JPS62253199A - Continuous voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] In a continuous speech recognition device that recognizes continuously uttered words and syllables, the distance between each subinterval pattern of an input continuous speech pattern and standard speech is determined by speech level processing, and When classifying the input continuous speech pattern in level processing, the continuous speech is recognized by segmenting it within a preset range of speech length corresponding to the number of segments and performing sentence level processing. This prevents unnatural classification, improves the recognition rate, and reduces the amount of processing.

[Industrial application field]

The present invention relates to a continuous speech recognition device that recognizes speech such as words and syllables that are continuously uttered, and in particular, to improving the recognition rate by preventing the occurrence of unnatural segmentation of input continuous speech patterns. This invention relates to a continuous speech recognition device that has been improved so as to allow continuous speech recognition.

[Conventional technology]

Since the recognition rate is good when recognizing continuously uttered word sounds, conventional continuous word speech recognition methods, especially the two-stage DP method, have been used to match input word sound patterns with pre-registered word standard patterns. (Two-1 level dy
namic programming match
ing) or similar methods are often used.

FIG. 4 is a block diagram showing the basic configuration of a conventional two-stage DP method for continuous word speech recognition.

In FIG. 4, when continuous word speech is input to the speech analysis section 210, the speech analysis section 210 creates a continuous word speech pattern by detecting parameters and intervals representing the characteristics of the continuous word speech, and then creates a continuous word speech pattern at the word level. It is input to the DP processing section 220.

The input continuous word speech pattern (represented by A) is expressed by the following equation (1,1) as time series data of feature vectors.

A=a(11,a(21,...a (7+)
-a (T) -(1) Here, a (7) is a parameter indicating the spectral characteristics of the input word speech pattern A in frame l, expressed as a vector quantity.

3 is the terminal frame of the human word speech pattern A, and also represents the word length (number of frames) of the input word speech pattern A. A frame is an audio section cut out by multiplying by a window function.

On the other hand, the word standard pattern dictionary 230 stores standard patterns for each word to be recognized. Similarly to the input word speech pattern A, the standard pattern of word n (referred to as Bn) to be recognized is also expressed as time-series data of feature vectors by the following equation (2).

B n = b(+, 1, b2(21,・”-b(
Jl-b (Jn)...(2) Here, b(J) is
Parameters indicating the spectral characteristics of the word standard pattern in frame j are expressed as vector quantities. Jn
is the final frame of the standard word pattern Bn, and also represents the word length (number of frames) of the standard pattern Bn of word n.

The word-level DP processing unit 220 cuts out all kinds of sub-interval patterns in the input continuous word speech pattern A, and, assuming that each sub-interval pattern corresponds to one word speech pattern, searches the word standard pattern dictionary 230. Perform DP matching with each word standard pattern to calculate the respective distances.

The distance between the subinterval pattern of one subinterval (/, m) and each word standard pattern is determined, and the optimal word with the minimum distance is set to a recognition rate of 1iN (Il, m), and this word N
Let the distance D (1+ m ) from (7!, m) be the distance between the partial interval pattern and the word.

For every subinterval pattern in the input continuous word speech pattern, that is, for each subinterval pattern formed by changing l and m in any combination under the condition 0≦12<m≦■, the subinterval pattern and the word and the corresponding optimal word are calculated and stored in a table (not shown).

The sentence level processing unit 240 calculates the distance of the word series corresponding to each subinterval pattern series formed by arbitrarily dividing the input continuous word pattern from the sum of the distances between each subinterval pattern and the word. The optimal word sequence corresponding to the segment using the sum as a reference is the recognized continuous word.

Now, assuming that the number of words (hereinafter sometimes referred to as the number of digits) of the input continuous word speech is known as kflil, the distance %1t of the optimal word sequence (referred to as Tk(1)) is calculated by the following recurrence. It is obtained by solving equation (3) using the DP method.

Here, in T (m) is the input continuous word speech pattern A
It is the minimum distance of the word sequence when it is assumed that there are X words in the partial section of frame m from the start of the frame m.

By calculating this recurrence formula (3) using the DP method for m from 1 to ■ (the number of frames of input continuous word speech pattern A) and for X from 1 to k, Tk (I) can be obtained. Desired.

If the number of words (number of digits) of the input continuous word speech is unknown, in Tk (1) found from recurrence formula (3),
Let the number of digits ko obtained from the following equation (4) using the number of digits k as a variable be the number of digits of the input continuous word.

ko= arg min Tk (1) −・−
(41) Here, arg min () is an operator that has the function of selecting k that satisfies the minimization condition in ().

