JPS60249199A - Voice recognition equipment - Google Patents

Abstract

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

Description

[Detailed description of the invention] Industrial applications 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, a device that performs P matching processing has been proposed, for example, as disclosed in Japanese Patent Application Laid-Open No. 50-96104.

First, a speech recognition device that performs speech recognition using this DPP matching process will be described.

In FIG. 1, (1) indicates a microphone as an audio signal input section, and an audio signal from this microphone (1) is supplied to an acoustic analysis section (2).
), the acoustic parameter time series Pi(nl) is obtained.

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.
(nl (i = l,...+Ii I is the number of channels of the bandpass filter bank, n=1...,
NUN is the number of frames cut out by voice section determination. ) is obtained as

The acoustic parameter time series Pi(n
) 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 DPP 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 needle section 5).

This DP matching distance needle section 5) performs DPP matching distance calculation processing between the input cover turn consisting of the acoustic parameter time series P i (n) of the voice input at that time and the standard pattern of the standard pattern memo January 4). is performed, and a distance signal indicating the DPP matching distance of the DP matching distance meter group unit 5) is supplied to the minimum distance determining unit (6),
In this minimum distance determination unit (6), D
A standard pattern with the minimum PP matching separation is determined, and from this determination result, 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 DPP matching process performs time axis normalization to deal with variations in speaking speed and differences in word length.

This DPP matching process will be explained below. Here, for simplicity, the dimension corresponding to the frequency axis direction i of the acoustic parameter time series P i (nl is omitted, and the parameter time series of the standard pattern is expressed as bl, . . . ).

bN, parameter time series of input pattern is a1+...
・As + aM, the DPP matching process in the case of two DP-passes with fixed end points will be explained.

Figure 2 shows a conceptual diagram of the DPP matching process, in which input parameters (M = 19) are arranged on the horizontal axis, and standard parameters (N = 12) are arranged on the vertical axis.
, N) There are MXN points on the grid plane, and one distance corresponds to each point. For example, the distance between a3 and b5 corresponds to the intersection of a straight line extending vertically from a3 and a straight line extending horizontally from b5. In this case, if we take the Chebyshev distance as the distance, a3 and b5
The distance from this point is Da5. As a DP-bus with fixed end points, the state before connecting to the grid point (m, n) is the grid point (m -1, n) on the left side, the diagonally lower left side The grid point (m-1, n-1) and the lower grid point (m
, n-1), the starting point, that is, al
Chebyshev distance lit D between and bl Starting from the point ■ indicating 1'1, calculate the distances of points located in the three directions allowed as a path (route) -↑/, and take the path in the minimum direction among the three distances. In the same way, the bus is advanced one after another until it reaches the end point, that is, the point ◎ indicating the Chebyshev distance D 1912 between al9 and bl2. ) is the input pattern parameter time series a1+..., aM and the standard pattern parameter time series b1.・・・＋ bN
DPP matching separation is performed. Therefore, if the number of standard patterns is L, the DP for the input cover turn is
L P matching distances are determined, and the standard pattern with the minimum distance among the L DP matching distances is determined as the recognition result.

As described above, the speech recognition device using P matching processing can deal with variations in speaking speed and differences in word length, that is, perform speech recognition with time axis normalization.

However, when speech recognition is performed using P matching processing like this, the stationary parts of the speech are largely reflected in the DPP matching distance, and it is easy to misrecognize between words that are partially similar. It became clear that

In other words, the acoustic parameter time series Pi (nl) can be thought of as drawing a trajectory in its parameter space.Actually, since the parameter of each frame n corresponds to one point in the parameter space, although it is a point sequence, it can be thought of as drawing a trajectory in the parameter space. By connecting curves in the series direction, one trajectory from the starting point to the ending point can be considered.

