JPH1091186A - Voice recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a voice recognizing method which is tolerative for increasing of the number of vocabulary and noise and a recognition rate is high by accumulating distance obtained from an input vector and a partial pattern with a statistical distance scale, obtaining the accumulated distance, and making a word of the minimum accumulation distance the recognized result from the accumulated distance. SOLUTION: A voice recognizing device is constituted with an acoustic analyzing section 1, a feature parameter extracting section 2, a plural frame buffer 3, partial distance calculating section 4, a partial standard pattern storing section 5, a path discriminating section 6, a distance accumulating section 7, a discriminating section 8, a voice section detecting section 9, and a frame synchronizing signal generation section 10. And a partial distance between an input vector formed by plural frames and a partial (standard) pattern of a voice of a word is obtained with a statistical distance scale based on posterior probability, the input vector is updated shifting a frame, distance between each partial vector is accumulated, and a word in which accumulated distance is minimum is made the recognized result. Thereby, a high recognition rate can be obtained for voice recognition for an unspecified talker.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for recognizing a human voice by a machine.

[0002]

2. Description of the Related Art In recent years, recognition devices for unspecified speakers that can recognize any voice without registering the voice of the user have come into practical use. As a practical method for an unspecified speaker, two patents (JP-A-61-188599 and JP-A-62-111293) filed by the present applicant before will be described as conventional examples. Japanese Patent Application Laid-Open No. 61-188599
The second conventional example is disclosed in Japanese Patent Application Laid-Open No. 62-111293.

In the first conventional method, a voice section is determined by finding the start and end of an input voice, and the voice section is linearly expanded and contracted by a fixed time length (I frame). This is a method of recognizing words by calculating the degree by performing pattern matching using a statistical distance scale.

[0004] A word standard pattern is obtained by uttering a word to be recognized by many people to collect voice samples, and all voice samples are I-frames of a fixed time length (I = 1 in the embodiment).
6), and thereafter, a statistic (an average value vector and a covariance matrix) between speech samples is obtained for each word, and this is processed by processing. That is, the time length of all word standard patterns is constant (I frame), and one standard pattern is prepared for one word in principle.

In the first conventional example, it is necessary to detect a voice section before pattern matching, but in the second conventional example, a portion that does not require voice section detection is different. This is a method that enables a method (word spotting method) of extracting and recognizing a voice part from a signal containing noise by pattern matching. That is, in a sufficiently long input section including speech, a partial area is set in the input section, and matching with the standard pattern is performed while expanding and contracting the partial area. Then, the partial area is shifted in the input section by the unit time, and the operation of performing the matching with the standard pattern is similarly performed in the entire input section, and the word having the smallest distance in all the matching calculations is performed. The standard pattern name is used as the recognition result. In order to enable the word spotting method, a statistical distance measure based on a posterior probability is used as a distance measure for pattern matching.

[0006]

The method of the prior art is a practical method capable of miniaturization. In particular, the second conventional example is:
Since it is strong against noise, it is beginning to be used for practical use.

However, a problem of the prior art is that a sufficient word recognition rate cannot be obtained. For this reason, it can be used for applications where the number of vocabularies is small, but if the number of vocabularies is increased, the recognition rate will be reduced and it will not be practical. Therefore, there is a problem that the use of the recognition device is limited in the method of the related art.

An object of the present invention is to solve the above-mentioned conventional problems, and an object of the present invention is to provide a speech recognition method which is robust against an increase in the number of words and noise and has a high recognition rate.

[0009]

In order to solve this problem, the present invention uses voice data uttered by many people,
A step of dividing the recognition target word into partial sections sharing an adjacent frame, and previously generating a standard pattern of the recognition target word in which partial (standard) patterns expressing the partial sections are connected for all the recognition target words; Analyzing the input voice for each fixed time length (frame) to obtain a characteristic parameter, obtaining an input vector using the characteristic parameters of a plurality of frames, and calculating a partial distance between the input vector and each of the partial patterns as a posterior probability. A step of calculating a cumulative distance obtained by accumulating partial distances between the input vector generated while shifting the frame and the partial pattern, and a step of calculating the cumulative distance of all recognition target words with respect to the standard pattern. Comparing the word with the minimum cumulative distance as a recognition result.

