JPH09305195A - Speech recognition device and speech recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition device and a speech recognition method capable of processing recognition with a high speed without losing precision. SOLUTION: In a speech section part of a digital sound data sampled in a speech input part 1, a cepstrum is calculated in a feature quantity extraction part 3 and generated, and is further normalized in a SPT(successive processing along trajectory) processing part 4'. The normalized cepstrum are linearly matched to a standard pattern in a preliminary selection part 5 by a linear matching and the order of a roughness of the standard pattern is decided, and also some of the high orders are made candidates, which are sent to a pattern recognition part 7' by DP(Dynamic programming) matching as word numbers for the preliminary selection candidates. In this pattern recognition part 7', the pattern recognition by the DP matching is executed between standard patterns of the word numbers becoming the above-mentioned candidates and the normalized cepstrum outputted from the SPT processing part 4', and the most similar pattern is outputted to a result output part 8 as an identification candidate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus and method for preliminarily narrowing down standard pattern candidates and then outputting final candidates in the apparatus and method for recognizing inputted speech.

[0002]

2. Description of the Related Art A feature amount of input voice data is calculated, a standard pattern is narrowed down by matching with a standard pattern, and then voice recognition is performed based on the similarity between the narrowed down standard pattern and the feature amount. As a conventional device for carrying out the above, for example, JP-A-63-104098 and JP-A-58-176698 are shown.
The former speech recognition apparatus disclosed in Japanese Patent Laid-Open No. 63-104098 uses a normal DP (Dynamic Programming) matching dedicated apparatus as a preliminary selecting section, and the standard pattern selected by this preliminary selecting section is similar to the input speech data. Degree
The calculation is performed by a highly accurate DP matching device that takes the phoneme duration into consideration. The latter speech recognition device disclosed in Japanese Patent Laid-Open No. 58-176698 is a general-purpose means (software) when a primary (preliminary) DP matching is performed by a dedicated computing means and the matching is not successful. The second DP matching is performed.

[0003]

However, in the above-mentioned conventional speech recognition apparatus, since DP matching is used for both preliminary and final matching, the processing speed must be sufficiently high. However, there is a problem in that the accuracy of voice recognition is not so good because normal DP matching is performed twice.

It is an object of the present invention to provide a voice recognition device and a voice recognition method which can solve the above-mentioned drawbacks and can perform recognition processing at high speed without lowering accuracy.

[0005]

A voice recognition device of the present invention comprises a feature amount calculation means for calculating a feature amount of input voice data, and the calculated feature amount and a standard pattern stored in advance. Preselection means for preliminarily selecting a standard pattern according to the result of linear matching, and voice recognition means for performing voice recognition by performing non-linear matching with the feature amount and the preliminarily selected standard pattern. , Is provided.

In the above speech recognizer, since linear matching is used to preselect a standard pattern,
It does not require complicated calculation such as DP matching. Therefore, the recognition processing can be performed at high speed with a simple configuration. Preliminarily selecting a standard pattern means, most typically, ranking standard patterns and selecting a standard pattern from a first rank to a predetermined rank. In addition, it is also possible to select all or a part of those satisfying a certain condition (having a close pattern) without ranking.

The speech recognition apparatus of the present invention further comprises a cepstrum calculating means for calculating a cepstrum for each short-time frame of the input voice data, and the intervals between the cepstrums on the trajectory of the calculated cepstrum in advance. The cepstrum normalization means for normalizing each cepstrum by linear interpolation so that the set fixed length is obtained, and the linear matching is performed by the normalized cepstrum pattern and the standard cepstrum pattern stored in advance, and according to the result. A preselection means for preliminarily selecting a standard cepstrum pattern, and a voice recognition means for performing voice recognition by performing non-linear matching with the normalized cepstrum and the preliminarily selected standard cepstrum pattern. It is characterized by

In the above speech recognition apparatus, the cepstrum normalizing means normalizes the intervals between the cepstrums so as to have a preset fixed length. According to this normalizing means, the intervals between the cepstrums are not equal on the time axis, and the intervals are small in places where consonants and the like change drastically.
In the place where the vowel changes slowly, the intervals become coarse and the number of cepstrums, which are feature vectors, does not change greatly with the length of word occurrence. For this reason, the preliminary selection means for preliminary selection of the standard cepstrum pattern does not need to use DP matching in which patterns having different numbers of vectors are compared with each other, and linear matching can be easily performed, and the processing speed can be increased. .