The distance of the optimal word sequence with this number of digits ko is Tko
It is represented by (I).

The minimum distance 11iTk (T
) (or the optimal word sequence corresponding to Tko(1) is output as a recognized word.

By the way, when solving the above-mentioned recurrence formula (3), the search range of the interval l that determines the word boundary is performed within a certain range, as shown in the following formula (5).

If x=l: j! =O When X>1: (5) m-1mB× ≦β≦m-/min Here, it min is the minimum number of frames allowed for one 'R8% in the input continuous word speech pattern. ,R
ma is the maximum number of frames allowed for one word in the input continuous word speech pattern. This βmin and ρm
Ax can be given in advance, and the DP of the first stage
In some cases, it is automatically determined as a result of bus limitations in processing, but in any case, when the -transformation is applied, it takes a fixed value regardless of the number of digits in the word.

Therefore, the above recurrence formula (3) can be expressed in more detail as follows (
3) It is expressed by the equation.

Figure 5 shows the number of frames 1 to the input continuous word speech pattern.
=180. This figure shows the text HELD processing method when nmax=200, 48m1n=50.

1max-200, j! Since min = 50, the number of digits X of the input continuous word speech pattern A is ■.

2.3 are possible.

Figure 5 (al is the same figure when the number of digits x = l (bl is the same figure when the number of digits x = 2) (C1 is the sentence level DP processing when the number of digits x = 3) be.

Figure 5 (at at, x=l, so m-1(=180
), n=o, and the recurrence formula (3) or (3)' [
T+ (m) −′D (0, m)) m=1 applies.

Therefore, a search range of l does not exist.

Figure 5 (In bl, x=2, so m=1 (=180
), maximum search range of 1st digit l N+maX=m-Rmi
n is 130 (=180-50). Also, the minimum search range/+min of the first digit l is equal to Rmin, so 5
It is 0.

In Figure 5 tel, x=3, so m=T(-180
), maximum search range of second digit β A2max=ml m
in is 130 (=180-50), and the first digit β
Maximum search range j! + may =l12max 1
The max is 80 (130-50). Also, the minimum search range of the first digit l is 7! +min is the value of l min 50, and the minimum search range of the second digit is 12mfn = 121m
1n+l1m1n is 100.

[Problem that the invention seeks to solve]

In the sentence level processing of the conventional input continuous word speech recognition method, when the distance of the optimal word sequence is calculated by the recurrence formula (3)', the search range of l that determines the word boundary is fixed l min and 7 ! Using ma, solve the recurrence formula (3)′ at once, which is related to the number of digits of the word (m-1maχ≦l≦m-ffm1n, and input X that gives the minimum value of Tx N). Words of continuous words It was expressed as a number (number of digits).

For this reason, as shown in (bl, (C1) in FIG.
When divided by , it means that the word in the first digit is pronounced extremely long, and the word in the second digit is pronounced extremely short. Also, when the words are divided by /++nin, the first digit means that the word is pronounced extremely short, and the second digit means that the word is pronounced extremely long. However, when humans actually occur naturally, this usually does not occur.

In the conventional method, since such an unnatural classification method, that is, an unnatural P path is allowed, there is a problem that the recognition rate decreases and the amount of calculation increases accordingly.

SUMMARY OF THE INVENTION An object of the present invention is to provide an input continuous speech recognition device that prevents such unnatural classification from occurring, improves the recognition rate, and reduces the amount of calculation.

Furthermore, words are made up of syllables.
A syllable is usually made up of one vowel and one consonant. There are approximately 100 types of syllables in Japanese.

In the present invention, continuous speech includes not only continuous word speech but also continuous monosyllabic speech.

[Means for solving problems]

When conveying information or intentions by uttering multiple sounds such as word sounds or syllable sounds in succession, it is important to avoid uttering only one sound extremely slowly or, conversely, uttering only one sound extremely short. , which normally cannot occur.

The present invention focuses on this fact, and in sentence-level processing performed during continuous speech recognition, the search range for word boundaries is changed in accordance with the number of digits of the word. This method is designed to prevent the natural occurrence of P passes, thereby increasing the recognition rate and reducing the amount of processing.

Hereinafter, the means taken by the present invention to solve the above-mentioned problems in the conventional continuous word speech recognition system will be explained with reference to FIG.

FIG. 1 is a block diagram showing the basic configuration of the present invention.

