JPH0573036B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD The present invention relates to a speech recognition device that recognizes speech. BACKGROUND TECHNOLOGY AND PROBLEMS Conventionally, as a speech recognition device that copes with variations in speech rate, there has been proposed a device that performs DP matching processing as disclosed in, for example, Japanese Unexamined Patent Publication No. 50-96104. First, a speech recognition device that performs speech recognition using this DP matching process will be described. In FIG. 1, 1 indicates a microphone as an audio signal input section, and the audio signal from this microphone 1 is supplied to an acoustic analysis section 2, where an acoustic parameter time series Pi(n) is obtained. It will be done. In this acoustic analysis section 2, for example, the rectified and smoothed output of the bandpass filter bank is converted into the acoustic parameter time series Pi(n) (i=1,...,I; I is the number of channels of the bandpass filter bank, n=1,...,
N; N is the number of frames cut out by voice section determination. ) is obtained as Acoustic parameter time series Pi(n) of this acoustic analysis section 2
is stored in the standard pattern memory 4 for each recognition target word in the registration mode by the mode changeover switch 3, and is supplied to one end of the DP matching distance calculation unit 5 in the recognition mode. Further, in this recognition mode, the standard pattern stored in the standard pattern memory 4 is supplied to the other end of the DP matching distance calculation section 5. The DP matching distance calculation unit 5 calculates the DP matching distance between the input pattern consisting of the acoustic parameter time series Pi(n) of the audio input at that time and the standard pattern in the standard pattern memory 4, and calculates the DP matching distance. A distance signal indicating the DP matching distance from the calculation unit 5 is supplied to the minimum distance determination unit 6, and the minimum distance determination unit 6 determines a standard pattern with the minimum DP matching distance for the input pattern, and the determination result is A recognition result indicating the input voice is obtained at the output terminal 7. Incidentally, the number N of frames of the standard pattern stored in the standard pattern memory 4 generally varies depending on variations in speaking speed and differences in word length. The DP matching process performs time axis normalization to deal with variations in speaking speed and differences in word length. This DP matching process will be explained below. Here, for simplicity, the acoustic parameter time series Pi
The dimension corresponding to the frequency axis direction i of (n) is omitted, and the parameter time series of the standard pattern is b ₁ ,...,b _N , and the parameter time series of the input pattern is a ₁ ,..., a _M , and the end point is fixed. DP matching processing in the case of DP-path will be explained. Figure 2 shows a conceptual diagram of the DP matching process.
Input parameters (M = 19) are arranged on the horizontal axis, standard parameters (N = 12) are arranged on the vertical axis, and the points on the (M, N) grid plane shown in Fig. 2 are M x There are N points, and each point corresponds to one distance. For example, the distance between a ₃ and b ₅ corresponds to the intersection of a straight line extending vertically from a ₃ and a straight line extending horizontally from b ₅ . In this case, if we take, for example, the Tiebishiev distance as the distance, the Tiebishiev distance d(3,5) between a ₃ and b ₅ is d(3,5)= _I 〓 ⁱ⁼¹ |P ^a _i (3)−P ^b _i (5)| (In this case, I=1 because the dimension corresponding to the frequency axis direction i is omitted.) and,
As a DP-path with fixed end points, grid point (m, n)
As for the state before connecting to this grid point (m, n), the left grid point (m-1, n), the diagonally lower left grid point (m-1, n-1), and the lower grid point If only the three 〓 of (m, n-1) are allowed, start from the starting point, that is, the point ○, which indicates the Tievisiev distance D ₁₁ between a ₁ and b ₁ , and select 3 directions 〓 as the path. , find the path that minimizes the sum of the distances of each grid point passed through, and enter the sum of this _{distance .} The value 1 was subtracted from the sum of the number of parameters M and the number of standard parameters N (M+N-
1) is used as the DP matching distance between the parameter time series a ₁ , ..., a _M of the input pattern and the parameter time series b ₁ , ..., b _N of the standard pattern. The initial conditions and recurrence formula showing this kind of processing are: Initial condition g(1,1)=d(1,1) Recurrence formula g(m,n)=mind(m,n)+g(m-1,
n) 2・d(m,n)+g(m-1,n-1) d(m,n)+g(m,n-1) From this, the DP matching distance D(A,
B) is expressed as D(A,B)=g(M,N)/(M+N-1).G(M,N) is divided by (M+N-1), which is the number of frames of the standard pattern N. This is to correct the difference in distance values due to the difference in . ). Through such processing, when there are L standard patterns, L DP matching distances for the input pattern are obtained, and the standard pattern having the minimum distance among the L DP matching distances is determined as the recognition result. A speech recognition device using such DP matching processing can deal with variations in speaking speed and differences in word length.
That is, it is possible to perform speech recognition that has been normalized on the time axis. However, in systems that perform speech recognition using such DP matching processing, the stationary parts of the speech are largely reflected in the DP matching distance, and it is easy to misrecognize between words that are partially similar. It became clear. That is, the acoustic parameter time series Pi(n) can be considered to draw a trajectory in the parameter space. Actually, since the parameters of each frame n correspond to one point in the parameter space, if the points are connected by curves in the time series direction, one trajectory from the starting point to the ending point can be considered. For example, when two types of words "SAN" and "HAI" are registered, the respective standard patterns A' and B' are "S", "S", and "HAI" as shown in Figure 3.