Further, the speech recognition apparatus of the present invention has a cepstrum calculating means for calculating a cepstrum for each short-time frame of input speech data, and an interval L of each cepstrum on the calculated trajectory of each cepstrum. In order to have preset fixed lengths L1 and L2 (L1> L2),
Cepstrum normalization means for normalizing each cepstrum into two types by linear interpolation, linear matching or non-linear matching is performed by a cepstrum pattern normalized at the fixed length L1 interval and a standard cepstrum pattern stored in advance, Non-linear matching is performed by preliminary selecting means for preliminarily selecting a standard cepstrum pattern according to the result, non-linear matching by the cepstrum normalized at the fixed length L2 interval, and the preliminarily selected standard cepstrum pattern. And a voice recognition means for performing recognition.

In the above speech recognition apparatus, in the preliminary selection of the standard cepstrum pattern, the fixed length L1 for normalization is lengthened to perform rough matching. The matching in this case may be linear matching or non-linear matching such as DP matching. When continuous DP matching, which is a kind of DP matching, is used, so-called word spotting becomes possible by extracting and recognizing words from a sentence. As described above, when the preliminary standard cepstrum pattern is selected, the interval between the cepstrum on the cepstrum locus is set to a preset fixed length, and the fixed length L1 is increased so that the normalized cepstrum is roughly generated. With this, the amount of calculation required for matching at the time of preliminary selection is reduced, and the recognition processing can be speeded up.

Further, the voice recognition apparatus of the present invention uses a fuzzy operation in which the number of standard cepstrum patterns preliminarily selected by the preselection means is the background noise information at the time of voice input as the input information of the antecedent part. A fuzzy arithmetic means for deciding or a fuzzy arithmetic means for deciding by a fuzzy arithmetic operation in which numerical information in which a part of each word of the standard cepstrum pattern is a part of another word is used as input information of the antecedent part. .

In the above speech recognition apparatus, the number of standard cepstrum patterns to be preliminarily selected is determined by the fuzzy calculation, so that the calculation amount for matching in the speech recognition means is reduced and the fuzzy calculation is used. This enables arithmetic processing with the optimal number of standard cepstrum patterns that dynamically corresponds to the environment. As the fuzzy operation, the background noise information at the time of voice input is used as the input information of the antecedent part, or the numerical information in which a part of each word of the standard cepstrum pattern is a part of another word is used as the input information of the antecedent part. You can

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows a block diagram of a voice recognition device according to an embodiment of the present invention.