In FIG. 1, reference numeral 110 denotes an audio level processing means, which cuts out a plurality of sub-interval patterns from an input continuous audio pattern created from input continuous audio, and divides each sub-interval pattern into each pre-registered standard audio pattern. The distance between the subinterval pattern and the standard voice is determined from the minimum value of these distances.

Reference numeral 120 denotes a voice length range setting means which, when dividing the input continuous voice pattern into a plurality of partial section patterns, sets the voice length range of each partial section pattern in accordance with the number of sections.

Reference numeral 130 denotes a sentence level processing means which, when dividing the input continuous speech pattern into a plurality of sub-segment patterns, divides the input continuous speech pattern into a plurality of sub-segment patterns within the speech length range set by the speech length range setting means 120 corresponding to the number of segments; The distance of the speech sequence corresponding to each subinterval pattern series formed by this division is calculated from the sum of the distances between each subinterval pattern and the standard speech, and the optimal speech sequence corresponding to the division that minimizes the sum is recognized. Continuous audio.

[For production]

When a human continuous voice pattern created from a human continuous voice is input, the voice level processing means 110 cuts out a plurality of partial interval patterns from the input continuous voice pattern, and for each partial interval pattern, each of the pre-registered The distance from the standard speech pattern is determined, and the distance between the partial section pattern and the standard speech is calculated from the minimum value of the distance.

As a result, assuming that each extracted subinterval pattern corresponds to one standard speech pattern,
The distance between the subinterval pattern and the standard CS voice is determined.

The sentence level processing means 130 divides the input continuous speech pattern and creates a series of partial interval patterns. At that time, the audio data is classified within the audio length range set by audio length range setting means 120 corresponding to the number of classifications.

The distance of the voice sequence corresponding to each subinterval pattern series formed by this division is calculated from the sum of the distances between each subinterval pattern forming the subinterval pattern series and the standard voice, and the division minimizes the sum. The speech sequence corresponding to the subinterval pattern series formed by , that is, the optimal speech sequence is defined as the recognized continuous speech.

Note that, as already mentioned, continuous speech includes not only continuous word speech but also continuous syllable speech.

As described above, when classifying an input continuous speech pattern, by limiting the classification range according to the number of classifications, unnatural classification is prevented from occurring, the recognition rate is improved, and the calculation The amount of G can be reduced.

〔Example〕

An embodiment of the present invention will be described with reference to FIGS. 2 and 3.

FIG. 2 is a block diagram showing the configuration of an embodiment of the present invention, and FIG. 3 is an explanatory diagram of the sentence level DP processing method of the embodiment.

(A) Structure of the Embodiment In FIG. 2, the voice level processing means 110, the voice length range setting means 120, and the sentence level processing means 130 are as described in FIG.

Reference numeral 140 denotes a microbon, into which continuous voice uttered by a speaker (not shown) or voice for registration is input.

Reference numeral 150 denotes a parameter extraction unit that extracts parameters representing the characteristics of the continuous voice input from the microphone 140 or the voice for registration.

Reference numeral 160 denotes a section detecting section that detects sections based on the parameters extracted by the parameter extracting section 150 and creates an input continuous speech pattern or a standard speech pattern for registration.

170 is a switching circuit that performs switching according to the input continuous voice pattern and the standard voice pattern for registration.

In the speech level processing means 110, reference numeral 111 is a standard speech pattern dictionary in which standard patterns of speech to be recognized are registered in units of words or syllables.

Reference numeral 112 denotes a voice level DP calculation unit that cuts out a plurality of partial interval patterns from the input continuous voice pattern inputted from the interval detection unit 160, and extracts a plurality of partial interval patterns from the input continuous voice pattern inputted from the interval detection unit 160, and extracts each standard voice registered in the standard voice pattern dictionary 111 for each partial interval pattern. The distance to the pattern is determined, and the distance between the subsection pattern and the standard voice is calculated from the minimum value of the distance.

In the sentence level processing means 130, 1311 to 131p
are recurrence formula calculation units, and 132 is a continuous speech determination unit.

The recurrence formula calculation unit 1311 calculates the input continuous speech back
(Divide by flit to create a one-digit partial interval pattern series. At this time, the voice range setting means 1 corresponds to the number of digits i.
20 within the range of audio length set. The distance of the i-digit voice sequence corresponding to the i-digit partial pattern sequence formed by this division is calculated from the sum of the distances between each partial interval pattern forming the 1-digit partial pattern sequence and the standard voice, and the sum is minimized. The i-digit voice sequence corresponding to the i-digit subinterval pattern sequence formed by the segmentation is set as a 1-digit voice candidate sequence, and the minimum value of the sum is output as the distance.

