JPS6129519B2 - - Google Patents

Abstract

PURPOSE:To increase a processing speed by extracting a sampling point which corresponds to a vowel part in a monosyllable, by selecting a recognition-expectant monosyllable, through the collation with the previously-registered reference vowel feature quantity, and by collating the monosyllable with an unknown voice. CONSTITUTION:By making use of the steadiness of the extracted feature quantity, steady-vowel-part extraction circuit 6 extracts a sampling point which corresponds to a vowel part in a monosyllable and then collates 10 a vowel feature quantity, which corresponds to this sampling point, with the reference vowel feature quantity previously registered for each monosyllable. As a result, the recognition-expectant monosyllable is determined and then collated 13 again with an unknown voice. Consequently, the processing speed can be increased.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention provides a monosyllabic speech recognition device, particularly a speech recognition device that performs recognition processing by extracting features based on frequency analysis results of a speech signal. is used to extract sample points corresponding to the vowel part in a monosyllable, select candidate monosyllables to be recognized using the vowel features corresponding to the sample points, and perform secondary matching on the candidate monosyllables. The present invention relates to a monosyllabic speech recognition device that improves processing speed in this way.

In the speech recognition device, parameters that effectively represent the characteristics of each phoneme are extracted using the frequency analysis results of the speech signal, and the extracted parameters and the parameters of the pre-registered phoneme corresponding to the registered word are used. The unknown input voice is recognized by comparing it with the unknown input voice. That is, as the above parameters, for example, the first formant frequency and the second formant frequency are sampled and used. However, in the above-mentioned verification, if a large number of sampling points is used to perform more precise verification, the time required for the verification processing becomes long.

For this reason, we focused on the fact that there are sections where the above parameters change rapidly over time and sections where they change gently, and by sampling densely in the former section and coarsely sampling in the latter section, A method is being considered that increases the recognition rate with a smaller number of samples by sampling at non-uniform sampling points (Patent Application No. 43972 of 1972).

Although the present invention is not limited to the above method of sampling at non-uniform sampling points, in the case of this method, the point where the above parameter changes gradually over time, that is, the point with stationarity, corresponds to a vowel. The purpose is to improve processing speed by narrowing down recognition target candidates in advance by using the fact that they correspond to syllables, and then performing more detailed matching processing. Furthermore, it is an object of the present invention to shorten the verification processing time when performing more detailed verification processing than the above. Therefore, the monosyllabic speech recognition device of the present invention extracts the feature quantity of the speech signal using the frequency analysis result of the speech signal, and performs monosyllabic speech recognition in accordance with the feature quantity. The device includes: a vowel stationary part extraction function unit that extracts sample points corresponding to each sound part in a single syllable by utilizing the stationarity of the extracted feature values; In the unknown input speech obtained based on the sample points extracted by the vowel constant part parameter registration unit which registers the vowel reference feature amount for each single syllable in advance, and the vowel constant part extraction function unit. at least a vowel constant part matching/candidate selection function unit that determines a recognition target candidate monosyllable by comparing the vowel feature quantity and the vowel reference feature quantity registered in the vowel constant part parameter registration unit; The vowel constant part extraction function unit is configured to perform a secondary comparison of the determined recognition target candidate monosyllable with the unknown input speech, and the vowel constant part extraction function unit sequentially extracts the cumulative variation amount AV (tn) of the input speech. A weight corresponding to the cumulative number of times the non-uniform sampling point has been calculated and accumulated, and the non-uniform sampling point at which the cumulative amount of variation has reached a predetermined threshold is determined, and the non-uniform sampling point for each non-uniform sampling point has been determined. is characterized in that it is configured to extract the sample points with the maximum number of non-uniform sampling points. This will be explained below with reference to the drawings.

Figure 1 shows the configuration of one embodiment of the present invention, and Figure 2 shows the configuration of the first embodiment.
An example configuration of the vowel stationary part extraction circuit shown in the figure,
FIG. 3 shows an embodiment of the structure of the portion shown in FIG. 1 shown in dotted lines.