Draw a trajectory that passes through each phoneme region of "A", "N", "H", "A", and "I". When uttering "SAN" in the recognition mode, overall there are very few similarities between standard pattern B' and input pattern A, but the "A" part of "SAN" in input pattern A is is more similar to the "A" part of "HAI" of standard pattern B' than the "A" part of "SAN" of standard pattern A', and there are cases where there are many points in that part (quasi-stationary part). Here, as shown in Fig. 3, when the parameters of input pattern A are similar to the parameters of standard pattern A' as a whole, and partially similar to the parameters of standard pattern B', DP matching processing causes erroneous recognition. The case where this occurs will be explained using a one-dimensional parameter as an example. In this case, the situation shown in FIG. 3, that is, the input pattern A shown in FIG. 4 as a one-dimensional parameter time series similar to the relationship between partially similar words; 2, 4, 6, 8, 8, 8, 8, 6,
4, 4, 4, 6, 8, and the standard pattern A' as shown in Figure 5; 3, 5, 7, 9, 9, 9, 9, 7,
5, 5, 7, 9, and the standard pattern B' as shown in FIG.
Consider 4,4. As is clear from the patterns shown in FIGS. 4 to 6, input pattern A is a pattern that is desired to be determined as standard pattern A'.
However, the standard pattern for input pattern A
Calculating the DP matching distance of A' and B' shows that input pattern A is close to standard pattern B'. That is, the standard pattern for input pattern A
As shown in FIG. 7, the horizontal axis shows the parameter time series of input pattern A; 2, 4, 6, 8, 8, 8, 8, 6,
4, 4, 4, 6, 8 are lined up, and the standard pattern is on the vertical axis.
Parameter time series of A′; 3, 5, 7, 9, 9,
9, 9, 7, 5, 5, 7, and 9 are arranged, and the Thievisiev distances of the individual parameters of the standard pattern A' with respect to the individual parameters of the input pattern A are determined corresponding to the intersections on the lattice plane. and,
The starting point is the point where the Tievishev distance d (1, 1) = 1 between the first parameter 2 of the parameter time series of input parameter A and the first parameter 3 of the parameter time series of standard parameter A', and input The end point is the point at which the Tievisiev distance d (13, 12) = 1 between the 13th parameter 8 of the parameter time series of pattern A and the 12th parameter 9 of the parameter time series of standard pattern A', and DP- As in the case of Figure 2, if we allow the path to take the point to the left, the point below, and the point diagonally to the lower left of any point as the previous state for that point (this path is represented by the solid line arrow). ), and the point on the path is d(1,
1)-d(2,2)-d(3,3)-d(4,4)-d
(5,5)-d(6,6)-d(7,7)-d(8,8
)
-d(9,9)-d(10,10)-d(11,10)-d(12
，
10) - d (13, 11) - d (13, 12), the total distance of which is 24, and this DP matching distance D (A, A') is 1. On the other hand, the standard pattern for input pattern A
The DP matching process for B' is performed as shown in FIG. 8, similar to the case shown in FIG. 7 above. That is, the individual parameters of input pattern A; 2, 4, 6, 8,
Individual parameters of standard pattern B' for 8, 8, 8, 6, 4, 4, 4, 6, 8; 7, 6, 6,
Find the Tievisiev distance of 8, 8, 8, 8, 6, 4, 4, 4, and as the DP path, the previous state for any point is the point to the left of that arbitrary point, the point below, and the diagonally lower left side. (This path is indicated by a solid arrow), the points on the path are d(1,1) - d(2,2) - d(3,3) - d(4 ，
4)-d(5,5)-d(6,6)-d(7,7)-d
(8,8)-d(9,9)-d(10,10)-d(11,11
)
-d(12,11)-d(13,11), and the total distance is 15, and this DP matching distance D(A,B') is 0.65. As is clear from the result of setting this DP-path in three directions, the input pattern A is determined to be the standard pattern B' whose DP matching distance is small, and the result to be determined cannot be obtained. In this way, in the DP matching process, it is easy to misrecognize words that are partially similar. In addition, in the DP matching process, as mentioned above, the number of frames N of the standard pattern is undefined, and all standard patterns are DP matched to the input pattern.
Matching processing is required, and as the number of words increases, the amount of calculation increases dramatically, causing problems in terms of the storage capacity of the standard pattern memory 4 and the amount of calculation. For this reason, it is a speech recognition device that has relatively few erroneous recognitions even between words that are partially similar, and that requires a relatively small storage capacity of the standard pattern memory 4 and a relatively small amount of calculation for processing. The one shown in FIG. 9 has been considered. In FIG. 9, reference numeral 1 indicates a microphone as an audio signal input section.The audio signal from this microphone 1 is supplied to an amplifier 8 of an acoustic analysis section 2, and the audio signal of this amplifier 8 is cut off at a cut-off frequency of 5.5K.