The continuous speech determination unit 132 performs each recurrence formula calculation 1311 to 13
The voice candidate sequence having the smallest distance from the 1-digit voice candidate series to the p-digit voice candidate series inputted from 1p, that is, the optimal voice sequence, is determined to be the recognized continuous voice.

The audio length range setting means 120 includes each recurrence formula calculation unit 1311
~131p, the voice length range is set and input in accordance with the number of digits, that is, the number of sections of the partial section pattern.

(B) Operation of the Embodiment The operation of the embodiment will be explained by dividing it into each operation performed during continuous speech recognition.

(B-1) Registration Operation Before recognition processing is performed on continuous speech uttered by a speaker, standard patterns for each speech are registered in the standard speech pattern dictionary 111 on a word-by-word or syllable-by-syllable basis.

When registering a standard voice pattern in the standard voice pattern dictionary 111, the switching circuit 170 is connected to the standard voice pattern dictionary 111 side, and the voice uttered from the microphone 140 in units of words or syllables is input to the parameter extraction unit 150. .

The parameter extraction unit 150 extracts parameters representing the characteristics of each input voice, and the interval detection unit 150 performs interval detection to create a standard pattern (standard voice pattern) for each voice, and the standard voice pattern dictionary 111 Register. The configurations and operations of the parameter extracting section 150 and the section detecting section 160 are all well known, so a detailed explanation thereof will be omitted.

(B-2) Input continuous speech pattern creation operation When recognizing input continuous speech, the switching circuit 170 is connected to the speech level DP calculating section 112 side.

When continuous sound is input from the microphone 140, the parameter extraction unit 150 and section detection unit 160 create an input continuous sound pattern and send it to the sound level DP calculation unit 112 in the same way as in the case of registering the standard sound pattern described above. input. Furthermore, the section detection unit 160 inputs the detected frame number of the input continuous audio pattern to the audio range setting unit 120.

The created input continuous speech pattern (represented by S) is expressed by time-series data of feature vectors, similarly to the above-mentioned human-powered continuous word speech pattern A, and is expressed by the following equation (6).

5=s(11,5f21.-"s+m)-...s (
U) - (6) Here, s (ml is a parameter indicating the spectral characteristics of the input speech pattern S in frame m, expressed as a Pertl quantity. U is the terminal frame of the input speech pattern S, It also represents the audio length (number of frames) of pattern S.

Similarly, the standard voice pattern Rn of voice n is also expressed as time series data of feature vectors by the following equation (7).

Rn=r(11, r(21,...r(J)・-
・r (Vn) - (71 Here, rol is a parameter indicating the spectral characteristics of the standard voice pattern Rn in frame j, expressed as a vector quantity.Vn is the terminal frame of the standard voice pattern Rn, and the voice n It also represents the audio length (number of frames) of the standard audio pattern Rn.

(B-3) Setting the audio length range When the audio length range setting means 120 receives the number of frames U of the input audio pattern S from the section detection unit 160, the audio length range setting means 120 determines the number of frames allowed for one audio in the input continuous audio pattern S. Minimum number of frames is 7! The number of digits assumed for the input audio pattern S is determined based on min and the maximum number of frames (1 max). The maximum assumed number of digits is determined from the nearest integer value that is less than U/ffm1n, and the minimum assumed number of digits is determined from the nearest integer value that is greater than U//may. Determined from numerical values. These e min and llmaX can be given in advance, or may be automatically determined as a result of DP path restriction in audio level DP processing.

Next, corresponding to the assumed number of digits, a range of voice length in the partial interval pattern corresponding to each digit is set and inputted to the recurrence formula calculation unit corresponding to the assumed number of digits.

The audio length range is set, for example, by the following equation (8) or (9).

f min (k)=-X (1cr)...(8) f max (kl--X (1+ α)f min
(kl =--xα...(9) (J f max (kl = -x tx-') where k: the assumed number of digits, from the minimum value k min to the maximum value k
Takes values up to m8K.

fma: Upper limit value of audio length range fmir+: Lower limit value of audio length range U: Number of frames of input continuous audio pattern α: Constant, usually takes a value in the range 1>α>0.