In the figure, 1 is the extracted feature amount, and 2 is a group of band filters, which input the input audio into N channels, for example, 15 channels of frequency signals P ₁ (t), P ₂ (t), ...,
P _N (t), and 3 is a parameter extraction unit that extracts feature quantities (parameters) of the input speech, that is, monosyllabic speech, such as the moment M ₁ corresponding to the first formant frequency and the moment corresponding to the second formant frequency. 4 is a sampling time determination circuit that extracts feature quantities effective for matching monosyllables based on the parameter time series matching judgment unit described later, including M ₂ and further including low frequency power and high frequency power; It represents what determines the non-uniform sampling points T ₀ , T ₁ , . . . as shown in relation to the illustrated feature quantity 1. 5 is a non-uniform sampling circuit section, which is a sampling point shown in feature amount 1 above.
T ₀ , T ₁ ...T〓...corresponding to the timing T〓 shown in the above feature amount 1. 7 and 8 are respectively switching circuits that switch between a registration mode for registering information corresponding to registered speech and a recognition mode for recognizing unknown input speech. , 9 is a vowel constant part parameter registration unit which stores the vowel reference feature corresponding to each registered monosyllable, and 10 is a vowel constant part matching/candidate selection unit which stores the vowel feature in the above-mentioned vowel feature in the recognition mode. This represents the selection of recognition target candidate monosyllables based on this. Furthermore, 11 is a parameter time series registration unit which stores the above-mentioned sampling points T ₀ ,
12 is a candidate selection circuit section which stores the feature amounts (parameters) corresponding to T ₁ . 13 is a parameter time series matching/determining unit that guides the time series information of reference feature values (parameters) corresponding to the plurality of parameters) to the next parameter time series matching/judging unit 13, which guides the time series information of reference feature values (parameters) corresponding to Comparison of feature quantity (parameter) time series information with that of the above candidate monosyllable, 1
Reference numeral 4 represents an output circuit that outputs the recognized category name, and 15 represents a control unit that controls the entire apparatus.

As is well known, the parameter extraction circuit 3 extracts data every time a clock pulse of a certain period occurs. The feature amount corresponding to the first formant frequency, the second formant frequency, etc. is calculated in accordance with the calculation shown in , and the result is stored in a register (not shown). In the above equation (1), Pi (tn) is the output of the i-th filter sampled at time tn every 10 msec, Wij is its load,
Fi represents its center frequency. The load Wij may be considered to be determined experimentally so that the quantities M ₁ and M ₂ coincide with the first and second formant frequencies. Of course, in the parameter extraction circuit 3, the first formant frequency and the second
In addition to the formant frequency, other feature quantities are extracted, but in order to simplify the explanation, the first formant frequency and the second formant frequency will be described below as typical feature quantities.

The sampling time determination circuit 4 is The cumulative variation amount AV(tn) defined by is calculated at the cycle of calculating the feature amounts M ₁ and M ₂ shown in the above equation (1), and the non-uniform sampling point tnk is determined. That is, it is monitored whether the cumulative variation amount AV(tn) exceeds a predetermined threshold value, and the time point tnk at which the threshold value is exceeded is determined as the k-th non-uniform sampling point.

In the above equation (2), V(tn) is the amount of variation in the output of the filter, and is defined by the following equation.

From the above, it can be seen that the cumulative variation amount AV(tn) is an accumulation of parameter changes after a certain non-uniform sampling point occurs. It can be seen that when the cumulative variation AV(tn) exceeds a certain threshold, the next non-uniform sampling point is determined, and at that point the previous cumulative variation AV(tn) is reset. As a result, it can be seen that the non-uniform sampling points T ₀ , T ₁ . . . T〓, .

The non-uniform sampling circuit 5 samples the extracted feature quantities M ₁ and M ₂ for each non-uniform sampling point and stores them in the parameter time series registration unit 11 shown in FIG. 1 in the registration mode, Also, in the recognition mode, the parameters are input to the parameter time series comparison determination section 13.

The vowel stationary part extraction circuit unit 6 (i) counts the cumulative number of times while the non-uniform sampling points T ₀ , T ₁ . . . are being determined, and (ii) uses the counted value as a weight W. Then, the maximum non-uniform sampling point of the weight W (in the case of feature amount 1, the sampling point T) is determined. That is, sample points are determined. Then, the vowel features obtained corresponding to the sample point are
In registration mode, vowel stationary part parameter registration part 9
Also, in the recognition mode, it is supplied to the vowel constant part matching/candidate selection unit 10.

In the recognition mode for recognizing monosyllabic speech that is unknown input speech, processing is performed as follows. At this time, the feature amount time series information corresponding to the registered monosyllable is stored in the registration unit 11 shown in FIG. 1, and the vowel feature amount corresponding to the registered monosyllable is stored in the registration section 11 shown in FIG. It is stored in 9.

(1) Corresponding to the input unknown input voice, the non-uniform sampling point T is determined as described above via the band filter group 2, the parameter extraction circuit 3, the sampling time determination circuit 4, and the non-uniform sampling circuit section 5. ₀ , T _{1 , .} . . are supplied to the parameter time series comparison determination unit 13.

