JP3083855B2 - Voice recognition method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method and apparatus.

[0002]

2. Description of the Related Art As a conventional example of a speech recognition method and apparatus, there is a speech recognition system disclosed by the present applicant in Japanese Patent Application Laid-Open No. Sho 64-23299. This is because, in advance, standard patterns corresponding to partial patterns of various audio patterns, which are time series of feature vectors obtained by converting audio signals, are set in the pattern storage means with durations and the like, and the input audio signals are input to a microphone or the like. The duration of the partial pattern is detected by converting the pattern into a voice pattern by the voice conversion means, and the distance between patterns corresponding to the similarity between the partial pattern and the standard pattern according to the duration of the partial pattern and the duration of the standard pattern. Is calculated, and the phoneme of the voice signal is recognized based on the calculation result.

That is, in the above-described voice recognition system, when the duration of a partial pattern of a voice pattern corresponding to a predetermined state of the standard pattern is X, the duration of the standard pattern is Y, and a predetermined constant is w, The distance D between patterns is
D = w (Lx−Lj) ² is calculated.

[0004]

In the voice recognition system disclosed in the above publication, a voice is recognized based on the duration of a standard pattern preset in the device and the duration of a partial pattern of the voice pattern input to the device. It recognizes the phoneme of the signal.

However, in the above-described speech recognition system, the same arithmetic processing is always performed irrespective of the extension of the duration of the partial pattern of the speech pattern, so that erroneous recognition tends to occur. In other words, vowels at the end can be extremely long, but rarely extremely short,
In the above-described voice recognition method, since the duration of the standard pattern is set to 50 (ms) or the like in advance, if the duration of the voice pattern changes to 10 (ms) or 100 (ms), a good calculation result is obtained. I can't get it. That is, in the above-described speech recognition method, as illustrated in FIG. 8, even if the duration of the speech pattern is shorter than the distribution range, a Since the same arithmetic processing is performed even for a long b in the distribution range, erroneous recognition is likely to occur.

[0006]

According to the first aspect of the present invention,
The voice conversion means converts the input voice signal into a voice pattern which is a time series of feature vectors, and a standard pattern corresponding to a partial pattern of various voice patterns is determined in advance by at least a duration and a plurality of weighting factors for the duration. Set in the pattern storage means, the time detection means detects the duration of the partial pattern of the voice pattern, and the time comparison means compares the duration of the detected partial pattern with the duration of the standard pattern. The coefficient selecting means selects a predetermined one from the weight coefficients of the standard pattern according to the sign of the standard pattern. The selected weight coefficient is referred to as the partial pattern and the standard pattern.
The similarity calculating means calculates the inter-pattern distance corresponding to the similarity between the partial pattern obtained by multiplying the square of the difference from the standard pattern and the standard pattern, and recognizes the phoneme of the voice signal based on the calculation result. Was made to recognize.

According to a second aspect of the present invention, an input voice signal is provided.
The sound conversion means is a sound pattern that is a time series of feature vectors.
Converts to turns and turns them into partial patterns of various voice patterns
The corresponding standard pattern at least for the duration and said continuation
Multiple weighting factors with respect to time are used to
Set in the turn storage means, partial pattern of voice pattern
Time detecting means detects the duration of the
The duration of the minute pattern and the duration of the standard pattern
Comparison means, and the standard pattern is
Coefficient selecting means selects a predetermined one from the weight coefficients of the
The partial pattern and the standard pattern are
Calculate similarity distance between patterns corresponding to similarity with turn
Is calculated according to the type of phoneme, and based on the calculation result.
The recognition means recognizes the phoneme of the voice signal based on the
Further, the calculated inter-pattern distance is compared with a preset threshold value by the distance comparing means, and the coefficient updating means updates the weight coefficient of the standard pattern according to the comparison result.

According to a third aspect of the present invention, the phoneme detecting means detects a phoneme whose inter-pattern distance is smaller than the inter-pattern distance calculated by the similarity calculating means according to the type of the phoneme of the partial pattern of the voice pattern. Then, the pattern updating means updates at least the weighting factor of the detected standard pattern of phonemes.