For example, when two types of words "IAI" and "SAI" are registered, the respective standard patterns A' and B' are "H", "A", "■", and "S" as shown in Figure 3. Draw a trajectory that passes through each phonological region. When uttering "SAI" in recognition mode, the part "A→■" of "HAT" in standard pattern B' is closer to the part "A→■" of "SAI" in input pattern B' than the part "A→■" of "SAI" in input pattern AA'. In the case of specific speaker recognition, this can occur if the sound quality changes slightly between the time of registration and the time of recognition.

Here, as shown in FIG. 3, the input cover pattern A generally follows the standard pattern A', and partially follows the standard pattern B'.
A case will be explained using a one-dimensional parameter as an example where DP matching processing causes erroneous recognition when drawing a trajectory along . In this case, the situation shown in Figure 3, that is, the fourth
Inlet cover turn A as shown in the figure, 2.4.6.8.8
．． 8.8.6°4, 4.4.6.8 and standard pattern A' as shown in FIG. 5; 3.5.7.9.9.9. Hunger? , 5.5.7.9 and standard pattern B' as shown in Figure 6; 7.6.6.8°8, 8.8.6.4.4
．． 4. As is clear from the patterns in FIGS. 4 to 6 (the input cover turn A is a pattern that is desired to be determined as the standard pattern A'), the DP matching distance of the standard patterns A' and B' with respect to the input cover turn A is When calculated, input cover pattern A is standard pattern B.
′ is shown to be close to .

That is, the DP of the standard pattern A' for the input cover pattern A
Similar to FIG. 2, the matching process is performed as shown in FIG.
8.8.8.8.6.4.4.4.6.8 are arranged, and the vertical axis is the parameter time series 3.5.7.9 of standard pattern A'.
．． 9.9.9.7.5.5.7.9 are arranged and find the Chebyshev distance of each parameter of standard pattern A' with respect to each parameter of input cover pattern A corresponding to the intersection in the grid plane. Ru.

Then, starting from the point where the Chebyshev distance D11 = 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', the input cover turn A is The 13th parameter 8 in the parameter time series and the 12th parameter 9 in the parameter time series of standard pattern A'
Chebyshev distance 111Dttt2= 1) point woH
Point Toshi, DP-Pass as the second [i! ! As in the case of I, when the previous state for any point is allowed to take the point to the left, the point below, and the point diagonally to the lower left of the arbitrary point (this path is shown by a solid arrow). ), the points on the path are D 1l-D22 033 D44 D66 Des
Dl-Des D99D 1o 1o D 11ro
There are 14 points, D 1210-D 1311D 1312, and the sum of their distances is 14, and (sum of distances/number of points) is l. Also, in this FIG. 7, if it is allowed to take a point to the left of an arbitrary point and a point below it as the previous state for an arbitrary point as a DP-path (this path is indicated by a broken line arrow). , the points on the path are D xl−D 2
1- D 22 D 32 D 33- D 43D5
3 DG3 D73-DP4 076 Dvs D77
-Doll Dss Dss-Dss Dtos D1
29 Dlxl.

There are 24 points of D121'1-D131 Sue D 1a x2, the total distance is 24, and (total distance/score) is 1.

On the other hand, DP of standard pattern B' for input cover pattern A
The matching process is performed as shown in FIG. 8, similar to the case shown in FIG. 7 above. That is, the individual parameters of input cover pattern A 2.4.6.8.8.8.8°6, 4.4.4.
Individual parameters 7 of standard pattern B' for 6.8
, 6.6.8.8.8.8.6.4.4 Determine the Chebyshev distance of 4, and as the previous state for any point as a DP-path, the point on the left of that arbitrary point, the lower side If you allow points and points on the diagonal lower left side (solid arrow), the points on the path are DllD 22 D 33 D 44 D 6
6-Dss-D? ? -D 5s-D 99
D so Io D 1111D 121'I D 1
There are 13 points of 3x'x, and the total distance is 13, and (total distance/number of points) is 1. In addition, in this Figure 8, if it is allowed to take a point to the left of any point and a point below it as the previous state for any point as a DP-path (dashed line arrow), the point on the path will be D. 1'l D
21-D 31- D 32- D 33 D 43-
D 44 D 64-D 64-D? 4 D vs.
D? GD? T -D at -D a@-D
so D 5sDs 5o-Ds 11-D+osID
! '114 D121'I D131'1 There are 23 points, the total distance is 21, and (total distance/number of points) is 0. !
1) It is 13.