Thus, a speech recognition method that is robust against an increase in the number of words and noise and has a high recognition rate can be obtained.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 of the present invention comprises the steps of: forming a standard pattern of a word to be recognized by concatenating partial patterns; and obtaining an input vector while shifting a frame from input speech. A step of accumulating a distance obtained by a statistical distance scale from the input vector and the partial pattern to obtain a cumulative distance; and a step of obtaining a word of a minimum cumulative distance from the cumulative distance as a recognition result. The partial distance between the input vector obtained from the input voice while shifting the input voice and the part (standard) pattern constituting the standard pattern of the word voice is calculated using a statistical distance scale, and the distance is accumulated. A recognition result is obtained, and a recognition rate is obtained for speech recognition for an unspecified speaker.

According to a second aspect of the present invention, a word to be recognized is divided into sub-intervals using voice data uttered by many people,
A step of connecting a partial (standard) pattern expressing the partial section to generate a standard pattern of the recognition target word in advance for all the recognition target words, and analyzing the input voice for each predetermined time length (frame) Calculating an input vector with the feature parameters of a plurality of frames, obtaining a partial distance between the input vector and each of the partial patterns constituting the standard pattern on a statistical distance scale, and shifting the frame. Calculating a cumulative distance obtained by accumulating partial distances between the input vector and the partial pattern generated while generating, and comparing the cumulative distances of all the words to be recognized with the standard pattern to each other to obtain a word having a minimum cumulative distance as a recognition result. Is divided into an input vector formed by a plurality of frames and a word voice into sub-intervals,
A partial distance with a partial (standard) pattern representing the partial interval is calculated using a statistical distance scale based on a posterior probability, an input vector is updated while shifting a frame, and a distance between each partial vector is accumulated. The recognition result is a word that minimizes the cumulative distance.In speech recognition for unspecified speakers, a robust and high recognition rate can be obtained for an increase in the number of vocabulary words and noise, and the processing is simple. Using a processing processor (DSP) or the like, it has the effect of realizing a small-sized recognition device capable of real-time operation.

[0013] The invention according to claim 3 is the invention according to claim 1 or 2.
In the above, the partial section of the recognition target word is divided so as to include a section overlapping each other, and it is possible to surely obtain the motion information of the boundary of the section and to generate a more detailed partial pattern.

According to a fourth aspect of the present invention, a recognition target word is divided into a plurality of partial sections using voice data uttered by a large number of persons, and a partial (standard) pattern expressing the partial section is connected. Generating a standard pattern of the recognized words to be recognized in advance for all the words to be recognized, and analyzing the input voice for each fixed time length (frame) to obtain feature parameters. Calculating a partial distance between the input vector and each of the partial patterns on a statistical distance scale based on a posteriori probability, and accumulating a partial distance between the input vector generated while shifting a frame and the partial pattern. The step of calculating the accumulated distance and the accumulated distance of all the words to be recognized with respect to the standard pattern are compared with each other, and the word of the minimum accumulated distance is recognized. And the result step, as a partial pattern composed of a plurality of frames, and when obtaining an input vector and a partial distance, by using a statistical distance scale based on the posterior probability, the difference between the input position and the partial pattern can be obtained. Regardless, there is an effect that the partial distance can be obtained.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the first embodiment, the starting point of the input voice
This is an embodiment in a case where the termination is detected in advance. In this case, pattern matching may be performed only in the voice section. The second embodiment is an example in which the start and end of the input voice are unknown. In this case, a method is used in which matching of the input signal and the standard pattern is performed while shifting the unit time by unit time over the entire area of a sufficiently wide section including the input voice, and a partial section having a minimum distance is cut out. This type of method is commonly called word spotting.

(Embodiment 1) FIG. 1 is a functional block diagram of a speech recognition apparatus according to Embodiment 1 of the present invention, which will be described.

In FIG. 1, an acoustic analysis unit 1 converts an input signal into an analog signal, fetches the signal (sampling frequency 10 kHz), and analyzes the signal every fixed time length (called a frame, 10 ms in this embodiment). In this embodiment, linear prediction analysis (LPC analysis) is used. The feature parameter extracting unit 2 extracts a feature parameter based on the analysis result. In this embodiment, the LPC
The twelve parameters of the cepstrum coefficient (C _{0 to} C ₁₀ ) and the difference power value V ₀ are used. Input parameters per frame

[0018]

[Outside 1]

In this case, the characteristic parameters are as shown in (Expression 1).