Generally, as α approaches 0, the amount of calculation decreases, but the recognition rate decreases. However, if the length of each voice is uniform, α
Even if , the recognition rate decreases little. Therefore, it is determined experimentally by taking into account the recognition rate, the amount of calculation, and the category of continuous input speech.

If the number of digits of the input continuous speech pattern is known, the process of calculating the assumed number of digits described above is not necessary.

Furthermore, f min and f ma obtained in this way
The usage of x will be explained in the recurrence formula 00 below (
B-4) Audio level DP processing The audio level DP calculation unit 112 cuts out a plurality of partial interval patterns from the input continuous audio pattern S input from the interval detection unit 160, and each partial interval pattern is
Each standard voice pattern registered in the standard voice pattern dictionary 111 and D
Perform P matching and calculate respective distances.

Similar to the conventional method described above, one partial interval (7!
, m) and each standard speech pattern, and the optimal speech with the minimum distance is defined as the recognized speech N5(A, m), and the distance from this speech N5(n, m) is 1.
iIt! Let Ds(7!, m) be the distance between the subsection pattern and the voice.

Any subinterval pattern in the input continuous speech pattern S, that is, under the condition O≦l<m≦U, l and m
For each subsection pattern formed by changing all combinations of , the distance between the subsection pattern and the voice and the corresponding optimal voice are determined and stored in a table (not shown).

(B-5) Calculation of recurrence formula in sentence level DP processing The recurrence formula calculation unit 1311 divides the input continuous speech pattern S into one piece to create a one-digit subinterval pattern series, and
An i-digit voice candidate sequence corresponding to the digit subsection pattern sequence is determined.

Now, assuming that the distance of the i-digit voice candidate sequence is Tj (U), this Ti (U) can be found by solving the following recurrence formula 00) using the DP method.

This recurrence formula 00) is the recurrence formula (00) in the conventional method mentioned above.
3) It corresponds to '. That is, T i x f
ml is the minimum distance of a voice sequence when the input continuous voice pattern S is divided into one subsection pattern, assuming that there are X voices in the subsection of frame m from the starting end of the pattern.

By calculating this recurrence formula (10) using the DP method for m from 1 to U (the number of input continuous speech pattern S frames) and for X from 1 to i (number of digits), Ti (U ) is required.

Each recurrence formula calculation unit 1311 to 131p solves the recurrence formula OI according to the number of digits, calculates the distance of the voice candidate series for each digit, and sends the voice candidate series together with the continuous voice determination unit 132. input. Note that the number of digits i of the recurrence formula 00) changes depending on the number of digits of each recurrence formula calculation unit.

Figure 3 shows the number of frames U = 1 of the input continuous audio pattern S.
80.7! may = 200.6m1n = 50, α-
This figure shows an example of the sentence level DP processing method in the case of 0, 2. In the figure (a), when the number of digits is 1, fbl in the same figure shows the case where the number of digits is 2, and in the figure (c+ is the number of digits). This is an example when is 3.

Since /may=200 and #m1n=50, the number of digits of the input continuous speech pattern S is assumed to be 1, 2.3, or 3.

In addition, f min and f max are determined by formula (8), and f min and f max in the case of ■ digit, 2 digit, and 3 digit are f m1nt , f max + i f
m1n2. f maX2 and f m1na ,
When f max3 is assumed, it is calculated as follows from equation (8).

Q f maxl = 216 f max2
=108 f max3 =72 f min
+ = 144 f m1n2 = 72
f m1n3 = 48 In Figure 8 fa), the number of digits is 1, so m = U (-1, 80), # = O, and CTC (ml = D (0, m ))
m=U applies.

Therefore, there is no search range for β.

In Figure 3(b), the number of digits is 2, so m=U (=1
80), maximum search range of 1st digit l 11maX-m −
f m1n2 is 108 (=180-72).

Also, the minimum search range Lmin of the first digit β is Lmin=m−f111
aX2 is 72 (=180-108).

In Figure 3(C), the number of digits is 3, so m=U (=1
80), maximum search range of second digit β 12rnax=m-
f m1n3 is 132 (-18(14B), 2
Minimum search range of digit p 12m1n = m - f m
ax3 is 108 (180-72). Also, the maximum search range/ImaX of the first digit l is 72 since it is the range f tnax3 from the starting end of the input continuous speech pattern, and the minimum search range A+min of the first digit l is from the beginning end 0. Since it is in the range of f m1n3, it is 48.