Hz low-pass filter 9 to a 12-bit A/D converter 10 with a sampling frequency of 12.5 KHz, and the digital audio signal of this A/D converter 10 is
It is supplied to 15 channels of digital bandpass filter banks 11 _A , 11 _B , ..., 11 _O. This 15
The channel's digital bandpass filter banks _11A , _11B ,..., _11O are configured with, for example, Butterworth 4th order digital filters, and
Bands up to 5.5KHz are distributed at equal intervals on the logarithmic axis. Then, the output signals of the digital bandpass filters 11 _A , 11 _B , ..., 11 _O are transferred to the 15-channel rectifiers 12 _A , 12 _B , ...,
12 _O , and these rectifiers 12 _A , 12 _B ,
. . , 12 _O are supplied to 15 channels of digital low-pass filters 13 _A , 13 _B , . . . , 13 _O , respectively. These digital low-pass filters _13A , _13B ,..., _13O have a cutoff frequency of 52.8.
Consists of a Hz FIR (finite impulse response type) low-pass filter. And each digital low pass filter _13A ,
13 _B ,...,13 _O output signal sampling period
5.12ms sampler 14. This sampler 14 filters the digital low-pass filter 13.
_A , 13 _B , ..., 13 _O output signals with frame period
This sampler 1 samples every 5.12ms.
4 sampling signals are supplied to the sound source information normalizer 15. This sound source information normalizer 15 removes differences in vocal cord sound source characteristics depending on the speaker of the speech to be recognized. That is, the sampling signal Ai(n) (i=1,...,
15; n: frame number) is subjected to logarithmic transformation as follows: A′i(n)=log(Ai(n)+B) (1). In this equation (1), B
Set the bias to a value that hides the noise level. Then, the vocal cord sound source characteristics are yi=a・i+b
Approximate it with the formula: The counts of a and b are determined by the following equation.

[ka]

[ ] Then, the normalized parameters of the sound source are Pi(n)
Then, when a(n)<0, the parameter Pi(n) is expressed as Pi(n)=A′i(n)−{a(n)·i+b(n)} (4). Also, when a(n)≧0, only level normalization is performed, and the parameter Pi(n) is Pi(n)=A′i(n)− _I 〓 ⁱ⁼¹ ′A′i(n)/I( I=15) ...(5) Through such processing, the normalized parameters Pi(n) of the vocal cord sound source characteristics are supplied to the intra-speech interval parameter memory 16. The intra-speech-segment parameter memory 16 receives a speech-segment determination signal from a speech-segment determining section 17, which will be described later, and stores a normalized parameter Pi(n) of the vocal cord sound source characteristic for each speech period. On the other hand, the digital audio signal from the A/D converter 10 is supplied to a zero cross counter 18 and a power calculator 19 of the audio section determining section 17, respectively. This zero cross counter 18 counts the number of zero crosses of the digital audio signal at 64 points in that section every 5.12 ms, and uses the count value as the first
Supplied to the input end of the Also, the power calculator 19
The power of the digital audio signal in that section, that is, the sum of squares, is determined every 5.12 ms, and a power signal indicating the power within that section is supplied to the second input terminal of the speech section determiner 20. Furthermore, the sound source normalization information a(n) and b(n) of the sound source information normalizer 15 is supplied to the third input terminal of the speech segment determiner 20. Then, in the voice section determiner 20, the number of zero crossings, the power within the section, and the sound source normalization information a(n), b(n) are processed in a composite manner,
A process of determining silent, unvoiced, and voiced sounds is performed, and a voice section is determined. The voice interval determination signal indicating the voice interval of the voice interval determiner 20 is output from the voice interval determination unit 17, and the voice interval parameter memory 1
Supply to 6. NAT the normalized acoustic parameters Pi(n) of the vocal cord sound source characteristics for each voice interval stored in the voice interval parameter memory 16 in the chronological direction.