As is clear from this result, when the DP-path is 2 directions - 1 and its (sum of distances/scores) is the DP matching distance, the input cover turn A is determined to be the standard pattern B', and 2
(Total distance/number of points) in direction -T is the DP matching distance, or (Total sum of distance/number of points) in 3-direction force
If the DP matching distance is taken as the DP matching distance, the input cover pattern A will be equidistant 1 from the standard patterns A' and B', and the result that should be determined will not 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 moreover, it is necessary to perform the DP matching process on all the standard patterns for the input cover turn, and as the number of words increases, the amount of calculation increases accordingly. This has caused problems in terms of storage capacity and amount of calculation for standard pattern memos (January 4).

Purpose of the Invention In view of the above, the present invention has relatively few erroneous recognitions even between words that are partially similar, and the storage capacity of the standard pattern memory and the amount of calculation are relatively small. The purpose is to obtain a speech recognition device.

Summary of the Invention The present invention has an audio signal input section, converts an audio signal from the audio signal input section into an acoustic parameter series, estimates a trajectory in the parameter space from this acoustic parameter series, and estimates a trajectory in the parameter space based on this trajectory. The speech recognition device of the present invention is designed to recognize speech signals, and with the speech recognition device of the present invention, misrecognition is relatively rare even between partially similar words, and it is possible to recognize speech signals that are similar to the standard. There is an advantage that the storage capacity of the pattern memory and the amount of calculation can be relatively small.

Embodiment Hereinafter, an embodiment of the speech recognition apparatus of the present invention will be described with reference to FIGS. 9 to 17. In FIGS. 9 to 17, parts corresponding to those in FIGS. 1 to 8 are designated by the same reference numerals, and detailed explanation thereof will be omitted.

In FIG. 9, (11 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 Signal cutoff frequency 5.5KHz
A 12-bit A/D converter (+01) with a sampling frequency of 12.5 KHz is supplied through a low-pass filter (9) of
IIA), (IIB), old, (llo).

This 15-channel digital bandpass filter bank (IIA), (IIB), ..., (llo)
is constituted by, for example, a Butterworth fourth-order digital filter, and the bands from 250 Hz to 5.5 KHz are distributed at equal intervals on the logarithmic axis. And each digital bandpass filter (IIA), (liB)
.

..., (llo) through a 15-channel rectifier (12A), (12B), ..., (1
2o), and these rectifiers (12^) and (1
2B) ,...

The square output of (12o) is passed through a 15-channel digital low-pass filter (13A), (13s),...
, (13o), respectively. These digital low-pass filters (138), (13s), ..., (
13o) is constituted by an FIR (finite impulse response type) low-pass filter with a cutoff frequency of 52.8H2.

and each digital low-pass filter (13A).

The output signals of (13s), . . . , (13o) are supplied to a sampler (14) with a sampling period of 5.12 m5.

This sampler (14) allows digital low-pass filters (13A), (13th), ..., (13o)
The output signal of the sampler (14) is sampled at every frame period of 5.12 m5, and the sampling signal of this sampler (14) is supplied to the sound source information normalizer (15). This sound source information normalizer (15) removes differences in individual characteristics between speakers of speech to be recognized.

That is, the sampling signal At(nl (i = 1+...) supplied from the sampler (14) every frame period
....

15;n: frame number) -Ai(n)=log(At(n)+B)...(
1) A logarithmic transformation is performed. In this (11 formula,
B is set to a value such that the noise level is hidden by the bias. Then, the vocal cord sound source characteristics are approximated by the formula yi =a - i + b. The counts of a and b are determined by the following equation.