Comparing Figure 3 obtained in the above manner with the conventional Figure 5 under the same conditions, when the number of digits is 2, the size of the search range of l in the method of the present invention is 36 (= 108-7
2), whereas in the conventional method it is 80 (130-5
0). In addition, in the case of 3 digits, the method of the present invention is 24
(=132-108=72-48), whereas
In the conventional method, it is 30 (-130-100=80-50)
It is.

As a result, in the present invention, the unnatural classification method seen in the conventional method, that is, the unnatural P path does not occur, and the recognition rate can be improved and the amount of calculation can be reduced.

(B-6) Recognition of Continuous Speech The continuous speech determination unit 132 performs each recurrence formula calculation unit 1311 to 1.
From the voice candidate sequence of each digit input from 31p and their distances 11tT+ (U) to 'rp (U), the voice candidate sequence with the smallest distance, that is, the optimal voice sequence is detected and used as a recognized continuous voice.

The process of this continuous speech determination unit 132 is similar to the above-mentioned process of the conventional method (
The processing shown in the following equation (11) corresponding to equation 4) is performed.

The number of digits of the optimal speech sequence = arg min T i
(U) - (11) Recurrence formula calculation unit 1311 to 131p
By parallel calculation of recurrence formula 00), (11
) can be solved in one go, so the optimal speech sequence can be detected quickly, that is, continuous words can be recognized quickly.

Hereinafter, one embodiment of the present invention has been described, but each structure of the present invention is not limited to each structure of this embodiment. For example, the recurrence formula calculation unit may be configured with one unit, and the recurrence formula 00) may be solved for each digit in a serial manner.

Furthermore, as already mentioned, continuous speech includes not only continuous word speech but also continuous syllable speech.

According to the present invention, it is possible to improve the recognition rate of continuous input speech, and the effect is particularly great when the registered standard speech pattern has a small speech length range. For example, when recognizing continuous digits using a decimal number as a registered standard speech pattern, or when recognizing continuous syllabic speech using a monosyllabic speech as a registered standard speech pattern, the range of speech length of the registered standard speech pattern is small, and the time Since the length is relatively uniform, the effect of improving the recognition rate is large.

〔Effect of the invention〕

As explained above, according to the present invention, the following effects can be obtained.

(a) Unnatural segmentation of input continuous speech patterns is prevented from occurring, and the recognition rate of continuous speech can be improved. This is particularly effective when the range of voice lengths is small and the time lengths are uniform.

(b) When dividing an input continuous speech pattern into sub-segment patterns in sentence-level processing, the range of audio length of the sub-segment pattern is set corresponding to the number of segments, which reduces the amount of processing. I can do it.

, 31

[Brief explanation of drawings]

Fig. 1...Explanatory diagram of the basic configuration of the present invention, Fig. 2...
FIG. 3 is an explanatory diagram of the configuration of an embodiment of the present invention. FIG. 4 is an explanatory diagram of the sentence level DP processing method of the same embodiment. FIG. Figure: An explanatory diagram of a sentence level DP processing method in a conventional continuous word speech recognition method. 1 and 2, 110... Voice level processing means, 120... Voice length range setting means, 130... Sentence level processing means, I40
... Microphone, 150 ... Parameter extraction section, 1
60... Section detection unit, 170... Switching circuit.

Claims (3)

[Claims] (1) In a continuous speech recognition device that recognizes speech such as continuously uttered words and syllables, (a) A plurality of partial interval patterns are cut out from an input continuous speech pattern created from input continuous speech, and each partial interval is (b) audio level processing means (110) that calculates the distance from each pre-registered standard audio pattern for each pattern, and calculates the distance between the partial interval pattern and the standard audio from the minimum value of those distances; When an input continuous speech pattern is divided into a plurality of partial section patterns, a voice length range setting means (1
20) and (c) When dividing the input continuous speech pattern into a plurality of sub-segment patterns, the segmentation is performed within the range of the audio length set by the audio length range setting means 120 corresponding to the number of segments, and The distance of the speech sequence corresponding to each subinterval pattern series formed by the segmentation is calculated from the sum of the distances between each subinterval pattern and the standard speech, and the optimal speech sequence corresponding to the segment that minimizes the sum is recognized continuously. A continuous speech recognition device comprising: sentence level processing means (130) for converting speech into speech. (2) The continuous speech recognition device according to claim 1, wherein the input continuous speech pattern is an input continuous word speech pattern created from input continuous word speech. (3) The continuous speech recognition device according to claim 1, wherein the input continuous speech pattern is an input continuous syllable speech pattern created from input continuous syllable speech.