(Normalization Along Trajectory) Processing Unit 2
Supply to 1. As NAT processing, this NAT processing unit 21 estimates the trajectory in the parameter space from the acoustic parameter time series Pi(n) by linear approximation,
A new acoustic parameter time series Qi(m) is formed by linear interpolation along this trajectory. Here, this NAT processing section 21 will be further explained. Acoustic parameter time series Pi(n) (i=1,
..., I; n=1, ..., N) draws a point sequence in the parameter space. FIG. 10 shows an example of a point sequence distributed in a two-dimensional parameter space. As shown in FIG. 10, the point sequence of the non-stationary part of the voice is distributed coarsely, and the quasi-stationary part is densely distributed. This is clear from the fact that if it is completely stationary, the parameters will not change, and in that case the point sequence will remain in the parameter space. FIG. 11 shows an example in which a locus consisting of a smooth curve is estimated and drawn on a series of points as shown in FIG. If a trajectory can be estimated for a sequence of points as shown in FIG. 11, it can be considered that the trajectory remains almost unchanged with respect to variations in speech rate. because,
Differences in time length due to variations in speech rate are mostly due to temporal expansion and contraction of the quasi-stationary part (corresponds to differences in point sequence density in the quasi-stationary part in the point sequence shown in Fig. 10); This is because it is thought that the influence of the time length of the stationary part is small. The NAT processing unit 21 performs time axis normalization by focusing on the invariance of the trajectory with respect to such variations in speech rate. That is, first, a trajectory drawn as a continuous curve from the starting point Pi(1) to the ending point Pi(N) is estimated for the acoustic parameter time series Pi(n), and the curve representing this trajectory is called Pi^(s). (0
≦s≦S). In this case, Pi^(o)=Pi
(1), it is not necessary that Pi^(s) = Pi(N), and basically it is sufficient as long as Pi^(s) approximately passes through the entire point sequence. Second, the length SL of the trajectory is obtained from the estimated Pi^(s), and a new point sequence is resampled at a constant length along the trajectory as shown by the circle in FIG. For example, when sampling at M points, the trajectory is resampled using a constant length, that is, a resampling interval T=SL/(M-1). This resampled point sequence is defined as Qi(m) (i=1,...,I;m
=1,...,M), then Qi(1)=Pi^, Qi(M)=
Pi^(s). New parameter time series obtained in this way
Qi(m) has basic information on the trajectory, and is a parameter that is almost invariant to variations in speech rate. That is, the new parameter time series Qi(m) becomes a parameter time series subjected to time axis normalization. For such processing, the acoustic parameter time series Pi(n) in the intra-audio section parameter memory 16 is supplied to the trajectory length calculator 22. This trajectory length calculator 22 calculates the length of the trajectory drawn by the acoustic parameter time series Pi(n) in its parameter space by linear approximation, that is, the trajectory length. In this case, the distance between the one-dimensional vectors a _i and b _i is, for example, the Euclidean distance D (a _i , b _i ).

It is [ ]. Incidentally, as this distance, it is possible to use the Thiebishiev distance, the square distance, or the like. Therefore,
One-dimensional acoustic parameter time series Pi(n) (i=1,
…, I; n=1, …, N), the distance S(n) between adjacent parameters in the time series direction when the trajectory is estimated by linear approximation is S(n)=D(Pi(n+1), Pi (n)) (n=1,...,N
−1) …(7) Then, the first
The nth parameter from the th parameter Pi(1)
The distance SL(n) to Pi(n) is expressed as SL(n)= _o-1 〓 ⁿ ′ ⁼¹ S(n′) (8). Note that SL(1)=0. Furthermore, the trajectory length SL is expressed as SL=SL(N)= _N-1 〓 ⁿ ′ ⁼¹ S(n′) (9). The trajectory length calculator 22 uses this equation (7), (8)
The signal processing shown in equations and equations (9) is performed. A trajectory length signal indicating the trajectory length SL of the trajectory length calculator 22 is supplied to the interpolation interval calculator 23. This interpolation interval calculator 23 calculates a resampling interval T of a constant length for resampling a new point sequence by linear interpolation along the locus. In this case, if resampling is performed at M points, the resampling interval T is expressed as T=SL/(M-1) (10). The interpolation interval calculator 23 is configured to perform signal processing as shown in equation (10). A resampling interval signal indicating the resampling interval T of the interpolation interval calculator 23 is supplied to one end of the interpolation point extractor 24, and the acoustic parameter time series Pi(n) of the intra-speech interval parameter memory 16 is supplied to the interpolation point extractor 24. supply to the other end. This interpolation point extractor 24 resamples a new point sequence at a resampling interval T along a trajectory in the parameter space of the acoustic parameter time series Pi(n), for example, a trajectory obtained by linear approximation between parameters, and A new acoustic parameter time series Qi(m) is formed from the sequence. Here, the signal processing in this interpolation point extractor 24 will be explained along the flowchart shown in FIG. First, in block 24a, the value 1 is set to the variable J indicating the number in the time series direction of the resampling point, and the value 1 is set to the variable IC indicating the number in the time series direction of the acoustic parameter time series Pi(n). be done. Then, in block 24b, the variable J is incremented, and in block 24c, the number of the resampling point in the time series direction is resampled depending on whether the variable J at that time is less than or equal to (M-1). It is judged whether or not it has reached the last required number. If it has, the signal processing of this interpolation point extractor 24 is finished. If it has not reached the last number, the signal processing from the first resampling point is started in block 24d. The resampling distance DL to the Jth resampling point is calculated, and block 2
In block 4e, the variable IC is incremented, and in block 24f, the resampling distance DL is incremented from the first parameter Pi(1) of the acoustic parameter time series Pi(n).
Distance SL (IC) to ICth parameter Pi (IC)
It is determined whether the resampling point at that time is located closer to the starting end of the trajectory than the parameter Pi (IC) at that time on the trajectory, and if it is not located, block 24e is selected.
After incrementing the variable IC in block 24f, the resampling point and the parameter Pi are incremented again in block 24f.