According to a fourth aspect of the present invention, there is provided a voice converting means for converting an input voice signal into a voice pattern which is a time series of a feature vector, and at least a standard pattern corresponding to a partial pattern of various voice patterns is continued. Pattern storage means set with a plurality of weighting factors for time and duration, and time detection means for detecting the duration of a partial pattern of the voice pattern converted by the voice conversion means, and the time detection means detects Time comparison means for comparing the duration of the partial pattern with the duration of the standard pattern stored in the pattern storage means is provided, and a predetermined one is selected from the weight coefficients of the standard pattern according to the sign of the comparison result of the time comparison means. coefficients provided selecting means, the partial pattern and the standard pattern of weighting coefficients this coefficient selecting means has selected to
Means for calculating the distance between patterns corresponding to the similarity between the partial pattern obtained by multiplying the square of the difference from the standard pattern and the standard pattern, and based on the calculation result of the similarity calculating means, Recognition means for recognizing the phoneme of the signal is provided.

[0010] According to a fifth aspect of the present invention, an input voice signal is provided.
Signal into a speech pattern that is a time series of feature vectors
Audio conversion means to provide partial patterns of various audio patterns.
The standard pattern corresponding to the
Set for each phoneme with multiple weighting factors for duration
Is provided, and the voice conversion means converts the
When detecting the duration of a sub-pattern of a sound pattern
Interval detecting means, and the partial pattern detected by the time detecting means.
Turn duration and standard stored in pattern storage
Providing a time comparison means for comparing with the duration of the pattern,
According to the comparison result of this time comparison means, the weight of the standard pattern is
A coefficient selection means for selecting a predetermined one from the
The partial pattern is selected according to the weight coefficient selected by the coefficient selection means.
Pattern distance equivalent to the similarity between the pattern and the standard pattern.
A similarity calculation means for calculating the separation according to the phoneme type is provided.
Based on the result of the calculation by the similarity calculating means.
A recognition means for recognizing a phoneme of the same; a distance comparison means for comparing the calculated inter-pattern distance with a preset threshold; and a coefficient updating means for updating a weighting coefficient of the standard pattern according to the comparison result. .

According to a sixth aspect of the present invention, there is provided a phoneme detecting means for detecting a phoneme whose inter-pattern distance is smaller than the inter-pattern distance calculated by the similarity calculating means in accordance with the type of phoneme of the partial pattern of the voice pattern. And a pattern updating means for updating at least the weighting coefficient of the standard pattern of the phoneme detected by the phoneme detecting means.

[0012]

According to the first and fourth aspects of the present invention, a standard pattern corresponding to a partial pattern of various voice patterns is preset in at least a duration and a plurality of weighting factors for the duration, and the continuation of the partial pattern of the voice pattern is performed. A predetermined one is selected from the weighting factors of the standard pattern according to the result of comparison between the time and the duration of the standard pattern, and a pattern distance corresponding to the similarity between the partial pattern and the standard pattern is calculated according to the selected weighting factor. By recognizing the phoneme of the voice signal, one of a plurality of weighting factors is selectively used in the arithmetic processing according to the magnitude of the duration of the partial pattern of the voice pattern with respect to the duration of the standard pattern. A phoneme whose continuation time distribution is not uniform like a vowel can be recognized well.

According to the second and fifth aspects of the present invention, the inter-pattern distance between a standard pattern preset for each type of phoneme and a partial pattern of the voice pattern converted by the voice conversion means is calculated according to the type of phoneme. By updating the weighting factor of the standard pattern according to the comparison result between the calculated distance between the patterns and a preset threshold, the weighting factor is updated for the standard pattern having the same voice pattern and phoneme, and the distance between the patterns is updated. Is shortened.

According to the third and sixth aspects of the present invention, a phoneme whose inter-pattern distance is smaller than the inter-pattern distance calculated according to the phoneme type of the partial pattern of the voice pattern is detected, and the detected phoneme is detected. The pattern updating means updates at least the weighting factor of the standard pattern of the standard pattern, so that the distance between the patterns is shortened by updating the weighting factor for the standard pattern having the same phoneme and phoneme, and the phoneme is different. Nevertheless, for a standard pattern having a small inter-pattern distance, the inter-pattern distance is expanded by updating the weighting coefficient.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition apparatus according to the first and fourth aspects of the present invention will be described with reference to FIGS. First, as illustrated in FIG. 1, the voice recognition device 1 is a voice conversion unit that converts a voice signal input to a voice input unit 2 including a microphone or the like into a voice pattern that is a time series of feature vectors. A feature sequence conversion unit 3 is connected, and a standard pattern storage unit 4 as a pattern storage unit in which various standard patterns are set in advance is stored in a RAM (Random Access Me).
mory) etc. Then, the collating unit 5 to which the standard pattern storage unit 4 and the feature sequence converting unit 3 are connected includes a time detecting unit for detecting the duration of the partial pattern, the duration of the partial pattern and the duration of the standard pattern. , A coefficient selecting means for selecting a predetermined one from the weighting coefficients of the standard pattern, a similarity calculating means for calculating a distance between patterns, and a recognition means for recognizing a phoneme of a voice signal (all of which are described below). CPU (Central Processing Unit) formed by firmware and the like
And so on. Then, the standard pattern storage unit 4
Is set to the average vector Y
And the duration 1 and the weight coefficients w ₁ and w ₂ .