[0020]

(Equation 1)

Here, j is the input frame number, and p is the order of the cepstrum coefficient (p = 10). The frame synchronizing signal generator 10 is a part that generates a synchronizing signal every 10 ms, and its output enters all blocks. That is, the entire system operates in synchronization with the frame synchronization signal.

The voice section detection section 9 is a section for detecting the start and end of the input signal voice. A simple and general method for detecting a voice section uses the power of voice, but any method may be used. In this embodiment, the recognition starts when the beginning of the voice is detected, and j = 1.

The multi-frame buffer 3 is a part that forms an input vector used for pattern matching (partial matching) by integrating feature parameters of frames adjacent to the j-th frame. That is, the input vector corresponding to the j-th frame

[0024]

[Outside 2]

Is represented by the following equation.

[0026]

(Equation 2)

That is, the input vector is a vector obtained by integrating the characteristic parameters of j−L ₁ to j + L ₂ frames every m frames. If L ₁ = L ₂ = 3 and m = 1, the number of dimensions of the input vector is (P + 2) × (L ₁ + L ₂ +1) = 12 × 7 = 84. Although the frame interval m is constant in (Equation 2), it need not be constant. When m is variable, it corresponds to thinning out frames non-linearly.

The partial standard pattern storage section 5 stores standard patterns of words to be recognized as a combination of partial patterns. Here, the standard pattern creation method in the present embodiment will be described in some detail.

In order to make the story easier to understand, the words to be recognized are now converted to Japanese numerals "Ichi", "Ni", "San", "Yon".
There are 10 types: “go”, “Roku”, “Nana”, “bee”, “kyu”, and “zero”. The use of such an example has no effect on the generality of the description.

For example, a standard pattern of "Sun" is created in the following procedure. (1) Prepare data in which many people (supposed to be 100 people) say “sun”.

(2) The duration distribution of 100 “suns” is examined, and the average time length I ₃ of the 100 persons is determined.

(3) Search for I ₃ samples with a length of time from 100 persons. When there are a plurality of samples, an average value of the plurality of samples is calculated for each frame. The representative sample obtained in this way is shown in (Equation 3).

[0033]

(Equation 3)

Where

[0035]

[Outside 3]

Is a parameter vector per frame, and is composed of 11 LPC cepstrum coefficients and differential power, as in (Equation 1).

(4) Pattern matching is performed between each of the 100 samples and the representative sample, and the correspondence between the representative sample and each of the 100 samples (correspondence between the most similar frames). ). The distance calculation uses the Euclidean distance. The distance d _{i, i ′} between the i frame of the representative sample and the i ′ frame of a certain sample is represented by (Equation 4).

[0038]

(Equation 4)

Here, t represents a transposed matrix.
The correspondence between frames can be efficiently obtained by using a dynamic programming (DP method) technique.

(5) Each frame of the representative sample (i = 1 to
In accordance with I ₃ ), a partial vector in the form of (Equation 2) is extracted from each of the samples for 100 persons. L ₁ for simplicity
= L ₂ = 3 and m = 1.

Corresponding to the i-th frame of the representative sample,
The partial vector of the n-th sample of the 100 persons is as follows.

[0042]

(Equation 5)

Here, (i) indicates that the frame corresponds to the i-th frame of the representative vector in the n-th sample.

[0044]

[Outside 4]

Is an 84-dimensional vector in this embodiment (n = 1 to 100). (6) Average value of the above vectors for 100 people

[0046]

[Outside 5]

(In this example, k = 3; 84 dimensions) and covariance matrix

[0048]

[Outside 6]

(84 × 84 dimensions) is obtained (i = 1 to
I ₃ ). The average value and the covariance matrix will exist by the number I ₃ of the standard frame length (however, these need not necessarily be created for all frames; they may be created by thinning them out).

In the same procedure as in the above (1) to (6), an 84-dimensional vector and a covariance matrix are obtained for words other than "sun".

The moving average is calculated for all sample data for 100 names for all words.

[0052]

[Outside 7]

(84 dimensions) and moving covariance matrix

[0054]

[Outside 8]

(84 × 84 dimensions) is obtained. These are called surrounding patterns. Next, a standard pattern is created using the average value and the covariance.