(2) On the other hand, the vowel constant extraction circuit section 6 supplies the vowel feature amount corresponding to the extracted sample point T as described above to the vowel constant section matching/candidate selection section 10.

(3) At this time, under the control of the control unit 15, the vowel reference features corresponding to the registered monosyllables are sequentially read out from the illustrated registration unit 9 to the illustrated matching/candidate selection unit 10 for each registered monosyllable. Ru.

(4) The matching/candidate selection unit 10 uses the read vowel reference feature amount and the vowel constant part extraction circuit unit 6.
The vowel features supplied from Then, a plurality of matched monosyllables are selected as candidate monosyllables, and the candidate selection circuit 12 is notified.

(5) Next, the candidates are narrowed down to the selected single syllables and then detailed, ie, secondary matching processing begins. That is, the control unit 15 reads the parameter time series registration unit 11. Through this reading, time series information is output from the registration unit 11 for each registered monosyllable, but at this time, the candidate selection circuit 12 selects the reference time series corresponding to the candidate monosyllable selected in the above process (4). Only the information is selected and transmitted to the parameter time series comparison determination section 13.

(6) In the determination unit 13, the feature amount time series information corresponding to the unknown input voice is previously input from the non-uniform sampling circuit unit 5. The information has been input. Then, the information is compared with the reference time series information. The verification process is performed using, for example, known dynamic programming.

(7) Then, if the registered monosyllable is matched well, the unknown input speech is determined to belong to the registered monosyllable, and is set in the output circuit 14 as a recognition result.

Figure 2 shows the vowel stationary part extraction circuit 6 shown in Figure 1.
The configuration of one embodiment is shown. In the figure, 16 represents a comparison circuit, 17 represents a maximum weight value register, and 18 represents a maximum weight value parameter register.

The arrangement shown in FIG. 2 operates as follows. That is, (8) Now, the non-uniform sampling circuit 5 shown in FIG.
Based on the above non-uniform sampling points
It is assumed that weights and parameters corresponding to T ₀ , T ₁ , . . . have been obtained.

(9) In this state, the uneven sampling point T ₀ ,
Weights and parameters corresponding to T ₁ , . . . are input in order. Initially, the weight maximum value register 17 is cleared to zero.

(10) When the weight corresponding to the non-uniform sampling point T ₀ is input to the comparison circuit 16, the weight
W ₀ is compared with the contents of register 17 Rmax.
Since the weight corresponding to the non-uniform sampling point _T0 is larger than the contents of the register circuit 16, the weight is set in the register 17, and the parameter P0 corresponding to the non-uniform sampling point _T0 is set in the register circuit ₁₆ . It is set to 18.

(11) Below, when the weight W ₁ corresponding to the non-uniform sampling point T ₁ is input, the condition that W ₁ ≧Rmax is checked. If the condition is satisfied, the weight W ₁ is set in the register 17 and the parameter P ₁ is set in the register 18. However, if the above conditions are not satisfied, the contents of registers 17 and 18 are not changed.

(12) In this way, the weights and parameters corresponding to the non-uniform sampling points are inputted one after another, and finally the weight Wα corresponding to the non-uniform sampling point T〓 having the maximum weight is stored in the register 17.
The parameter Pα is also set in the register 18.

(13) Then, the parameter Pα corresponding to the uneven sampling point T with the largest weight is supplied via the switching circuit 8 to the registration section 9 or the selection section 10 shown in FIG.

FIG. 3 shows an embodiment of the structure of the part shown in FIG. Reference numerals 9 and 10 in the figure correspond to those in FIG.
19 and 20 are address counters, 21 is a collation processing unit, 22 is a comparison circuit, 23 is a candidate vowel parameter register, 24 is parameter time series registration unit address information, 25 is a selection circuit, and 26 is candidate registration unit address information. It represents the storage section.

In the recognition mode, vowel reference feature amounts of a plurality of registered single syllables are stored in the vowel constant part parameter registration unit 9. In the illustrated case, the operation is as follows. That is, (14) In this state, when the vowel feature of the unknown input speech is supplied through the switching circuit 8, the counter 19 sequentially reads out the vowel reference feature of each registered monosyllable from the registration section 9. Ru. Then, the distance from the vowel feature of the input voice is measured. That is, the verification processing unit 21 sequentially performs verification.

(15) If a relatively good match is obtained as a result of the matching, the vowel reference feature is supplied to the comparison circuit 22. At this time, the counter 20 starts counting, and the previously stored candidate vowel feature quantities are sequentially read out from the candidate vowel parameter register 23 and supplied to the comparison circuit 22.