In such a configuration, in the speech recognition apparatus 1, the speech signal input from the speech input unit 2 is converted by the feature sequence conversion unit 3 into a speech pattern X which is a time series of feature vectors. Here, various methods can be considered as a means for extracting a feature vector useful for such a speech recognition apparatus 1. For example, the output of a band-pass filter group (not shown) of 15 channels is output every 10 (ms). Or FFT (Fast Fourier) with a 5.0 (ms) frame period.
Transform) Extracting a 20-dimensional vector obtained by equally dividing the spectrum on a logarithmic frequency axis and calculating an average, and a 14th-order LPC (Linear Predictive Codin)
g) It is possible to extract cepstrum every 10 (ms). Then, the feature vector x _i of the i-th frame of the speech pattern X = x ₁ obtained as described _{above, x 2 ... x i ... x} I (I is the number of frames), 15-dimensional corresponding to the number of channels, etc. Becomes

Then, the collating unit 5 compares the input voice pattern X with the N stored in the standard pattern storage unit 4.
N-th reference pattern Y _n of pieces of reference pattern Y ₁ to Y _N
The inter-pattern distance Dist (X, Y _n ) corresponding to the degree of similarity is calculated sequentially, and the category (here, word) which is the phoneme to which the standard pattern having the smallest calculation result belongs is detected as the recognition result. Become.

Each standard pattern Y _n is formed as J time series, and the j-th state includes a feature vector y _nj representing the j-th state of the standard pattern Y _n ,
The duration l _nj and two weighting factors w _nj1 and w _nj2 are registered respectively. Therefore, when collated with the voice pattern X and the standard pattern Y _n, for example, part of the speech pattern X pattern _{x n + 1, x n +} 2 ... x i is the j-th reference pattern Y _n
If it corresponds to the state, (im ≧ lnj) in this case
The distance sd (m, i, n, j) from (im <lnj) is

(Equation 1) Becomes Note that dist ₁ used in the above equation is an arithmetic expression for calculating the distance between two vectors, and dist ₂ is an arithmetic expression for calculating the distance between two scalars. Therefore,
Using the Euclidean distance as these equations,
As illustrated in the flowcharts of FIGS. 2 and 3, the distance sd (m, m) between (im ≧ lnj) and (im <lnj) in this case.
i, n, j)

(Equation 2) Becomes Therefore, the pattern distance Dist (X,
Y _n )

(Equation 3) Becomes In the above equation, 0 = S (0) ≦ S (1).
(N) = I, and DT used in the flowchart of FIG. 3 is a variable used temporarily.

That is, in the speech recognition apparatus 1, one of the two weighting coefficients is used in the arithmetic processing according to the magnitude of the duration of the partial pattern of the speech pattern with respect to the duration of the standard pattern. Furthermore, even a phoneme whose distribution of duration is not uniform can be recognized well.

Here, assuming a case where dynamic programming is used as a solution of the above-described mathematical formula, an array D (i, 1) (1 ≦ j ≦ I, 1 ≦ j ≦ J)
Procedure D (i, 1) = sd (0, i, n, 1) (1 ≦ j ≦ I)

(Equation 4) It is conceivable that the procedure Dist (X, Y _n ) = D (I, J). In this manner, a mathematical expression for calculating the inter-pattern distance Dist can be easily obtained.

FIG. 4 shows an embodiment of the invention according to claims 2 and 5.
It will be described based on. First, in this voice recognition device 6,
A standard pattern is set for each type of phoneme in a standard pattern storage unit 7 which is a pattern storage means such as a RAM, and a matching unit 8 such as a CPU stores a pattern distance between the partial pattern and the standard pattern. Distance comparing means for comparing with a preset threshold is formed by firmware or the like. The matching unit 8 is connected to a weight coefficient updating unit 9 that is a coefficient updating unit that updates the weight coefficient of the standard pattern in the standard pattern storage unit 7. The other structure is the same as that of the speech recognition device 1 described above.