(N=15) ... (2) (N = 15) ... (3) Then, if the normalized parameter of the sound source is Pi (nl), then a (
When nl < 0, the parameter Pi(n) is Pi(n)
−＾1(n) −(a[n)・i +b(nl) −−
−+41.

Also, when a (n)≧0, only level normalization is performed,
The parameter pttn) is expressed as... (5).

Through this type of processing, the normalized parameters P i (nl) of the sound source characteristics are stored in the voice interval parameter memory (1
6) Supply to. This voice interval parameter memory (
16) receives a speech section determination signal from a speech section determination section (17), which will be described later, and stores a normalized parameter Pi(n) of the sound source characteristic for each speech section.

On the other hand, the digital audio signal from the A/D converter Ql 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 m5, and supplies the count value to the first input terminal of the audio section determiner (20). In addition, the power calculator (19) calculates the power of the digital audio signal in that section every 5.12 m5, that is, the sum of squares,
The power signal indicating the power within the interval is sent to the voice interval judger (
20). Furthermore, the sound source normalization information a (nl and b (+) of the sound source information normalizer (15)
1) 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 (nl,
b (processes nl in a complex manner, performs processing to determine silent, unvoiced, and voiced sounds, and determines a speech section. A speech section determination signal indicating the speech section of this speech section determiner (20) is sent to the speech section determination section. The output of (17) is supplied to the voice section parameter memory (16).

Normalized acoustic parameters P of sound source characteristics for each voice section stored in this intra-voice section parameter memory (16)
Hn) is supplied to the NAT processing unit (21) in the chronological direction. This NAT processing unit (21) estimates a trajectory in the parameter space from the acoustic parameter time series P l (nl) as a NAT process, and forms a new acoustic parameter time series Qi (b) based on this trajectory.

Here, this NAT processing section (21) will be further explained. Acoustic parameter time series Pi(nl (i −1,・
...+Ii n = L..., 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 variable space.

FIG. 11 shows an example in which a locus is drawn as a smooth curve on the point sequence as shown in FIG. As shown in Fig. 11 (if a trajectory can be estimated for a sequence of points, it can be considered that the trajectory remains almost unchanged even when the speech rate changes. Most of the differences are due to the temporal expansion and contraction of the quasi-constant part (corresponding to the difference in the density of the quasi-stationary part in the point sequence shown in Fig. 10), and the influence of the time length of the unsteady part is This is because it is considered to be 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 start point Pi(1) to the end point Pi(g) for nl is estimated, and the curve representing this trajectory is expressed as Pi(sl()).
It is sufficient if it approximately passes through the entire 0≦S point sequence.

Determine the trajectory length SL from the second estimated Pits),
As shown by circles in FIG. 12, a new point sequence is resampled at a constant length along the trajectory. For example, when sampling at M points, the resampling interval T=S
The trajectory is resampled using L/(M-1) as a reference. The resampled point sequence of each is converted into the old fml (i
=L-,I; m =1. =・,M), then []
H1l-Pi(01, Qi(M) = Pi(S).

The new parameter time series Qi obtained in this way)
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 entry itml becomes a parameter time series subjected to time axis normalization.

For this kind of processing, parameter memory (
The acoustic parameter time series Pi (nl) of 16) is supplied to the trajectory length calculator (22). In other words, the Euclidean distance 1itlD(ai+bi) between the ■dimensional vectors a1 and bi is... (6).Then, the ■dimensional acoustic parameter time series Pi(
n) (i = 1.-=, I; n = 1.-, N), when the trajectory is estimated by linear approximation, the distance S fn) between adjacent parameters in the time series direction is 5(n) = D
(Pi (nlx), Pi(nl) (fi=1.
..., N)... (7) It is expressed as. Then, from the first parameter Pifl) to the nth parameter Pif in the time series direction
The distance 5L(n) to nl is expressed as. Note that 5L(1) = O. Furthermore,
The trajectory length SL is expressed as. The trajectory length calculator (22) uses this equation (7),
The signal processing shown in equations (8) and (9) is performed.