(16) In the comparison circuit 22, the verification processing section 2
The vowel features supplied from 1 are stored in register 2.
It is compared with the candidate vowel features from 3. If there is no matching vowel feature, the vowel feature is stored in the register 23 as a candidate vowel feature. Further, if a matching item is already a candidate, it will not be stored in the register 23.

(17) In the above process (16), when a candidate is newly stored in the register 23, the comparison circuit 22 issues an instruction to the selection circuit 25. As a result, the selection circuit 25 selects a registration unit (registration unit 1 shown in FIG.
The address information 24 of 1) is set in the address information storage section 26 of the candidate registration section. That is, in the information storage section 26, storage address information of feature time series information of registered monosyllables corresponding to the candidate vowel feature amount stored in the register 23 is set. The address information is used for selectively supplying the reference feature amount corresponding to the candidate monosyllable to the matching determination unit 13 shown in FIG. 1 during the secondary matching.

As described above, candidate monosyllables are selected, and thereafter, secondary matching is performed on the monosyllables narrowed down as candidates. In this case, needless to say, the feature amount time-series information of the unknown input speech and the reference feature amount time-series information of the registered monosyllable are compared in chronological order. The matching processing time in this case is roughly proportional to the square of the number N of time series information.

In the case of the present invention, the following processing mode is adopted during the secondary verification processing to further shorten the verification processing time. That is, in performing the secondary verification, verification is performed starting from the first information in the feature quantity time-series information, but verification is performed using the time-series information up to the sample point T〓. By doing this, the matching time is shortened (for example, if the number of information is halved, the time is shortened to 1/4).
Furthermore, the power of the feature values after the sample point is low and there is generally a lot of variation. This has sometimes been one of the causes of recognition errors when matching is performed using the widely varying feature quantities, but this point is also improved. Of course, what kind of feature quantity is used in the secondary matching process is arbitrary, but in the so-called primary matching, which is directly related to the present invention, only vowels are matched.

As explained above, according to the present invention, the processing time for speech recognition processing is significantly reduced. In particular, when a sampling method using non-uniform sampling points is adopted, sample points can be easily extracted.

[Brief explanation of the drawing]

Figure 1 shows the configuration of one embodiment of the present invention, and Figure 2 shows the configuration of the first embodiment.
An example configuration of the vowel stationary part extraction circuit shown in the figure,
FIG. 3 shows an embodiment of the structure of the portion shown in FIG. 1 shown in dotted lines. In the figure, 1 is the extracted feature amount, 5 is the non-uniform sampling circuit part, 6 is the vowel stationary part extraction circuit part, 9
10 represents a vowel constant part parameter registration unit, 10 represents a vowel constant part matching/candidate selection unit, 11 represents a parameter time series registration unit, 12 represents a candidate selection circuit, and 13 represents a parameter time series matching determination unit.

Claims (1)

[Scope of Claims] 1. A monosyllabic speech recognition device that uses frequency analysis results of a speech signal to extract feature quantities of the speech signal and recognizes unknown input speech in accordance with the feature quantities, a vowel stationary part extraction function unit that extracts sample points corresponding to the vowel part in a single syllable by using the stationarity of the feature values; A vowel stationary part parameter registration unit that registers vowel reference feature quantities, and a vowel feature quantity in the unknown input speech obtained based on the sample points extracted by the vowel stationary part extraction function unit and the vowel stationary part. a vowel stationary part matching/candidate selection function unit that determines a recognition target candidate single syllable by comparing it with vowel reference feature quantities registered in a part parameter registration unit; The vowel constant part extraction function section sequentially calculates and accumulates the cumulative variation amount AV (tn) of the input speech, and performs a secondary comparison with the unknown input speech for a single syllable. A non-uniform sampling point whose cumulative variation amount has reached a predetermined threshold is determined, and the non-uniform sampling point has the maximum weight corresponding to the cumulative number of times that the non-uniform sampling point has been determined for each non-uniform sampling point. A monosyllabic speech recognition device characterized in that it is configured to extract the sample points as described above. 2 The above-mentioned secondary verification is performed using the reference feature time series information from the tip of the recognition target candidate monosyllable to the sample point where the vowel reference feature was determined, and the reference feature time series information from the tip of the unknown input speech to the vowel feature. 2. The monosyllabic speech recognition device according to claim 1, wherein the monosyllabic speech recognition device is configured to compare the feature amount time series information up to the determined sample point.