In such a configuration, the speech recognition device 6 calculates the inter-pattern distance between the partial pattern of the speech pattern and the standard pattern in substantially the same manner as the speech recognition device 1 described above. That is, a partial pattern a of the number of frames x
_{i + 1} , a _{i + 2} ... a _{i + x} and the inter-pattern distance Dist between the standard pattern y with the feature vector b,

(Equation 5) Calculated as At this time, the first term on the right side of the above equation is the Euclidean distance related to the spectrum, and the second term on the right side is the duration of the partial pattern. That is, the duration D is the distance between the x frame and the duration y of the standard pattern, and is obtained by two weighting factors w ₁ and w ₂ of the standard pattern.
In the case of (x ≧ y), D = w ₁ (xy) ² , and in the case of (x <y), D = w ₂ (xy) ² .

Here, in the voice recognition device 6, for example, when the type of the phoneme of the partial pattern x of the voice pattern is known, the standard pattern y of the phoneme k is read from the standard pattern storage unit 7 by the collating unit 8. The distance D is calculated again. Next, this distance D is compared with a threshold value D _0, and D
> D ₀ , two weighting factors w ₁ and w ₂ of the standard pattern
The update constant w ₀ is subtracted from one of the predetermined constants, and if D <D ₀ , the update constant w ₀ is added to _one of the two weight coefficients w ₁ and w ₂ of the standard pattern. By doing so, the restriction on the expansion of the voice pattern is reduced for each phoneme, so that the distance between the standard pattern and the voice pattern is reduced. Therefore, in the voice recognition device 6, the above-described operation is repeated, so that the weight coefficient w of the standard pattern is sequentially updated to an appropriate value, so that the accuracy of voice recognition is improved.

FIG. 5 shows an embodiment according to the third and sixth aspects of the present invention.
7 will be described with reference to FIG. First, in the speech recognition apparatus 10, a phoneme detecting means for detecting a phoneme whose inter-pattern distance is calculated smaller than the inter-pattern distance calculated according to the phoneme type of the partial pattern is stored in a matching unit 11 including a CPU or the like by firmware or the like. The matching unit 11 includes a standard pattern updating unit 1 that is a pattern updating unit that updates a weight coefficient and a duration of the standard pattern.
2 are connected. The other structure is the same as that of the speech recognition device 6 described above.

In such a configuration, the speech recognition apparatus 10 calculates the inter-pattern distance between the partial pattern of the speech pattern and the standard pattern in the same manner as the speech recognition apparatus 6 described above. That is, for the j-th phoneme of the number J, the feature vector is b _j , the duration is y _j , and the weighting factors are w _j1 and w _j.
Standard pattern y of _j2 and partial pattern a of the number of frames x
_{i + 1} , a _{i + 2} ... the distance Dist between patterns with a _{i + x}

(Equation 6) Calculated as At this time, the first term on the right side of the above equation is the Euclidean distance related to the feature vector, and the second term on the right side is the duration of the partial pattern. That is, the duration D is the distance between the x frame and the duration y of the standard pattern, and is calculated by two weighting factors w _j1 and w _j2 of the standard pattern.
In the case of (x ≧ y _j ), D = w _j1 (x−y _j ) ² , and (x <y
_In the case of _j ), D = w _j2 (x−y _j ) ² .

Here, in the voice recognition device 10, for example, when it is known that the partial pattern of the voice pattern is the phoneme k, the presence of another phoneme m whose weighting coefficient should be updated is searched. The update condition in this case is D
_{(x, y m) <D} (x, y k) and all elements of the set of become phonemes _{m, D (x, y m} ) <D (x, y k) and Dist (m) <Dist
it is possible to like all elements of the set of phonemes m as a (k), D (x, y m) is also possible to impart a condition referred to as M pieces from smaller ones to these conditions It is. When the phonemes m ₁ , m ₂ ... _{M M} to be updated are detected in this way, the duration yk of the standard pattern Yk is updated as yk = yk + α (x−yk) as illustrated in FIG.
(α is a positive constant), two weighting factors wk of the standard pattern Yk
₁ and wk ₂ are also wk ₁ = wk ₁ -a and wk ₂ = wk ₂ -b (a,
b is a positive constant). Similarly, m = m ₁ , m ₂
... even for m _M, each weight factor wm _1, wm ₂ are each wm ₁ =
wm ₁ + c, wm ₂ = wm ₂ + d (c and d are positive constants) are updated.