A trajectory length signal indicating the trajectory length SL of this trajectory length calculator (22) is supplied to an interpolation interval calculator (23). This interpolation interval calculator (23) resamples a new point sequence by linear interpolation along the trajectory at a fixed length sampling interval T.
is calculated. In this case, if resampling is performed at M points, the resampling interval T is expressed as T=SL/(M-1) .

The interpolation interval calculator (23) is configured to perform signal processing as shown by this αΦ formula.

This interpolation interval calculator (23) supplies a resampling interval signal indicating the sampling interval T to one end of the interpolation point extractor (24), and also supplies the voice interval parameter memory (
16) is supplied to the other end of the interpolation point extractor (24).
) resamples a new point sequence at a resampling interval T along the trajectory of the acoustic parameter time series Pi(n) in its parameter space, for example, a trajectory that is a linear approximation between the parameters, and from this new point sequence, a new acoustic It forms the parameter time series Qi(-).

Here, the signal processing in this interpolation point extractor (24) will be explained along the flowchart shown in FIG. First, in block (24a), a value 1 is set to a variable J indicating the number of the resampling point in the chronological direction, and a variable IC indicating the number of the acoustic parameter time series Pl(n) in the chronological direction is set. is set to the value 1. And block (
24b), the variable J is incremented, and the block (
In 24c), depending on whether the variable 1 at that time is less than or equal to (M-1), it is determined whether the number of the sampling point in the time series direction is the last number that needs to be sampled. If it is not, the glue sample distance pL from the first glue sampling point to the fifth glue sampling point is calculated in block (24d), and the variable IC is calculated in block (24e).
is incremented, and in block (24f), the resampling distance 111 [IL is the acoustic parameter time series Pi (
Depending on whether the distance from the first parameter Pi(1) of nl to the first Cth parameter PiQc) is smaller than the distance 5LOc), the sampling point at that time is smaller than the parameter P1(Ic) at that time on the trajectory. If it is not located, the variable IC is incremented in block (24e), and then the trajectory between the resampling point and the parameter Pi (Ic) is determined in block (24f) again. When the above positions are compared and it is determined that the resampling point is located closer to the starting point than the parameter Pi (IC) on the trajectory, a new acoustic nomameter Qi ( J) is formed. That is, first, the distance 5LOc-1) due to the (IC-1)th parameter PiθC-1) located closer to the starting point than the third glue sampling point is subtracted from the resampling distance 111DL based on the third glue sampling point. and the (IC-1)th parameter Pi(Ic-1
) to the third glue sampling point 1IIISS
I put it on. Next, the parameter P i (Ic-1
) and the distance 5Trll between the parameter P i (Ic )
(Divide this distance aSS by this distance S (nl is obtained by signal processing shown in equation (7)) ss/S
(IC-1) L, this division result SS/S (IG-
1), the difference (PiQc) Pi(+c-t)) between the parameter PiQc) located on both sides of the third glue sampling point on the trajectory and PiθC-1 is (subtracted (Pi
Gc) Pi(Ic-1)) * SS/ S (+c
-t) Then, the (IC-1)th parameter Pi(IC-1
), and calculate the interpolation amount from this interpolation amount and the third (IC) located adjacent to the starting point side from the third
-1)th parameter PiOc-1),
A new acoustic parameter Qi(J) along the trajectory is formed. Figure 14 shows the two-dimensional acoustic parameter time series Pt1)
, PI2).

..., P(81, a trajectory is estimated by linear approximation between the parameters, and along this trajectory, new acoustic parameter time series Q(1), Q(21°・
..., Q(61) is shown. Also, in this block (24g), one dimension (
i=1. ..., I) signal processing is performed.

In this way, the start and end points (these are old f1>=PilO) in blocks (24b) to (24g),
gi(M) = Pi(s). ) except < (M-
2) A new acoustic parameter time series Qifm) is formed by point sampling.