A. The covariance matrix is shared by (Equation 6).

[0057]

(Equation 6)

Where K is the type of the word to be recognized (K = 10),
I _k represents a standard time length of a word k (k = 1, 2,..., K). Also,
g is the mixing ratio of the surrounding pattern, and normally g = 1.

B. Partial pattern of each word

[0060]

[Outside 9]

And

[0062]

[Outside 10]

Is created.

[0064]

(Equation 7)

[0065]

(Equation 8)

The derivation of these equations will be described later. FIG. 2 shows a conceptual diagram of the standard pattern creation method. FIG. 2A shows a power pattern of audio when the input signal is “sun”. FIG.
(B) conceptually shows a method of creating a partial pattern. In between the start and end of the speech samples, seeking frame correspondence between the representative sample, thereby dividing the speech samples to I _3. In the figure, the corresponding frame with the representative sample is indicated by (i). Then, for each of the start (i) = 1 to the end (i) = I ₃ of the voice, (i) −
L ₁ ~ (i) + with 100 persons data L ₂ of section calculates the mean value and the covariance, the partial pattern

[0067]

[Outside 11]

[0068]

[Outside 12]

Is obtained. Therefore, the standard pattern of the word k is a concatenation (collection) of Ik partial patterns including sections that overlap each other. FIG. 2 (c)
Indicates a method of creating a surrounding pattern. The surrounding pattern is as shown in the figure for all the data used to create the standard
The moving average and the moving covariance are obtained while shifting the partial section of L1 + L2 + 1 frame by one frame. The range of creating the surrounding pattern may be not only within the voice section but also the preceding and following noise sections. In a second embodiment described later, it is necessary to include a noise section in the surrounding pattern.

Next, the calculation of the partial distance will be described. The distance (partial distance) between the partial standard pattern of each word created in advance as described above and the plurality of frame buffers 3 is calculated by the partial distance calculation unit 4.

The calculation of the partial distance is performed by using a statistical distance scale between an input vector including information of a plurality of frames represented by (Equation 2) and a partial pattern of each word. Since the distance as a whole word is obtained by accumulating the distance to the partial pattern (referred to as partial distance), the partial distance can be compared with each other regardless of the input position and the difference in the partial pattern. Needs to be calculated. For this, it is necessary to use a distance measure based on the posterior probability.
An input vector of the form (Equation 2)

[0072]

[Outside 13]

(For the sake of simplicity, i and j will be described for the time being.) Posterior probability of word k for partial pattern ω _k

[0074]

[Outside 14]

Is as follows using the Bayes theorem.

[0076]

(Equation 9)

The first term on the right side treats the appearance probabilities of the respective words as the same and treats them as constants. The prior probability of the second term on the right side is based on the assumption that the parameter distribution is a normal distribution,

[0078]

(Equation 10)

Is represented by

[0080]

[Outside 15]

Is the sum of probabilities for all possible input conditions, including the word and its surrounding information. When the parameter is an LPC cepstrum coefficient or band-pass filter output, the distribution shape is considered to be close to a normal distribution. be able to.

[0082]

[Outside 16]

Assume that the mean follows a normal distribution, and

[0084]

[Outside 17]

The covariance matrix is

[0086]

[Outside 18]

When using the above, the equation (11) is obtained.

[0088]

[Equation 11]

By substituting (Equation 10) and (Equation 11) into (Equation 9), taking the logarithm, omitting the constant term, and further multiplying by -2, the following equation is obtained.

[0090]

(Equation 12)

[0091] This (Equation 12) is an equation in which the Bayes distance is posteriorly probabilized, and has a disadvantage that the discriminating ability is high but the amount of calculation is large. This equation is developed into a linear discriminant as follows. Assume that the covariance matrices, including all subpatterns and surrounding patterns for all words, are equal. Under this assumption, the covariance matrix is shared by (Equation 6), and

[0092]

[Outside 19]

[0093]

[0094]

[Outside 20]

Instead of

[0096]

[Outside 21]

By substituting the first and second terms of (Equation 12),
The terms can be expanded as follows:

[0098]

(Equation 13)

[0099]

[Equation 14]

In (Equation 13) and (Equation 14)

[0101]

(Equation 15)

[0102]

(Equation 16)

Is as follows. The third term of (Equation 12) becomes zero. Therefore, (Equation 12) becomes a simple primary discriminant as follows.