Compare the position on the trajectory with (IC) and set the resampling point on the trajectory with the parameter Pi
When it is determined that the position is closer to the starting end than (IC), a new acoustic parameter Qi (J) along the trajectory is formed by resampling in block 24g. That is, first, from the resampling distance DL at the J-th resampling point, the distance (IC-
1) Distance SL by the th parameter Pi (IC-1)
(IC-1) is subtracted to find the distance SS from the (IC-1)th parameter Pi (IC-1) to the Jth resampling point. Next, the parameter Pi (IC-1) and the parameter Pi (IC) located on both sides of this J-th resampling point on the trajectory
The distance between S(IC-1) (this distance S(IC-1) is (7)
It is obtained by signal processing shown in the formula. ), divide this distance SS by SS/S (IC-1), and add the parameters located on both sides of the J-th resampling point on the trajectory to this division result SS/S (IC-1).
The difference between Pi(IC) and Pi(IC−1) (Pi(IC)−Pi(IC−)
1) Multiply by (Pi(IC)-Pi(IC-1))*SS/S
(IC-1), and the (IC-1)-th parameter Pi (IC-1) located adjacent to the starting end side of the J-th resampling point on the trajectory. Calculate the interpolation amount, add this interpolation amount and the (IC-1)th parameter Pi (IC-1) located adjacent to the start end side of the J-th resampling point, and add it to the trajectory. A new acoustic parameter Qi(J) is formed. 1st
Figure 4 shows two-dimensional acoustic parameter time series P(1), P
(2),...,P(8), a trajectory is estimated by linear approximation between the parameters, and along this trajectory, linear interpolation is performed to create new acoustic parameter time series Q(1),Q of 6 points.
An example of forming (2),...,Q(6) is shown below. Further, in this block 24g, signal processing for I dimensions (i=1, . . . , I) is performed in the frequency series direction. In this way, the starting and ending points of blocks 24b to 24g (these are Qi(1)=Pi^(o), Qi(M), Pi^
(s). A new acoustic parameter time series Qi(m) is formed by resampling (M-2) points excluding ). This new acoustic parameter time series Qi(m) of the NAT processing unit 21 is stored in the standard pattern memory 4 for each recognition target word in the registration mode by the mode changeover switch 3, and in the recognition mode, one end of the Tievisiev distance calculation unit 25 supply to. Further, in this recognition mode, the standard pattern stored in the standard pattern memory 4 is supplied to the other end of the Tievisiev distance calculating section 25. This Chiebishiev distance calculation unit 25 calculates the Chiebishiev distance between the input pattern consisting of the new acoustic parameter time series Qi(m) whose time axis of the currently input audio is normalized and the standard pattern in the standard pattern memory 4. Processing is done. Then, the distance signal indicating this Chiebishiev distance is supplied to the minimum distance determining unit 6, and the minimum distance determining unit 6 determines a standard pattern that minimizes the Chiebishiev distance with respect to the input pattern, and based on this determination result, it indicates the input voice. The recognition result is supplied to the output terminal 7. The operation of the speech recognition device constructed in this way will be explained. The audio signal of the microphone 1 is converted into an acoustic parameter time series Pi(n) in which the vocal cord sound source characteristics are normalized for each voice section in the acoustic analysis unit 2, and this acoustic parameter time series Pi(n) is converted to an acoustic parameter time series Pi(n) in the NAT processing unit 21. supplied to,
In this NAT processing unit 21, acoustic parameter time series
A trajectory is estimated from Pi(n) by linear approximation in the parameter space, and a new acoustic parameter time series Qi(m) is formed by linear interpolation and time axis normalization along this trajectory. This new acoustic parameter time series Qi(m) is transferred to the standard pattern memory 4 via the mode changeover switch 3.
is stored in In addition, in the recognition mode, the NAT processing unit 21
The new acoustic parameter time series Qi(m) of
5, and the standard pattern in the standard pattern memory 4 is also supplied to the Tiebisiev distance calculating section 25. Figures 15 to 17, Figures 4 to 6
Parameter time series of one-dimensional input pattern A shown in the figure; 2, 4, 6, 8, 8, 8, 8, 6, 4,
4, 4, 6, 8, parameter time series of standard pattern A'; 3, 5, 7, 9, 9, 9, 9, 7, 5,
5, 7, 9, Parameter time series of standard pattern B'; 7, 6, 6, 8, 8, 8, 8, 6, 4, 4,
Parameter time series of a one-dimensional input pattern A whose trajectory is estimated by linear approximation in the NAT processing unit 21, and the resampling points are set to 8 points;
2, 4, 6, 8, 6, 4, 6, 8, standard pattern
Parameter time series of A′; 3, 5, 7, 9, 7,
5, 7, 9, parameter time series of standard pattern B'; 7, 6, 7, 8, 7, 6, 5, 4 are shown, respectively. In this case, a trajectory in the parameter space is estimated from the acoustic parameter time series Pi(n), and a new acoustic parameter time series Qi(m) is formed along this trajectory. Time axis normalization is performed by the series Pi(n) itself. Then, the Tiebishiev distance 8 between the input pattern A and the standard pattern A' is calculated in the Tiebishiev distance calculation unit 25, and the Tiebisiev distance 16 between the input pattern A and the standard pattern B' is calculated, and these distances 8 and distance 16 are supplied to the minimum distance determination section 6, and since the distance 8 is smaller than the distance 16, the minimum distance determination section 6 determines that the standard pattern A is the input pattern.
