for inputting voice. The first pitch extraction unit 503 converts this digital signal into a pitch frequency (also referred to as a fundamental frequency) of the first voice by using a predetermined analysis window length. The cepstrum method and the partial autocorrelation method are generally used for extracting the pitch frequency.
The second pitch extracting section 504 converts this digital signal into a pitch frequency (basic frequency) of the second voice by a predetermined analysis window length 2. The inter-pitch correlation calculation unit 505 calculates a correlation value or similarity between the first and second pitch frequencies. Analysis window length setting unit 50
6 selects the optimum analysis window length from the two analysis window lengths based on the correlation value or the similarity, and the speech analysis unit 507 uses the analysis window length (hereinafter, referred to as the optimum analysis window length) as A. The voice analysis is performed on the digital signal output from the / D conversion unit 502, and its feature vector (hereinafter, referred to as an optimum feature vector) is output. The HMM storage unit 509 is a standard pattern storage unit that stores a model for the vocabulary to be recognized. The HMM recognition unit 508 stores a model of each recognition target vocabulary stored in the HMM storage unit 509 and a sequence of optimal feature vectors. This is a recognition means for calculating likelihood. The recognition result determination unit 510 is a recognition result determination unit that determines the one with the highest likelihood as a recognition result.

In the speech recognition apparatus of the present invention configured as described above, the case where the correlation value or the similarity is large is considered as an example in the case where the determination of the optimum analysis window length is considered as the magnitude of the correlation value or the similarity. The operation will be described separately for the case of the small case. However, the analysis window length of the pitch extraction units 503 and 504 is 1,
Here, 2 is described as W and twice as much as 2W. (Step 5-1) Digital signal U = ｛u ₁ , u ₂ , u ₃ ,..., U sampled by A / D conversion section 502
The _T}, series C = {c ₁ of the first pitch frequency by a first pitch extraction unit 503 in the analysis window length W, c _2, ···, c
_i ,..., c _I }. In addition, the analysis window length 2W (W
), A second pitch frequency sequence D = ｛d ₁ , d ₂ ,..., D _j ,.
Convert to _J ｝.

[0048] (Step 5-2) The correlation value r (c _i, d _j) between the pitch frequency at the same time the first pitch frequency sequence C and a second pitch frequency sequence D, for example, be calculated according to Equation 19 .

[Equation 19] Here, s (c, d) indicates the variance of c and d. In the case of the similarity h (c _i , _dj ), the similarity is calculated according to Equation 20.

(Equation 20) The magnitude of the correlation value or the similarity is considered to have the following meaning. The fact that the correlation value or similarity is large means that the same pitch frequency is calculated even if the voice at the same time is analyzed using the long analysis window length 2W and the short analysis window length W. (Vowel portion), and conversely, a small correlation value or similarity can be regarded as a consonant portion with a large frequency change.

(Step 5-3) (When the correlation value or similarity is larger than a predetermined threshold)
Correlation value r (c _i, d _j) or similarity h (c _i, d _j) larger than the threshold which is determined in advance, so do I select a longer analysis window length 2W analysis window length setting unit 506 in the analysis The window length 2W is set as the optimum window length W '''. (When the correlation value or similarity is small) On the other hand, the correlation value r (c _i
, D _j ) or the similarity h (c _i , d _j ) is smaller than a predetermined threshold, the analysis window length setting unit 506 sets the analysis window length W to the optimal window. Length W '''. As a result, the optimum analysis window length for the vowel / consonant is automatically selected.

(Step 5-4) Using the optimum analysis window length W ′ ″ obtained as described above, the speech analysis unit 507 calculates a feature vector, and the feature vector sequences G and H
The likelihood L ″ ″ (w) of each vocabulary to be recognized stored in the MM storage unit 509 with the hidden Markov model λ (w) is calculated by the HMM recognition unit 508 according to Formula 21.

(Equation 21) (Step 5-5) When the likelihood L ″ ″ (w) of each vocabulary to be recognized is calculated, the recognition result determination unit 510 determines the recognition result y ″ ″ according to Expression 22.

(Equation 22) From the above (Step 5-1) to (Step 5-
By making the analysis window length variable according to 5), a speech recognition method with high recognition accuracy can be realized without significantly increasing the recognition processing time.

In this embodiment, the case where there are two pitch extraction units is described. However, the analysis window length of the pitch extraction units is made to correspond to the correlation value or similarity by using more pitch extraction units. You may make it set.