The voice recognition apparatus shown in FIG. 1 is a word type voice recognition apparatus for recognizing a voice for each word. For example, when the pronunciation of "Tokyo" is pronounced, "Tokyo" is recognized. The voice input unit 1 is a microphone that converts the vibration of the air emitted through the human vocal tract into an electric analog signal, or a digital signal by sampling the electric analog signal converted by the microphone at a predetermined cycle. The A / D converter etc. which convert into audio data are provided. The voice section extraction unit 2 determines a voice section from the digital voice data output from the voice input unit 1 using power information and the like. The power information is threshold information for identifying a sound part and a silent part. Next, in the feature quantity extraction unit 3, a linear prediction method or a fast Fourier transform (FF) is performed from the digital voice data of the determined voice section.
T), the discrete Fourier transform (DFT) method or the like is used to calculate the time series of the spectral feature amount for each short-time frame. A cepstrum obtained by using FFT or DFT can be used as the feature amount. As is well known, the cepstrum calculates the logarithm of the power spectrum obtained by FFT (or DFT) processing speech data, and further
T) Defined as having undergone the inverse transformation process. The time series of the cepstrum, which is sequentially calculated for each short-time frame, is input to the preliminary selection unit 5 by linear matching, or is input to the preliminary selection unit 5 after being normalized by the feature amount normalization unit 4, and this preliminary selection is performed. In Part 5, it is compared with a standard cepstrum pattern (hereinafter referred to as a standard pattern). The preliminary selection unit 5 compares a large number of standard patterns stored in the standard pattern storage unit 6 with the input cepstrum by linear matching, and preliminarily selects (narrows down) some standard patterns as candidates. .
In this case, each standard pattern is ranked, some of the standard patterns in the upper ranks are selected as candidates, and output by word number. In the preliminary selection unit 5 that performs such an operation,
The calculation amount is small and the calculation cost is low because the calculation of the linear matching is sufficient and the standard pattern candidates can be selected with some rough ranking.

The standard pattern storage unit 6 stores a large number of standard patterns in advance. These patterns are output to the preliminary selection unit 5 and the pattern recognition unit 7 based on nonlinear matching.

Pattern recognition unit 7 by nonlinear matching
Then, the pre-normalized or post-normalized cepstrum output from the feature quantity extraction unit 3 or the feature quantity normalization unit 4 and some of the upper standard patterns of the ranked standard patterns, that is, the preliminary Pattern recognition by DP matching is performed based on the standard pattern corresponding to the word number output from the selection unit 5, and the result is output to the result output unit 8 as a recognition candidate. That is, the standard pattern that is closest to the input cepstrum is selected and output to the display unit such as a display. In this embodiment, DP matching is used for the non-linear matching performed by the pattern recognition unit 7.
It is also possible to use a probabilistic model of HMM (Hidden Markov Model) or a method such as a neural network.

The preliminary selection unit 5 based on the linear matching may perform preliminary selection by comparing the cepstrum output from the feature amount extraction unit 3 with the standard pattern stored in the standard pattern storage unit 6. In this case, it is necessary that the number of cepstrum output from the feature amount extraction unit 3 be approximately the same when the same word is used. In this embodiment, each cepstrum output from the feature amount extraction unit 3 is temporally normalized by the feature amount normalization unit 4. The normalization method in the feature quantity normalization unit 4 will be described below.

The cepstrum obtained by the feature amount extraction unit 3 is extracted for each frame for a short time during the voice section T. If the time series of this cepstrum is not normalized, the time axis is calculated. Since the cepstrum will be calculated and generated for each frame in the short time above,
As shown in FIG. 2, when the same word occurs at different times, the number of extracted cepstrum differs. That is, in the example shown in FIG.
Although a total of three cepstrums of 1, T2, and T3 are extracted and generated, in the example shown in FIG. 1B, a total of four cepstrums of T1, T2, T3, and T4 are extracted and generated. Will be done. Therefore, if this is left as it is, there is a problem that the recognition performance at the time of performing pattern recognition is deteriorated and the reliability is deteriorated.

Therefore, in the embodiment of the present invention, the feature amount normalization unit 4 is provided and the time series of the cepstrum obtained by the feature amount extraction unit 3 is normalized along the locus in the cepstrum space. This normalization method is referred to here as SPT (Succ
essive Processing along Trajectory) method.
FIG. 3 shows a configuration diagram of a voice recognition device in which an SPT processing unit 4 ′ is arranged as the feature amount normalization unit 4. In addition, as the pattern recognition unit 7 by nonlinear matching,
A pattern recognition section 7'by DP matching is arranged.

FIG. 4 is a diagram for explaining a method of normalizing the SPT type cepstrum. A normalization method of the cepstrum in the SPT processing unit 4 '(hereinafter referred to as SP
The T method) includes the following procedures (1) to (6).