A' is determined, and from this determination result, a recognition result indicating that the input voice is the standard pattern A is obtained at the output terminal 7. Therefore, speech recognition can be performed with relatively few erroneous recognitions even between words that are partially similar. Here, the voice recognition device that performs NAT processing and the DP
The difference in the amount of calculation with a speech recognition device that performs matching processing will be explained. When the average amount of calculations in the DP matching distance calculation unit 5 per standard pattern for the input pattern is α, the average amount of calculations in the Thiebishiev distance calculation unit 25 is β, and the average amount of calculations in the NAT processing unit 21 is γ. , the amount of calculation C ₁ due to the DP matching process for J standard patterns is C ₁ =α·J (11). Also, NAT for J standard patterns
The amount of calculation C ₂ in the case of processing is C ₂ =β·J+γ (12). Generally, the average amount of calculations α has a relationship α≫β with respect to the average amount of calculations β. Therefore, _the following relationship holds: J≫γ/α−β (13) That is, as the number of words to be recognized increases, the amount of calculation C ₁ becomes smaller than the amount of calculation C ₂ ≫
The relationship is _C2 , and by using a voice recognition device that performs NAT processing, the amount of calculation can be significantly reduced. Furthermore, since the new acoustic parameter time series Qi(m) obtained from the NAT processing unit 21 can be set to a constant number of parameters in the time series direction, the storage area of the standard pattern memory 4 can be used effectively, and its storage capacity can be reduced. It can be done relatively little. By the way, there is generally a fluctuation y at the end of a voice as shown in FIG. It is difficult to detect and remove this fluctuation y at the end of speech by speech segment determination. For this reason, conventionally, the fluctuation of the ending of this voice
It has become clear that the recognition rate decreases due to fluctuations in the endings of the speech. OBJECTS OF THE INVENTION In view of the above, an object of the present invention is to obtain a relatively high recognition rate by relatively reducing the influence of fluctuations in the endings of speech. Summary of the Invention The present invention has an audio signal input section, supplies an audio signal from the audio signal input section to an acoustic analysis section, supplies an acoustic parameter time series from the acoustic analysis section to a NAT processing section, and processes the NAT processing section. The section estimates the trajectory in the parameter space from the acoustic parameter time series, forms a new acoustic parameter time series along this trajectory, and processes this new acoustic parameter time series in the NAT processing section to generate audio. In the speech recognition device of the present invention, the NAT processing section removes a predetermined amount of the part corresponding to the ending of the speech to form a new acoustic parameter time series. Depending on the device, there is an advantage that the influence of fluctuations in the endings of speech is relatively small and a relatively high recognition rate can be obtained. Embodiment Hereinafter, an embodiment of the speech recognition apparatus of the present invention will be described with reference to FIGS. 19 and 20. In FIGS. 19 and 20, parts corresponding to those in FIGS. 1 to 18 are designated by the same reference numerals, and detailed explanation thereof will be omitted. In this example, as shown in FIG. 19, the interpolation point extractor 24 of the NAT processing unit 21 extracts the resampling interval from the acoustic parameter time series Pi(n) along a trajectory in the parameter space, for example, a trajectory obtained by linearly approximating the parameters. Resample a new point sequence at T, and remove from this new point sequence, for example two points (this is determined by the statistics of the fluctuation of the ending of the voice) that correspond to the ending of the voice, and add a new point sequence. Acoustic parameter time series Qi(m) (i=1,...,I; m=
1,...,M-2). Here, signal processing in this interpolation point extractor 24 will be explained along the flowchart shown in FIG. 19. First, in block 24h, the value 1 is set to the variable J indicating the number in the time series direction of the resampling point, and the value 1 is set to the variable IC indicating the number in the time series direction of the acoustic parameter time series Pi(n). be done. Then, in block 24i, the variable J is incremented, and in block 24j, the number of the resampling point in the time series direction is resampled depending on whether the variable J at that time is less than or equal to (M-2). It is determined whether or not it has reached the last required number. If it has, the signal processing of this interpolation point extractor 24 is finished. If it has not reached the last number, the signal processing from the first resampling point to the Jth resampling point is completed. The resampling distance DL to the th ring sampling point is calculated, and the block 2
At block 4l, the variable IC is incremented, and at block 24m, the resampling distance DL is incremented from the first parameter Pi(1) of the acoustic parameter time series Pi(n).
Distance SL (IC) to ICth parameter Pi (IC)
Based on whether the resampling point is smaller than , it is determined whether the resampling point at that time is located closer to the starting end of the trajectory than the parameter Pi (IC) at that time on the trajectory, and if it is not located, block 24l is
After incrementing the variable IC at block 24m, the resampling point and parameter Pi are again set at block 24m.