Next, FIG. 6 is a block diagram showing a configuration of a speech recognition apparatus according to a sixth embodiment of the present invention. In the present embodiment, a case will be described in which the past pitch frequency stored in the buffer is only one before one time. FIG.
In, the microphone 600 is for inputting voice, and the input voice (hereinafter referred to as input voice) is input to the low-pass filter 601. The low-pass filter 601 cuts a frequency component equal to or more than の of the sampling frequency.
2 given. The A / D converter 602 is for sampling an input signal at a predetermined sampling frequency (for example, 8 kHz) and converting it into a digital signal.
Microphone 600, low-pass filter 601, A /
The D conversion unit 602 constitutes a voice input unit for inputting voice. The pitch extracting unit 603 converts this digital signal into a pitch frequency (fundamental frequency) of a voice according to a predetermined analysis window length. For pitch extraction, a cepstrum method, a partial autocorrelation method, and the like are generally used. The buffer 604 stores the pitch frequency one time before. The pitch-to-pitch correlation calculation unit 605 calculates a correlation value or similarity between the two frequencies from the pitch frequency, which is the calculation result of the pitch extraction unit 603, and the pitch frequency one time before stored in the buffer 604. is there. The analysis window length setting unit 606 sets the analysis window length longer than the predetermined analysis window length of the pitch extraction unit 603 when the correlation value or the similarity is larger than a predetermined first threshold value, If it is smaller than the second predetermined threshold smaller than the first threshold, the analysis window length is set to be shorter than the predetermined analysis window length of the pitch extraction unit 603. Then, the speech analysis unit 607 uses the A / D conversion unit 602 based on the analysis window length (hereinafter, referred to as an optimum analysis window length).
The voice analysis is performed from the digital signal which is the output of, and its feature vector (hereinafter, referred to as an optimum feature vector) is output. The HMM storage unit 609 is a standard pattern storage unit that stores a model for a vocabulary to be recognized. The HMM recognition unit 608 stores a model of each recognition target vocabulary stored in the HMM storage unit 609 and a sequence of optimal feature vectors. This is a recognition means for calculating likelihood. The recognition result determination unit 610 is a recognition result determination unit that determines the one with the highest likelihood as a recognition result.

In the speech recognition apparatus of the present invention configured as described above, the case where the determination of the optimal analysis window length selection is considered to be the magnitude of the correlation value or the similarity is taken as an example. The operation will be described separately for the case of the small case. However, the description will be made on the assumption that the analysis window length of the pitch extraction unit 603 is W. (Step 6-1) Digital signal U = ｛u ₁ , u ₂ , u ₃ ,..., U sampled by A / D conversion section 702
_T ｝ is converted into a pitch frequency c _i by the pitch extraction unit 603 of the analysis window length W. The buffer 604 stores a pitch frequency sequence ci _-1 one time before.

(Step 6-2) Pitch frequency sequence c _i
If one time before the pitch frequency c _i-1 of the correlation value _{r (c i, c i-} 1) between the pitch frequency, for example, be calculated according to Equation 23.

(Equation 23) Here, s (c, c) indicates the variance of c and c. In the case of the similarity h (c _i , c _i−1 ), the similarity is calculated according to Equation 24.

(Equation 24) The magnitude of the correlation value or the similarity is considered to have the following meaning. A large correlation value or similarity means that there is a correlation between adjacent pitch frequencies and a steady portion (vowel portion) with little change in frequency. Conversely, a small correlation value or similarity means that the frequency is small. It can be said that the consonant part has many dynamic changes.

(Step 6-3) Here, it is converted into the analysis window length W by the following equation. Correlation value _{r (c i, c i-} 1) or the similarity _{h (c i, c i-} 1) When the x, multiples function f (x) is to W, depending on the x used here The function to determine.

(Equation 25) For example, f (x) = (x−Th1) +1 (x ≧ Th1) f (x) = exp (x−Th2) (x ≦ Th2) f (x) = 1 (Th2 <x <Th1) If the correlation value or similarity is greater than the first threshold Th1, a function that returns a value that increases the analysis window length W is returned. If the correlation value or the similarity is smaller than the second threshold Th2, a function that returns a value that reduces the analysis window length is returned. And When a value between the threshold value Th1 and the threshold value Th2 is taken, 1 is returned. (If the correlation value or similarity is greater than a first threshold value Th1) correlation value _{r (c i, c i-} 1) or the similarity h (c _i, c
_{If i-1} ) is larger than the first threshold value Th1, the analysis window length is calculated and set using an analysis window length W '''' larger than W using Equation 25. (When the correlation value or similarity is smaller than the second threshold Th2)
On the other hand, the correlation value _{r (c i, c i-} 1) or the similarity h (c _i,
If c _i-1 ) is smaller than the second threshold value Th2, the analysis window length W ″ ″ shorter than W is calculated and set using Equation 25.