[0104]

[Equation 17]

Here, (Equation 17) is rewritten as a distance between the input j-th frame component (Equation 2) and the partial pattern of the i-th frame component of the word k.

[0106]

(Equation 18)

Here,

[0108]

[Outside 22]

Is (Equation 7).

[0110]

[Outside 23]

Is given by (Equation 8). L _k ^{i, j} is a word
This is the partial similarity between the i-th partial pattern of k and the vector near the input j-frame.

In FIG. 1, the distance accumulating section 7 is a section for accumulating the partial distances for each word in the section from i = 1 to I _k and obtaining the distance for the entire word. In that case, it is necessary to accumulate the input voice length (J frame) while expanding and contracting to the standard time length I _k of each word. This calculation can be efficiently performed using a dynamic programming method (DP method).

Now, for example, if the cumulative distance of "sun" is to be obtained, k is always 3, so k is omitted and the calculation formula will be described.

The partial distance L ^{i, j} between the input j-th frame part and the i-th partial pattern is expressed as l (i, j),
If the cumulative distance to the (i, j) frame is expressed as g (i, j),

[0115]

[Equation 19]

Is obtained. The route determination unit 6 selects the route with the smallest cumulative distance among the three routes in (Equation 19).

FIG. 3 illustrates a method for obtaining the cumulative distance by the DP method. Although a pen-shaped asymmetric path is used as shown in the figure, various other paths can be considered. In addition to the DP method, a linear expansion / contraction method may be used, or a hidden Markov model method (HMM method) may be used.

In this way, the distances are sequentially accumulated, and at the time when i = I _k and j = J, the accumulated distance G _k (I
_k , J) is determined for each word.

The determination section 8 finds the minimum value of the cumulative distance G _k (I _k , J) and obtains the recognition result by (Equation 20).

[0120]

[Outside 24]

Is output.

[0122]

(Equation 20)

(Embodiment 2) Next, FIG. 4 shows a functional block diagram of a speech recognition apparatus according to Embodiment 2 of the present invention, and will be described. In the first embodiment, the pattern matching is performed after the voice section detection. However, in the second embodiment, the voice section detection is not required. This is a method of recognizing a word by extracting a portion having a minimum distance from an input signal, and is one of the “word spotting methods”.

This method is based on the idea that "if an input signal contains a target voice, the distance (cumulative distance) to a correct standard pattern in the voice section is minimized." Therefore, a method is employed in which matching with the standard pattern is performed while shifting one frame at a time in a sufficiently long input section including a noise section before and after the input voice. In FIG. 4, blocks having the same numbers as those in FIG. 1 have the same functions. 4 differs from FIG. 1 in that it does not include the voice section detection unit 9 and that a distance comparison unit 12 and a temporary storage 11 are provided instead of the determination unit 8. Hereinafter, only portions different from the first embodiment will be described.

First, when the pattern matching starts (j
= 1) is before the beginning of the voice, and the time when the pattern matching ends (the time when j = J) is after the end of the voice. There are various methods for detecting the end of pattern matching, but in the present embodiment, the point in time at which the distance from all the standard patterns becomes sufficiently large is j = J.

The method of creating the standard pattern is exactly the same as in the first embodiment. However, it is necessary to use a sufficiently wide section before and after the voice section as a range in which the surrounding pattern is created using the voice sample. The reason is the denominator term of (Equation 9).

[0127]

[Outside 25]

Is defined as "probability density for all parameters to be subjected to pattern matching."

The greatest structural difference from the first embodiment is that the comparison of the cumulative distance for each word is performed for each frame. The distance comparison unit 12 calculates the cumulative distance G _k (I _k ,
j), the word having the smallest cumulative distance in the j-th frame

[0130]

[Outside 26]

Is obtained. Then, the minimum value at that time is also obtained at the same time. That is,

[0132]

(Equation 21)

[0133]

(Equation 22)

The temporary storage 11 stores the minimum value G _min of the cumulative distance appearing up to the j−1 frame and the standard pattern name k when the cumulative distance becomes the minimum.

With G _min

[0136]

[Outside 27]

Are compared,

[0138]

[Outside 28]

If so, the process proceeds to the next frame (j = j + 1) while keeping the temporary storage 11 as it is.