In the registration mode, the new acoustic parameter time series old (m) of the NAT processing unit (21) is stored in the standard pattern memory (
4), and in the recognition mode, it is supplied to one end of the Chebyshev distance calculation unit (25). Also, in this vgm mode, the standard pattern stored in the standard pattern memory (4) is supplied to the other end of the Chebyshev distance calculating section (25). In this Chebyshev distance calculation unit (25), a new acoustic parameter time series Qi which is normalized on the time axis of the audio input at that time is used. Chebyshev distance calculation processing is performed.

Then, the distance signal indicating this Chebyshev distance is supplied to the minimum distance determination section (6), and the minimum distance determination section (6)
A standard pattern with the minimum Chebyshev distance to the input cover pattern is determined, and from this determination result, a recognition result indicating the input voice 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 sound source characteristics are normalized for each voice section in the acoustic analysis unit (2), and this acoustic parameter time series PiTn) is converted to an acoustic parameter time series Pi(n) in which the sound source characteristics are normalized in the acoustic analysis unit (2). 21) is supplied to 1. In this NAT processing unit (21), the acoustic parameter time series Pi (
A trajectory is estimated by linear approximation in the parameter space from The acoustic parameter time series Qi) are stored in the standard pattern memo (January 4) via the mode changeover switch (3).

In the recognition mode, the new acoustic parameter time series old [ff1] of the NAT processing unit (21) is transmitted to the Chebyshev distance calculator (25) via the mode changeover switch (3).
and the standard pattern of the standard pattern memo (January 4) is supplied to the Chebyshev distance calculator (25). Parameter time series for one-dimensional input cover turns shown in FIGS. 4 to 6 in FIGS. 15 to 17; 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. Parameter time series of one-dimensional input cover 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, Parameter time series i of standard pattern A' 3.5.7.9.7.5.7.9, Parameter time series i of standard pattern B' 7.6. 7.8°7 and 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 O1) is formed along this trajectory, so the acoustic parameter time series obtained by converting the input speech Time axis normalization is performed by Pl(n+ itself.The Chebyshev distance calculator (25) then calculates the Chebyshev distance 8 between the input cover pattern A and the standard pattern A', and also calculates the Chebyshev distance 8 between the input cover pattern A and the standard pattern B. The Chebyshev distance 1i1116 between the Since it is smaller than , it is determined that the standard pattern A is the input pattern A',
From this determination result, a recognition result indicating that the input voice is 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.

As described above, the speech recognition device of the present invention has the microphone (11) as an audio signal input section, and converts the audio signal from the audio signal input section (11) into an acoustic parameter time series Pi(n). , this acoustic parameter time series Pi(
Since the trajectory is estimated by linear approximation in the parameter space from n) and the speech signal is recognized based on this trajectory, there is no possibility of erroneous recognition even between partially similar words. It has the advantage of being able to recognize relatively few voices. Further, there is an advantage that the amount of calculation for processing can be relatively reduced.

Here, the voice recognition device of the present invention that performs NAT processing and the DP
The difference in the amount of calculation compared to a conventional speech recognition device that performs matching processing will be explained.

Let α be the average amount of calculation in the DP matching distance calculation unit (5) per standard pattern for the input cover turn, let β be the average amount of calculation in the Chebyshev distance calculation unit (25), and calculate the average amount of calculation in the NAT processing unit (21). When the amount is γ, the calculation amount c1 by DP matching processing for 5 standard patterns is (, 1 = α ・ J ・ ・ ・ (11). Also, when NAT processing is performed for 5 standard patterns, The calculation amount C2 is C2-β·J+γ (12) Generally, the average calculation amount α has the relationship α)β with respect to the average calculation amount β. Therefore, J)□ ... (1
3) - The relationship α-β holds true, that is, as the number of words to be recognized increases, the amount of calculation C1 becomes the relationship C1>>C2 with respect to the amount of calculation C2, and this depends on the speech recognition device of the present invention that performs NAT processing. This has the advantage of significantly reducing the amount of calculation.