(Step 6-4) Using the second analysis window length W ″ ″ obtained as described above, the speech analysis unit calculates a feature vector, and the sequence H and HM of the feature vector are calculated.
The likelihood L ″ ″ (w) of each vocabulary to be recognized stored in the M storage unit 609 with the hidden Markov model λ (w) is calculated by the HMM recognition unit 608 according to Formula 26.

(Equation 26)

(Step 6-5) After the likelihood L ″ ″ (w) of each vocabulary to be recognized has been calculated, the recognition result determination unit 610 determines the recognition result y according to Equation 27.

[Equation 27]

By making the analysis window length variable from (Step 6-1) to (Step 6-5) as described above, a speech recognition method with high recognition accuracy can be realized without greatly increasing the recognition processing time. can do. Function f
(x) is not limited to the function shown here, and any function having such a tendency can be used.

[0059]

As described in detail above, the claims of the present application are described below.
According to the first to third and fifth and sixth aspects, speech can be analyzed using the optimum analysis window length. Therefore, the analysis window length can be automatically set by selecting a spectrum of a long analysis frame for a vowel with a high correlation and selecting a spectrum of a short analysis frame for a consonant with a low correlation. A recognition rate can be realized. Further, according to the invention of claim 4 of the present application, it is possible to obtain an optimum feature vector according to the feature of the input signal, and it is possible to improve the recognition rate.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a speech recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a speech recognition device according to a third embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a voice recognition device according to a fifth embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a voice recognition device according to a sixth embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a conventional speech recognition device.

[Explanation of symbols]

100, 200, 300, 400, 500, 600, 7
00 microphone 101, 201, 301, 401, 501, 601, 7
01 Low-pass filter 102, 202, 302, 402, 502, 602, 7
02 A / D converters 103, 104, 203, 303, 403, 703 Speech analyzers 105, 205, 405 Feature vector correlation calculator 106 Feature vector selectors 107, 208, 308, 408, 508, 608, 7
04 HMM recognition unit 108, 205, 209, 309, 409, 509, 6
09,705 HMM storage unit 109,210,310,410,510,610,7
06 recognition result determination unit 204, 304, 404, 604 buffer 206, 306, 506, 606 analysis window length setting unit 207, 307, 507, 607 voice analysis unit 305 linear regression coefficient calculation unit 406 voice enhancement coefficient calculation unit 407 voice enhancement Units 503, 504, 603 Pitch extraction unit 505, 605 Inter-pitch correlation calculation unit

Claims (6)