[0140]

[Outside 29]

Then,

[0142]

[Outside 30]

Then, the process proceeds to the next frame. in this way,
The temporary storage 11 always has the minimum value and the recognition result up to that frame. When the end of the pattern matching range (j = J) is reached, it is stored in the temporary storage 11.

[0144]

[Outside 31]

The result is the recognition result. Embodiment 2 is an effective method when it is difficult to detect a voice section such as utterance in noise.

In order to confirm the effect of the present invention, a total of 15
A recognition experiment was performed using 10 numeric data uttered by 0 persons. A standard pattern was created using data of 100 (50 men and women), and the remaining 50 were evaluated.
The evaluation conditions are shown in (Table 1),

[0147]

[Table 1]

The evaluation results are shown in (Table 2).

[0149]

[Table 2]

As described above, the improvement of the recognition rate in this embodiment is very remarkable.

[0151]

According to the present invention, a partial distance between an input vector formed by a plurality of frames and a partial (standard) pattern of a word voice is obtained by a statistical distance scale based on the posterior probability, and the input is performed while shifting the frames. This method relates to a method of updating vectors and accumulating the distances between each partial vector, and recognizing words that minimize the accumulated distance. Is obtained.

Furthermore, since the calculation method is simple, it can be easily realized as a small device using a signal processor (DSP).

As described above, the present invention is a practically effective method, and its effect is great.

[Brief description of the drawings]

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

FIG. 2 is a conceptual diagram illustrating a method of creating a partial pattern and a surrounding pattern in a standard pattern creation method according to the present invention

FIG. 3 is a schematic diagram showing a method of calculating a collation of a standard pattern in which an input voice and a partial pattern are connected by a dynamic programming method according to the present invention;

FIG. 4 is a functional block diagram of a voice recognition device according to a second embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Acoustic analysis part 2 Feature parameter extraction part 3 Multiple frame buffer 4 Partial distance calculation part 5 Partial standard pattern storage part 6 Path judgment part 7 Distance accumulation part 8 Judgment part 9 Voice section detection part 10 Frame synchronization signal generation part 11 Temporary storage 12 Distance comparison section

 ────────────────────────────────────────────────── ─── Continued on the front page (72) Inventor Tatsuya Kimura 3-10-1 Higashi Mita, Tama-ku, Kawasaki City, Kanagawa Prefecture Matsushita Giken Co., Ltd.

Claims (4)

[Claims] 1. A step of creating a standard pattern of a word to be recognized by concatenating partial patterns, a step of obtaining an input vector while shifting a frame from an input voice, and a statistical distance measure from the input vector and the partial pattern A speech recognition method, comprising: accumulating the distances obtained in (1) to obtain an accumulated distance; and obtaining a word having a minimum accumulated distance from the accumulated distance as a recognition result. 2. A recognition target word is divided into sub-intervals using voice data uttered by a large number of people, and all of the standard patterns of the recognition target word obtained by connecting partial (standard) patterns expressing the sub-intervals are obtained. Generating input words in advance, analyzing input speech at predetermined time lengths (frames) to obtain characteristic parameters, and obtaining input vectors using characteristic parameters of a plurality of frames; A step of obtaining a partial distance to each partial pattern on a statistical distance scale; a step of obtaining a cumulative distance obtained by accumulating partial distances between an input vector generated while shifting a frame and the partial pattern; Comparing the cumulative distances to the patterns with each other to obtain a word having the minimum cumulative distance as a recognition result. 3. The speech recognition method according to claim 1, wherein the partial patterns of the recognition target word include sections (frames) overlapping each other. 4. A recognition target word standard in which a recognition target word is divided into a plurality of partial sections using voice data uttered by a large number of persons, and a part (standard) pattern expressing the partial section is connected. Generating a pattern in advance for all recognition target words; analyzing input speech for each fixed time length (frame) to determine a feature parameter; and determining an input vector using feature parameters of a plurality of frames; Obtaining a partial distance between an input vector and each of the partial patterns on a statistical distance scale based on a posterior probability, and obtaining an accumulated distance obtained by accumulating partial distances between the input vector generated while shifting a frame and the partial pattern. And comparing the cumulative distances of all the recognition target words with respect to the standard pattern to each other, and using the word having the minimum cumulative distance as a recognition result. A voice recognition method.