(1) The length that divides the locus in the cepstrum space is determined as a fixed length.

(2) A locus between the start point and the end point of the input voice data is drawn and its length is obtained. However, the end point in this case is not the end point of the voice section but a position (position of the cepstrum) appropriately determined.

(3) The starting point is the position of the first normalized cepstrum.

(4) Until the end point is exceeded (5), (6)
repeat.

(5) Proceed a fixed length along the locus from the normalized cepstrum that has just been determined.

(6) The cepstrum at that position is generated by linear interpolation and used as the normalized cepstrum.

By this procedure, for example, if the start point is C1 and the end point is C5 as shown in FIG. 4 (C5 is not the end point of the voice section), the position of the normalized cepstrum S1 generated first is set. Becomes the same as C1 and the position of the second normalized cepstrum S2 is calculated and generated by linear interpolation based on the positions of at least three pre-normalized cepstrum C1 to C3 as shown in FIG. The linear interpolation is performed by a general internal division method, and in the example shown in FIG.
The length from S2 to S2 is a fixed length. In this way, the cepstrum normalized by linear interpolation by the internal division method is sequentially
By calculating and generating, in the example shown in FIG. 4, a total of four normalized cepstrum S1 to S4 are generated on the locus from the start point C1 to the end point C5.

FIG. 6 is a flowchart showing a specific processing method of the SPT method.

In the figure, C1, C2, ..., Cm
Are cepstrums that are vector quantities obtained from the feature quantity extraction unit 3, and S1, S2, ... Sn represent cepstrums of vector quantities that have been SPT-processed and normalized. Also, WholeLength indicates the total distance when moving along the trajectory of the cepstrum, and FixedLength is a preset fixed length.

First, in step ST1, initialization is performed,
Set FixedLength to the variable NextLength. S
At T2, it is determined whether or not the end point has been reached. Note that m
Is a value indicating the end point and can be appropriately determined. In the example shown in FIG. 4, it is 5. Next, in ST3, the distance d between the pre-normalized cepstrums is obtained. In the example shown in FIG. 4, this distance d in the first processing is the distance between C2 and C1. In addition,
For this distance d, for example, Euclidean distance is used. In ST3, WholeLength is updated by adding d. Next, in ST4, WholeLengt
The comparison between h and NextLength is performed. That is, it is determined whether or not the position (WholeLength) that has advanced on the cepstrum locus exceeds the position (NextLength) of the normalized vector to be generated next. At the first stage shown in FIG. 5, WholeLength represents the position of C2 and NextLength represents the position of S2, so in this ST4, it becomes no,
The counter i is incremented and the process returns to ST2. At the next stage, WholeLength becomes the position of C3, so ST4
Then, yes, and the process proceeds to ST5. ST5 is a process of performing linear interpolation by the internal division method to determine the position of the normalized cepstrum S2. In the same manner, S3 and S4 and the respective normalized cepstrums are sequentially calculated and generated.

In the above-mentioned SPT method, since the fixed length is constant, the number of cepstrum is large in a portion having a lot of information to be recognized, such as a consonant portion (a portion that changes drastically), so that speech recognition is high. Performance can be improved.
Further, since the cepstrum is sequentially normalized without waiting for the end point of the voice, there is an advantage that the sequential processing becomes possible and the processing speed of the system becomes faster.

In FIG. 3, the preliminary selection unit 5 by linear matching compares the cepstrum normalized by the SPT processing unit 4'with the standard pattern stored in the standard pattern storage unit 6 by linear matching, and A rough rank is assigned from similar standard patterns, and candidate word numbers from the first rank to a fixed rank are sent to the pattern recognition unit 7'by DP matching. In the pattern recognition unit 7 ', the cepstrum normalized by the SPT processing unit 4',
The standard pattern corresponding to the word number sent from the preliminary selection unit 5 is compared by DP matching, and the word number of the closest standard pattern is output to the result output unit 8 as a recognition candidate. The result output unit 8 outputs a word corresponding to the obtained word number as a recognition candidate to a display or the like.