By doing so, as shown in FIG. 6, the voice pattern and the phoneme k have the same standard pattern Yk.
, The duration yk and the weighting factors wk ₁ and wk ₂ are updated to shorten the inter-pattern distance D (x, yk), and the inter-pattern distance D ( x, with respect to the standard pattern Ym ym) is small, the weighting factor wm _1, between wm ₂ is updated pattern distance D (x,
ym) will be enlarged. Therefore, in the voice recognition device 10, the above-described operation is repeated, so that each standard pattern is updated to an appropriate content, so that the accuracy of voice recognition is further improved.

[0028]

According to the first and fourth aspects of the present invention, the input voice signal is converted into a voice pattern which is a time series of a feature vector by a voice conversion means, and a standard pattern corresponding to a partial pattern of various voice patterns is provided. Is set in the pattern storage means in advance with at least the duration and a plurality of weighting factors for the duration, the duration of the partial pattern of the voice pattern is detected by the time detection means, and the duration of the detected partial pattern and the standard pattern are detected. Time comparing means, and the coefficient selecting means selects a predetermined one from the weighting coefficients of the standard pattern according to the comparison result, and the similarity between the partial pattern and the standard pattern according to the selected weighting coefficient. Is calculated by the similarity calculating means, and the recognition means recognizes the phoneme of the voice signal based on the calculation result. As a result, one of a plurality of weighting factors is selectively used in the arithmetic processing according to the magnitude of the duration of the partial pattern of the voice pattern with respect to the duration of the standard pattern, so that the distribution of the duration like the vowel at the end Has an effect that even a non-uniform phoneme can be recognized well, and high-performance speech recognition can be easily realized.

According to the second and fifth aspects of the present invention, the standard pattern is preset in the pattern storage means for each phoneme, and the standard pattern stored in the pattern storage means and the part of the voice pattern converted by the voice conversion means are stored. The similarity calculation means calculates the distance between the patterns and the pattern according to the type of phoneme, and the distance comparison means compares the calculated distance between the patterns with a preset threshold value. Is updated by the coefficient updating means, the distance between the patterns is shortened by updating the weighting factor for the standard pattern having the same speech pattern and phoneme, so that the speech recognition operation is repeated every time the speech recognition operation is repeated. This has the effect that a voice recognition device with improved accuracy can be obtained.

According to the third and sixth aspects of the present invention, the phoneme for which the inter-pattern distance smaller than the inter-pattern distance calculated by the similarity calculating means in accordance with the type of the phoneme of the partial pattern of the voice pattern is calculated. And the pattern updating means updates the weighting factor and the duration of the detected standard pattern of the phoneme, so that the weighting factor can be updated for the standard pattern having the same voice pattern and phoneme. Since the inter-pattern distance is shortened and the inter-pattern distance is increased by updating the weighting factor for a standard pattern having a small inter-pattern distance despite different phonemes, the repetition of the speech recognition operation becomes extremely large every time the speech recognition operation is repeated. This has effects such as that a speech recognition device with improved speech recognition accuracy can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the invention described in claims 1 and 4;

FIG. 2 is a flowchart.

FIG. 3 is a flowchart.

FIG. 4 is a block diagram showing an embodiment of the invention described in claims 2 and 5;

FIG. 5 is a block diagram showing an embodiment of the invention according to claims 3 and 6;

FIG. 6 is a characteristic diagram.

FIG. 7 is a flowchart.

FIG. 8 is a characteristic diagram showing a conventional example.

[Explanation of symbols]

1, 6, 10 voice recognition device 2 voice conversion means 4, 7 pattern storage means 5 time detection means, time comparison means, coefficient selection means, and similarity calculation means 8 time detection means, time comparison means, coefficient selection means, and similarity Calculating means and distance comparing means 9 coefficient updating means 11 time detecting means and time comparing means and coefficient selecting means and similarity calculating means and distance comparing means and phoneme detecting means 12 coefficient updating means and pattern updating means

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/10 JICST file (JOIS)

Claims (6)