[Claims] A voice input unit for inputting voice; at least two voice analysis units for converting the input voice into feature vectors with different analysis window lengths at each time; Feature vector correlation calculating means for calculating at least one of a correlation value and a similarity between output feature vectors; and the voice analyzing means according to the correlation value or the similarity calculated by the feature vector correlation value calculating means. If the correlation value or similarity is large from among the feature vectors calculated in step (1), it is assumed that the change in the input voice is small.If the correlation value or similarity is small, the change in the input voice is large. A feature vector selecting means for selecting a feature vector having a small analysis window length (hereinafter, referred to as a selected feature vector); and a marker for storing a model for each vocabulary to be recognized. Pattern storage means; recognition means for calculating the likelihood between the time series of the selected feature vector and the standard pattern stored in the standard pattern storage means; and likelihood with each standard pattern output from the recognition means. And a recognition result determining means for determining a recognition result from the speech recognition device. 2. Voice input means for inputting voice, first voice analysis means for converting the input voice into a feature vector with a predetermined analysis window length, and at least one or more past time points A buffer for storing a feature vector, and a correlation between feature vectors for calculating at least one of a correlation value and a similarity between a feature vector output from the first speech analysis unit and a past feature vector stored in the buffer. Calculating means for comparing the correlation value or similarity with a predetermined first threshold, and when the correlation value or similarity is equal to or greater than the threshold, an analysis window length longer than the predetermined analysis window length, An analysis window length setting unit configured to set an analysis window length shorter than the predetermined analysis window length when the value is equal to or smaller than the predetermined second threshold value as compared with a small second threshold value; Minutes Second speech analysis means for converting the input speech into a feature vector using a window length; standard pattern storage means for storing a model for each vocabulary to be recognized; feature vector obtained from the second speech analysis means Recognition means for calculating the likelihood of the sequence of the reference pattern and the standard pattern stored in the standard pattern storage means, and recognition result determination means for determining a recognition result from the likelihood of each standard pattern output from the recognition means A voice recognition device comprising: 3. Voice input means for inputting voice, first voice analysis means for converting the input voice into a feature vector with a predetermined analysis window length, and at least two voices obtained by the voice analysis means. A buffer for storing one or more feature vectors, a linear regression coefficient calculating means for calculating a linear regression coefficient from the feature vectors stored in the buffer, and a first threshold in which the linear regression coefficient is predetermined. If the comparison is less than the threshold, the analysis window length longer than the predetermined analysis window length is compared with a predetermined second threshold larger than the first threshold. Analysis window length setting means for setting an analysis window length shorter than a predetermined analysis window length; and a second voice for converting the input voice into a feature vector using the analysis window length set by the analysis window length setting means. Analytical means and A standard pattern storage unit that stores a model for each recognition target vocabulary; and a likelihood between a series of feature vectors obtained by the second speech analysis unit and a standard pattern stored in the standard pattern storage unit. A speech recognition apparatus comprising: a recognition unit; and a recognition result determination unit that determines a recognition result from a likelihood of each standard pattern output from the recognition unit. 4. Speech input means for inputting speech, speech analysis means for converting the input speech into a feature vector with a predetermined analysis window length, and at least one or more past time feature vectors. A buffer for storing; a feature vector correlation calculating means for calculating at least one of a correlation value and a similarity between a feature vector output from the voice analyzing means and a past feature vector stored in the buffer; If the value or the similarity is less than or equal to a threshold value by comparing with a predetermined threshold value, an enhancement coefficient calculating unit that calculates an enhancement coefficient for enhancing the feature vector, and the feature vector is enhanced by the enhancement coefficient. Voice emphasis means for replacing with a feature vector; standard pattern storage means for storing a model for each vocabulary to be recognized; Recognition means for calculating the likelihood between the time series of feature vectors to be obtained and the standard pattern stored in the standard pattern storage means, and the recognition result is determined from the likelihood of each standard pattern output from the recognition means. A speech recognition device comprising: a recognition result determination unit. 5. Speech input means for inputting speech, at least two pitch extraction means for converting the input speech into a pitch frequency (fundamental frequency) of speech with a different analysis window length at each time, An inter-pitch correlation calculating means for calculating at least one of a correlation value and a similarity between pitch frequencies output from the pitch extracting means; and an analysis window used for the pitch extracting means in accordance with the correlation value or the similarity. If the correlation value or similarity is large from among the lengths, a large analysis window length is set as the change in the input voice is small, and if the correlation value or similarity is small, a small analysis window length is set as the change in the input voice is large. Analysis window length setting means for performing analysis, speech analysis means for converting the input speech into a feature vector using the analysis window length set by the analysis window length setting means, and a model for each vocabulary to be recognized. A standard pattern storage unit for storing, a recognition unit for calculating a likelihood between the sequence of feature vectors obtained by the voice analysis unit and a standard pattern stored in the standard pattern storage unit, and an output of the recognition unit. A speech recognition apparatus comprising: a recognition result determination unit that determines a recognition result from likelihood with a certain standard pattern. 6. A voice input unit for inputting voice, a pitch extracting unit for converting the input voice into a voice pitch frequency (basic frequency) with a predetermined analysis window length, and a pitch extracting unit. A buffer for storing at least one or more past pitch frequencies at least in the past, and at least one of a correlation value and a similarity between the pitch frequency output from the pitch extracting means and the past pitch frequencies stored in the buffer. Between the pitch value and the correlation value or similarity and a predetermined first threshold value. If the correlation value is equal to or greater than the first threshold value, an analysis longer than a predetermined analysis window length is performed. The window length is compared with a predetermined second threshold smaller than the first threshold, and when the window length is equal to or smaller than the second threshold, an analysis window length shorter than the predetermined window length is set. Setting means; speech analysis means for converting the input speech into a feature vector using the analysis window length set by the analysis window length setting means; standard pattern storage means for storing a model for each vocabulary to be recognized; Recognition means for calculating the likelihood between the series of feature vectors obtained by the voice analysis means and the standard pattern stored in the standard pattern storage means, and from the likelihood of each standard pattern output from the recognition means. A speech recognition device comprising: a recognition result determination unit that determines a recognition result.