By normalizing the cepstrum in the SPT processing section 4'as described above, the number of cepstrum is dense in the consonant part of the voice and is coarsely generated in the vowel part. On the other hand, the difference in the temporal length of words is mainly due to the length of vowels. Therefore, when normalizing the cepstrum by the SPT method,
Since the influence of the length of the vowel can be reduced, even if there is a difference in the temporal length of the word, the difference can be absorbed to some extent, and the number of normalized cepstrum numbers is often the same. FIG. 7 shows waveforms of "Tokyo" spoken by different speakers. The horizontal axis corresponds to the time axis, and the vertical line is SP
The generation point of the normalized spectrum by the T method is shown. The upper side is 0.64 seconds and the lower side is 0.57 seconds, but 22 cepstrums are generated by the SPT processing in both cases, and the difference in the length of time is absorbed.

That is, even if there is a difference in the time length of words, by normalizing the cepstrum by the SPT method, effective preliminary selection can be performed even in linear matching with a low calculation cost. Linear matching of vector sequences is performed by using two equal number of vector sequences v ₁ , v ₂ , ... _, v _n
And v ₁ ', v ₂ ', ... _, v _n '

[0035]

[Equation 1]

The calculation cost is only proportional to m and can be smaller than that of DP matching.
In DP matching, since m vector sequences and n vector sequences (generally m ≠ n) are handled, the calculation cost is proportional to m × n.

Now, the sequence of the cepstrum obtained after being normalized by the SPT processing unit 4'is s ₁ , s ₂ , ... _, S
_{Let n} be the standard pattern of a certain word (for example, “Tokyo”) be r ₁ , r ₂ , ... _, R _k . If d (v, w) is a function representing the distance between the vectors v and w,

[0038]

[Equation 2]

And n−k + 1 different vector sequences
Gives the distance of d₁, d_Two, ..._, d _{n-k + 1}The smallest of
Value d_minWith the standard pattern for this word
Use linear matching distance. Figure 8 shows a linear match
Time axis of time series of standard pattern r and input pattern s
Is shown. Such a procedure for all standards
Do about the pattern, d_min1 in ascending order
Each d from rank to predetermined rank_minStandard putter corresponding to
As a candidate and the word number as a result of the preliminary selection.
Output.

In FIG. 3, the word number of the preliminary selection candidate obtained by the linear matching by the preliminary selection unit 5 is D.
The pattern is sent to the pattern recognizing unit 7 ′ by P matching, and here, by the well-known DP matching, S
The pattern recognition of the normalized cepstrum obtained by the PT processing unit 4'is performed. The standard pattern used at this time is only the standard pattern corresponding to the word number of the preliminary selection candidate, and all the standard patterns stored in the standard pattern storage unit 6 are not used. Therefore, the calculation cost in the pattern recognition unit 7'is reduced, and the processing speed can be increased. Instead of DP matching, it is also possible to use a method such as pattern recognition using a probabilistic model of HMM or a neural network.

FIG. 9 shows another embodiment of the present invention.
This will be described with reference to FIG. In the speech recognition apparatus of the embodiment shown in the figure, the SPT processing unit 4'calculates and generates a normalized cepstrum (L1) having a fixed length L1 and a normalized cepstrum (L2) having a fixed length L2. Fixed length L
1 and L2 have a relationship of L1> L2. Further, in place of the preliminary selection unit 5 based on linear matching, a preliminary selection unit 5 ′ based on continuous DP matching is provided. Then, the normalization cepstrum (L1) having a long fixed length is input to the preliminary selection unit 5'by continuous DP matching, and the normalization cepstrum (L2) having a short fixed length is input to the pattern recognition unit 7'by DP matching. The continuous DP matching is a method of performing DP matching while shifting the standard pattern by one frame from the starting end of the input pattern.
You can do word spotting to extract words from sentences.