(57) [Claims] 1. An audio signal input means converts an input audio signal into an audio pattern which is a time series of a feature vector, and converts a standard pattern corresponding to a partial pattern of various audio patterns into at least a duration and a plurality of patterns corresponding to the duration. Is set in advance in the pattern storage means with the weight coefficient of the above, the duration of the partial pattern of the voice pattern is detected by the time detection means, and the duration of the detected partial pattern is compared with the duration of the standard pattern. Means for comparing, the coefficient selecting means selects a predetermined one from the weighting coefficients of the standard pattern in accordance with the positive or negative of the comparison result , and determines the selected weighting coefficient as the difference between the partial pattern and the standard pattern.
The similarity calculating means calculates the inter-pattern distance corresponding to the similarity between the partial pattern multiplied by the square and the standard pattern, and the recognition means recognizes the phoneme of the voice signal based on the calculation result. A speech recognition method characterized by doing so. 2. An audio conversion means for inputting an audio signal.
Is converted into a voice pattern that is a time series of
Standard pattern corresponding to partial pattern of voice pattern
At least the duration and multiple weights for said duration
Preset in the pattern storage means for each type of phoneme with the only coefficients
And the duration of the partial pattern of the voice pattern is
The detecting means detects the continuation of the detected partial pattern.
The time comparison means compares the time with the duration of the standard pattern.
And comparing the weight of the standard pattern according to the comparison result.
The coefficient selection means selects a predetermined one from the coefficients, and this selection
The partial pattern and the standard pattern are
Calculate similarity distance between patterns corresponding to similarity with turn
Is calculated according to the type of phoneme, and based on the calculation result.
The recognition means recognizes the phoneme of the voice signal
The calculated distance between patterns is compared with a preset threshold value by a distance comparing means, and the coefficient updating means updates the weight coefficient of the standard pattern according to the comparison result. Ruoto voice recognition method. 3. The phoneme detecting means detects a phoneme whose inter-pattern distance is calculated to be smaller than the inter-pattern distance calculated by the similarity calculating means in accordance with the type of phoneme of the partial pattern of the voice pattern. 3. The speech recognition method according to claim 2, wherein the pattern updating means updates at least a weight coefficient of the standard pattern of phonemes. 4. A voice conversion means for converting an input voice signal into a voice pattern which is a time series of feature vectors, wherein a standard pattern corresponding to a partial pattern of various voice patterns has at least a duration and a duration corresponding to the duration. Pattern storage means set with a plurality of weighting factors; time detection means for detecting the duration of the partial pattern of the voice pattern converted by the voice conversion means; and continuation of the partial pattern detected by the time detection means. A time comparing means for comparing the time with the duration of the standard pattern stored in the pattern storing means, and a coefficient for selecting a predetermined one from the weighting coefficients of the standard pattern according to the sign of the comparison result of the time comparing means Selecting means, the weighting coefficient selected by the coefficient selecting means is used as the partial pattern and
A similarity calculating means for calculating a pattern-to-pattern distance corresponding to the similarity between the partial pattern and the standard pattern, which is obtained by multiplying the square of the difference between the standard pattern and the standard pattern , based on the calculation result of the similarity calculating means And a recognition unit for recognizing a phoneme of the voice signal. 5. When an input audio signal is a feature vector
A voice conversion means for converting to a voice pattern
Standard for various patterns of audio patterns
The pattern is at least for the duration and the duration
Pattern set for each phoneme with multiple weighting factors
Storage means, and a voice pattern converted by the voice conversion means.
Detection means for detecting the duration of a partial pattern of a pattern
And a pattern of the partial pattern detected by the time detecting means.
Duration and standard pattern stored in the pattern storage means
Time comparison means to compare the duration of the
Weight of the standard pattern according to the comparison result of the inter-comparison means.
A coefficient selection means for selecting a predetermined one from coefficients is provided.
The partial parameters according to the weighting factors selected by the coefficient selecting means of
Putter equivalent to the similarity between the turn and the standard pattern
Similarity calculation means that calculates the distance between phonemes according to the type of phoneme
And based on the calculation result of the similarity calculation means,
A recognizing means for recognizing a phoneme of a voice signal; a distance comparing means for comparing a calculated inter-pattern distance with a preset threshold; a coefficient updating means for updating a weight coefficient of the standard pattern according to the comparison result the characteristics and to Ruoto voice recognition device that is provided. 6. A phoneme detecting means for detecting a phoneme whose inter-pattern distance smaller than the inter-pattern distance calculated by the similarity calculating means according to the phoneme type of the partial pattern of the voice pattern is provided. 6. A speech recognition apparatus according to claim 5, further comprising a pattern updating means for updating at least a weight coefficient of a standard pattern of phonemes detected by said means.