Compare the position on the trajectory with (IC) and set the resampling point on the trajectory with the parameter Pi
(IC), a new acoustic parameter Qi (J) along the trajectory is formed by resampling in block 24m. That is, first, from the resampling distance DL at the J-th resampling point, the distance (IC-
1) Distance SL by the th parameter Pi (IC-1)
(IC-1) is subtracted to find the distance SS from the (IC-1)th parameter Pi (IC-1) to the Jth resampling point. Next, the parameter Pi (IC-1) and the parameter Pi (IC) located on both sides of this J-th resampling point on the trajectory
The distance between S(IC-1) (this distance S(IC-1) is (7)
It is obtained by signal processing shown in the formula. ), divide this distance SS by SS/S (IC-1), and add the parameters located on both sides of the J-th resampling point on the trajectory to this division result SS/S (IC-1).
The difference between Pi(IC) and Pi(IC−1) (Pi(IC)−Pi(IC−)
1) Multiply by (Pi(IC)-Pi(IC-1))*SS/S
(IC-1), and the (IC-1)-th parameter Pi (IC-1) located adjacent to the starting end side of the J-th resampling point on the trajectory. Calculate the interpolation amount, add this interpolation amount and the (IC-1)th parameter Pi (IC-1) located adjacent to the start end side of the J-th resampling point, and add it to the trajectory. A new acoustic parameter Qi (J) is formed, and the starting point (this is Qi (1)
= Pi∧(o), which can be obtained by substitution without requiring calculation processing. ) and a predetermined number, for example, two new acoustic parameters Qi (M-1) corresponding to the endings of the speech.
By the calculation process excluding Qi (M), (M-2)
The new acoustic parameter time series Qi(m) (i=1,
..., I; m=1, ..., M-2) is formed.
In this block 24n, I-dimensional signal processing is performed in the frequency series direction. A new acoustic parameter time series excluding the parameter corresponding to the ending of the voice of this NAT processing unit 21
Qi(m) (i=1,...,I; m=1,...,M-2)
is supplied to the mode changeover switch 3. In addition, the acoustic analysis section 2, standard pattern memory 4, Tiebishiev distance calculation section 25, minimum distance determination section 6, etc. are the above-mentioned 9th
The configuration is similar to the speech recognition device shown in the figure. The operation of the speech recognition device constructed in this way will be explained. The audio signal of the microphone 1 is converted into an acoustic parameter time series Pi(n) in which the vocal cord sound source characteristics are normalized for each voice section in the acoustic analysis unit 2, and this acoustic parameter time series Pi(n) is converted to an acoustic parameter time series Pi(n) in the NAT processing unit 21. supplied to,
In this NAT processing unit 21, acoustic parameter time series
A trajectory is estimated from Pi(n) by linear approximation in the parameter space, and along this trajectory, the ending of the speech is determined by the corresponding acoustic parameters Qi (M-1) and Qi
New acoustic parameter time series Qi(m) excluding (M)
(i=1,...,I; m=1,...,M-2) is formed. Fluctuations at the end of a voice appear as a relatively short trajectory length backwards on the trajectory, and this trajectory length has extremely small fluctuations due to differences in utterances. ), it is possible to remove only the fluctuations at the end of the speech. Therefore, a new acoustic parameter time series Qi(m) (i=1,
. . , I; m=1, . FIG. 20 shows such a state in two dimensions. In this Figure 20, the 11 points indicate the acoustic parameter time series P(1),...,P(11), the solid line between these marks indicates the trajectory based on linear approximation, and the 9 points indicate the trajectory based on linear approximation. New acoustic parameter time series Q(1),...,Q(9)
shows. As is clear from Fig. 20, the acoustic parameters P(9), P(10), P(11) located at the rear of the trajectory
takes a relatively short trajectory length on the trajectory, and this part represents the fluctuation of the ending of the speech, and the new acoustic parameter time series Q(1),... except for the new acoustic parameter Q(9) corresponding to this part. , Q(8) is relatively unaffected by fluctuations in the endings of speech. New acoustic parameter time series Qi(m) (i=1,...,I; m=1,...,M
-2) is stored as a standard pattern in the standard pattern memory 4 via the mode changeover switch 3 in the registration mode. In addition, in the recognition mode, the NAT processing unit 21
A new acoustic parameter time series Qi(m) (i=1,
..., I; m=1,...,M-2) is supplied as an input pattern to the Tiebishiev distance calculation section 25 via the mode changeover switch 3, and the standard pattern in the standard pattern memory 4 is supplied to the Tiebishiev distance calculation section 25. Then, the Tiebyshiev distance between the input pattern and the standard pattern is calculated, and the minimum distance determination unit 6 determines the standard pattern with the minimum Tiebyshiev distance, and this standard pattern is sent to the output terminal 7 as a recognition result indicating the input voice. can get.