In the case of the SPT method, since the cepstrum is generated every time the fixed length advances on the cepstrum locus, the length of this fixed length determines the accuracy and at the same time, if the fixed length is lengthened, the number of cepstrums will increase. Is reduced. In the voice recognition apparatus according to the present invention, when the preliminary selection is performed, the rough order of the standard patterns may be set, and therefore the preliminary selection unit does not need the same accuracy as the final recognition accuracy performed by the pattern recognition unit 7 '. Therefore, in the SPT processing unit 4 ', the normalized cepstrum (L
Increase the fixed length to generate 1) and reduce the number of cepstrum generated. By reducing the number of cepstrums, it is possible to reduce the calculation cost due to the DP matching in the preliminary selection unit 5 '. Now, consider a case where the fixed length L1 in the SPT processing unit 4'is doubled with respect to the fixed length L2. Then, since the number of normalized cepstrums is halved, the position of the point to be calculated by DP matching in this case is only the position corresponding to the black point in FIG. 10, which is calculated as compared with the case where the fixed length is not doubled. The amount is one quarter.
The preliminary selection unit 5'decimates the standard patterns to reduce the number of cepstrum. FIG.
Shows the time series correspondence between the normalized cepstrum input pattern s ′ and the standard cepstrum pattern (standard pattern) r ′. In the present embodiment, the preliminary selection unit 5'performs continuous DP matching, but this is for performing word spotting for extracting words from a series of sentences. That is, the preliminary selection unit 5'performs word spotting, and the pattern recognition unit 7'performs DP matching on the extracted word. The actual procedure in the preliminary selection unit 5'by the continuous DP matching will be described below.

Sequence of normalized cepstrum (L1) for preselection obtained by lengthening fixed length L s ₁ ', s ₂ ', ... _, s
_{For n} 'and a standard pattern r ₁ ', r ₂ ', ... _, r _k ' for preselection of a word (for example, "Tokyo"), a continuous D
P matching is performed and d _1, d ₂ , ... Are calculated. The smallest of d _1, d ₂ , ... _Is d _min . This procedure is performed for all the standard patterns, d _min is arranged in ascending order, and the standard patterns corresponding to the first to the predetermined order among them are used as the preliminary selection candidate word numbers for pattern recognition by DP matching. Output to the section 7 '.

The pattern recognition process after the preliminary selection is the same as that shown in FIG.

Next, a voice recognition apparatus according to another embodiment of the present invention will be described with reference to FIG. The voice recognition device shown in the figure is different from the device shown in FIG. 9 in that the preliminary selection unit 5'by continuous DP matching includes a fuzzy calculation unit 50. The fuzzy operation unit 50 determines which of the standard patterns from the first rank to which rank among the standard patterns roughly ranked by the preliminary selection unit 5'as the preliminary selection candidate. Then
As the input information of the antecedent part of the fuzzy rule, the S / N ratio and the number of partial candidates which are background noise information at the time of voice input are used. The S / N ratio indicates the ratio of a voice signal and a noise signal, and is detected by the S / N detector 9 based on the digital voice data from the voice input unit 1. In addition, the number of partial candidates means “Chiba” and “Nishichiba” for “Higashichiba”
It represents the number of words such as "Honchiba" and "Higashiomiya" that are part of other recognition target words (in this case, the number of partial candidates for "Higashichiba" is 4). This information is obtained from the standard pattern storage unit 6. The rule in the fuzzy operation unit 50 is, for example,
It is set as follows:

IF (small number of partial candidates) THEN threshold value is small IF (normal number of partial candidates) THEN threshold value is medium IF (large number of partial candidates) THEN threshold value is large IF (S / N ratio is small = noise is large) Increases the THEN threshold IF (S / N ratio is normal = noise is normal) THEN Threshold is medium IF (S / N ratio is large (= Noise is small) The THEN threshold value is made small. The threshold value of the consequent part is a value for determining how many standard patterns up to the standard selection candidate.
The larger the number, the larger the number of candidates (the number of rounded up thresholds may be the number of candidates itself). Figure 1 shows the membership functions used for the antecedent part (input side) and consequent part (output side).
It is shown in FIG.

The fuzzy operation is performed as follows.

(1) From the number of partial candidates and the S / N ratio, PS
Find out how many c, PMc, PLc, PSs, PMsPLs are.

(2) S = max (PSc, PLs), M
= Max (PMc, PMs), L = max (PLc, P
Ss) (3) Threshold value t = (PSt × S + PMt × M + PLt
XL) / (S + M + L) For example, if the number of partial candidates is 2 and the S / N ratio is 100, (1) PSc = 0.5, PMc = 0.5, PLc = 0.
0, PSs = 0.0, PMs = 0.0, PLs = 1.0 (2) S = max (0.5,1.0) = 1.0, M = m
ax (0.5,0.0) = 0.5, L = max (0.
0,0.0) = 0.0 (3) Threshold value t = (10 × 1.0 + 30 × 0.5 + 5)
0 × 0.0) / (1.0 + 0.5 + 0.0) = 16.7, and the threshold value at this time is 16.7. If the number of rounded up thresholds is the number of standard patterns to be preselected, the number is 17.

The above fuzzy rule is used to graph the relationship between the S / N ratio and the threshold value in the case of partial candidate numbers 1, 2, 4, 5, and 6 as shown in FIG. It can be seen that the change is smooth according to the S / N ratio.

By providing the fuzzy operation section 50 in this way, a number of candidates suitable for the surrounding environment are sent to the pattern recognition section 7 '. Therefore, for example, in the noisy environment, the standard preselected is selected. The optimum number of standard patterns is preselected according to the environment, such as the number of patterns is large and the number of standard patterns preselected in an environment where there is not much noise is reduced, which improves system efficiency. There is a good effect.

[0052]

According to the present invention, a plurality of standard patterns are preliminarily selected from the feature amount calculated by the feature amount calculating means by linear matching, and then voice recognition is performed by nonlinear matching. This has the effect of increasing the processing speed.

Further, by calculating a cepstrum as the above feature amount and further normalizing the intervals of the cepstrum on the cepstrum locus to be a fixed length set in advance, preselection of a standard pattern is carried out, whereby The difference in time length can be absorbed to some extent, and the above preliminary selection can be made highly accurate even in the case of linear matching.

Further, by making the fixed length of the cepstrum input to the preliminary selecting means longer and normalizing it, the calculation cost in the preliminary selecting means can be reduced.

Further, by dynamically changing the number of preselected standard cepstrum patterns by fuzzy calculation, it is possible to perform voice recognition using an optimum number of candidates according to the situation at that time such as the surrounding environment. it can.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram for explaining that the number of cepstrum varies depending on the difference in the voice section T.

FIG. 3 is a configuration diagram of a voice recognition device according to another embodiment of the present invention.

FIG. 4 is a diagram for explaining a cepstrum normalization method in the SPT method.

FIG. 5 is a diagram illustrating details of a cepstrum normalization method in the SPT method.

FIG. 6 is a flowchart showing the operation of the SPT method.

FIG. 7 is a diagram for explaining that the number of cepstrum generated in the SPT method is not affected by the generation time length.

FIG. 8 is a diagram for explaining a method of obtaining a distance between vector sequences by linear matching.

FIG. 9 is a configuration diagram of a voice recognition device according to another embodiment of the present invention.

10 is a sequence diagram of the speech recognition apparatus shown in FIG.
The figure explaining the method of calculating | requiring the distance of vector sequences by P matching.

FIG. 11 is a configuration diagram of a voice recognition device according to another embodiment of the present invention.

FIG. 12 is a diagram showing a membership function used in a fuzzy operation unit.