In this case, the standard pattern and the input pattern are new acoustic parameter time series Qi(m) (i=1,...,I; m=
1, . As described above, according to this example, the microphone 1 is provided as an audio signal input section, the audio signal from this audio signal input section 1 is supplied to the acoustic analysis section 2, and the acoustic parameter time series Pi of this acoustic analysis section is provided. (n) to the NAT processing unit 21, the NAT processing unit 21 estimates a trajectory in the parameter space from the acoustic parameter time series Pi(n), and creates a new acoustic parameter time series Qi(n) along this trajectory. m) (i=1,...,I;
m=1,...,M), and the NAT processing section 21 processes this new acoustic parameter time series Qi(m) to recognize speech. Two new acoustic parameters Qi (M−
1) and Qi(M) to create a new acoustic parameter time series Qi(m)(i=1,...,I; m=1,...,
M-2), the effect of fluctuation on the ending of the speech is relatively small, and there is an advantage that a relatively high recognition rate can be obtained. In addition, a new acoustic parameter time series Qi(m) (i=1,...,I; m=
1, . . . , M-2) can be reduced by the number of parameters corresponding to the endings of the speech, so there is an advantage that the storage capacity of the standard pattern memory 4 can be reduced by that amount. In the above embodiment, a new acoustic parameter time series Qi(m) (i=1,...,I; m=1,
..., M-2), however, it is also possible to form a new acoustic parameter time series by removing an appropriate number of resampling points, such as one, according to the statistics of the fluctuations in the endings of speech. It is easy to understand that the same effects as in the example can be obtained. In addition, in the above embodiment, a trajectory in the parameter space is estimated from the acoustic parameter time series Pi(n), a new point sequence is resampled along this trajectory, and the ending of the speech in this new point sequence is Although we have described the case where a new acoustic parameter time series is formed from a point sequence excluding the point sequence corresponding to , the trajectory length obtained from the statistics of the part corresponding to the fluctuation in the ending of the voice on this trajectory is removed from the terminal side of the trajectory, and a new sound is generated along the trajectory excluding the trajectory length corresponding to the fluctuation in the ending of the speech. It is easy to understand that the same effects as in the above embodiment can be obtained even if a parameter time series is formed. Furthermore, in the above embodiment, the length of the trajectory is calculated from the acoustic parameter time series by linear approximation in the parameter space. However, it is easy to understand that the same effects as in the above embodiment can be obtained.
Furthermore, it goes without saying that the present invention is not limited to the above-described embodiments, and can take various other configurations without departing from the gist of the present invention. Effects of the Invention According to the speech recognition device of the present invention, the speech recognition device has a speech signal input section, supplies the speech signal of the speech signal input section to the acoustic analysis section, and transmits the acoustic parameter time series of the acoustic analysis section to the NAT processing section. This NAT processing unit estimates a trajectory in the parameter space from the acoustic parameter time series, forms a new acoustic parameter time series along this trajectory, and uses this new acoustic parameter time series in the NAT processing unit. In a speech recognition device that recognizes speech by processing the speech, the NAT processing unit removes a predetermined amount of the part corresponding to the end of the speech to form a new acoustic parameter time series. This method has the advantage of being relatively unaffected by fluctuations in the endings of words and achieving a relatively high recognition rate.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an example of a speech recognition device that performs speech recognition using DP matching processing, Fig. 2 is a conceptual diagram explaining the DP matching processing, and Fig. 3 is an explanation of a trajectory in acoustic parameter space. Diagrams provided for, Figures 4, 5 and 6
The figures are one-dimensional input pattern A and standard pattern, respectively.
A diagram showing examples of A' and standard pattern B', and Figure 7 is a diagram used to explain time axis normalization by DP matching processing between the parameter time series of input pattern A and the parameter time series of standard pattern A'. , 8th
The figure is a diagram to explain time axis normalization by DP matching processing between the parameter time series of input pattern A and the parameter time series of standard pattern B'. A configuration diagram showing an example of a speech recognition device, FIG.
Figures 11, 12, and 14 are diagrams explaining the NAT processing unit, Figure 13 is a flowchart explaining the interpolation point extractor, and Figures 15, 16, and 17 are diagrams explaining the NAT processing unit, respectively. A diagram showing the one-dimensional acoustic parameter time series of input pattern A, standard pattern A', and standard pattern B' processed by the NAT processing unit. Figure 18 is a diagram showing an example of audio waveform. Figure 19. 20 is a flowchart showing an embodiment of the main part of the speech recognition device of the present invention, and FIG. 20 is a diagram for explaining FIG. 19. 1 is a microphone as an audio signal input section; 2
is an acoustic analysis section, 3 is a mode changeover switch, 4 is a standard pattern memory, 6 is a minimum distance judgment section, _11A ,
11 _B ,..., 11 _O are digital band pass filter banks of 15 channels, 16 is a voice interval parameter memory, 21 is a NAT processing unit, 22 is a trajectory length calculator, 23 is an interpolation interval calculator, and 24 is an interpolation point extraction 25 is a Tievishiev distance calculation unit, 26 is a trajectory length signal adder, 27 is a standard pattern selection unit, y
is the fluctuation of the ending of the voice.

Claims (1)

[Claims] 1 has an audio signal input section, supplies the audio signal of the audio signal input section to an acoustic analysis section, supplies the acoustic parameter time series of the acoustic analysis section to a NAT processing section,
The NAT processing unit estimates a trajectory in the parameter space from the acoustic parameter time series, forms a new acoustic parameter time series along the trajectory, and processes the new acoustic parameter time series in the NAT processing unit. The speech recognition device is characterized in that the NAT processing unit removes a predetermined amount of a portion corresponding to the ending of the speech to form a new acoustic parameter time series. speech recognition device.