[FIG. 13] S / N by using a fuzzy operation unit
Graph showing that the threshold changes with the ratio

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kenya Miura 606, Omron Software Co., Ltd.

Claims (8)

[Claims] 1. Linear matching is performed by a feature amount calculating means for calculating a feature amount of input voice data, and the calculated feature amount and a standard pattern stored in advance.
It comprises a preliminary selecting means for preliminarily selecting a standard pattern according to the result, and a voice recognizing means for performing voice recognition by performing non-linear matching by the feature amount and the preliminarily selected standard pattern. Speech recognizer. 2. A cepstrum calculating means for calculating a cepstrum for each short-time frame of input voice data, and an interval between each cepstrum on the trajectory of each calculated cepstrum is a preset fixed length. Cepstral normalizing means for normalizing each cepstrum by linear interpolation, and linear matching is performed with the standardized cepstrum pattern and the standard cepstrum pattern stored in advance, and the standard cepstrum pattern is preliminarily determined according to the result. A voice recognition device comprising: preselection means for selecting; and voice recognition means for performing voice recognition by performing non-linear matching with the normalized cepstrum and the preliminarily selected standard cepstrum pattern. 3. A cepstrum calculating means for calculating a cepstrum for each short-time frame of input voice data, and fixed lengths L1 and L2 in which a distance L between the cepstrum is preset on the calculated trajectory of each cepstrum. (L1
> L2), the cepstrum normalization means for normalizing each cepstrum into two types by linear interpolation, the pattern of the cepstrum normalized at intervals of the fixed length L1 and the standard cepstrum pattern stored in advance Preselection means for preliminarily selecting a standard cepstrum pattern in accordance with matching or non-linear matching, a cepstrum normalized at the fixed length L2 interval, and the preliminarily selected standard cepstrum pattern. A voice recognition device comprising: a voice recognition means for performing non-linear matching by performing voice recognition. 4. A fuzzy calculation means for determining the number of standard cepstrum patterns preliminarily selected by the preliminary selection means by fuzzy calculation using background noise information at the time of voice input as input information of the antecedent part. The voice recognition device according to any one of claims 1 to 3, characterized in that. 5. The number information of the standard cepstrum patterns preliminarily selected by the preselection means, and the numerical information in which a part of each word of the standard cepstrum pattern is a part of another word. The voice recognition device according to any one of claims 1 to 3, further comprising fuzzy calculation means that is determined by fuzzy calculation. 6. A feature quantity of input voice data is calculated,
Linear matching is performed using the calculated feature amount and a standard pattern stored in advance, and a standard pattern is preliminarily selected according to the result, and the feature amount and the preliminarily selected standard pattern are further selected. A voice recognition method characterized by performing non-linear matching with and voice recognition. 7. A cepstrum for each short-time frame of input voice data is calculated, and each cepstrum is linearly interpolated on the trajectory of each calculated cepstrum so as to have a preset fixed length. Normalizing the cepstrum, performing a linear matching with the normalized pattern of the cepstrum and a prestored standard cepstrum pattern, and preliminarily selecting a standard cepstrum pattern according to the result, and further performing the normalization. A voice recognition method comprising performing non-linear matching by using the cepstrum and the preliminarily selected standard cepstrum pattern to perform voice recognition. 8. A cepstrum for each short-time frame of input voice data is calculated, and fixed intervals L1 and L2 (L1> L2) in which the interval L of each cepstrum is preset on the trajectory of each calculated cepstrum. ), Each cepstrum is normalized into two types by linear interpolation, and non-linear matching is performed using the pattern of the cepstrum normalized at the interval of the fixed length L1 and the standard cepstrum pattern stored in advance, and the result is obtained. Preliminarily select the standard cepstrum pattern in accordance with the fixed length L2
A voice recognition method characterized by performing a non-linear matching by a standardized cepstrum pattern at a predetermined interval and the preliminarily selected standard cepstrum pattern to perform voice recognition.