In addition, the new acoustic parameter time series Qi (2)) obtained from the NAT processing unit (21) can be set to a constant number of parameters in the time series direction, so the storage area of the standard pattern memo (January 4) can be effectively used. It has the advantage that its storage capacity can be relatively small.

Note that in the above embodiment, the acoustic parameter time series Pl[
Although we have described the case where the trajectory in the parameter space is estimated from nl by linear approximation and a new acoustic parameter time series Qi (e) is formed from this trajectory by linear interpolation, arc approximation, spline approximation It is also possible to obtain the same effects as in the above-mentioned embodiments by estimating the trajectory by estimating the trajectory and also by forming a new acoustic parameter time series i ((b) from the trajectory by circular interpolation, spline interpolation, etc. It is easy to understand.Also, in the above embodiment, the trajectory in the parameter space is estimated from the acoustic parameter time series Pi (nl), and a new acoustic parameter time series Q i hl is formed based on this trajectory. As described above, the frequency characteristics of the audio signal can be normalized by estimating the trajectory in the parameter space from the acoustic parameter frequency series and forming a new acoustic parameter frequency series based on this trajectory. Furthermore, in the above embodiment, a new acoustic parameter time series O1 (E) is formed by estimating a trajectory extending from the acoustic parameter time series Pi(nl) into the parameter space and resampling along this trajectory. Although the case has been described above, it is also possible to perform speech recognition by extracting the coefficients indicating the shape of the trajectory and using these coefficients as a new acoustic parameter time series.Furthermore, the present invention does not go beyond the above-mentioned embodiments. Of course, various other configurations can be adopted without departing from the above.

Effects of the Invention According to the speech recognition device of the present invention, it has a speech signal input section, converts the speech signal from the speech signal input section into an acoustic parameter series, and estimates a trajectory in the parameter space from this acoustic parameter series. Since the speech signal is recognized based on this trajectory, there are relatively few misrecognitions even between partially similar words, and the storage capacity of the standard pattern memory There is an advantage that a relatively small amount of calculation can be obtained.

[Brief explanation of drawings]

Fig. 1 is a configuration diagram showing an example of a speech recognition device that performs speech recognition by DP matching processing, and Fig.
A conceptual diagram for explaining the matching process, FIG. 3 is a diagram for explaining the locus in the acoustic parameter space, and FIGS. 4, 5, and 6 are one-dimensional input pattern A, standard pattern A', and A diagram showing an example of standard pattern B',
Fig. 7 is a diagram for explaining time axis normalization by DP matching processing between the parameter time series of input cover turn A and the parameter time series of standard pattern A', and Fig. 8 is a diagram showing the parameter time series of input cover turn A and the standard pattern A'. A diagram for explaining time axis normalization by DP matching processing with the parameter time series of pattern B', FIG. 9 is a block diagram showing an embodiment of the speech recognition device of the present invention, FIGS. 10, 11, FIG. 12 and FIG. 14 are diagrams for explaining the NAT processing section, respectively;
Figure 13 is a flowchart for explaining the interpolation point extractor;
16 and 17 are diagrams showing one-dimensional acoustic parameter time series of input cover turn A, standard pattern A', and standard pattern B' processed by the NAT processing section, respectively. (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 judger, and 11
8) (lla), ..., (Ilo) is a 15-channel digital bandpass filter bank, (16) c) voice interval parameter memory, (21) is a NAT processing unit, (22) is a trajectory length calculator , (23) is an interpolation interval calculator, (24) is an interpolation point extractor, and (25) is a Chebyshev distance calculator. Figure 1 2 J 5 6 Figure 13 Figure 14 Procedural amendment 1. Indication of the case 1981 Patent Application No. 106178 3, Person making the amendment Relationship to the case Patent applicant address 6-7-35, Kitashinyo, Tokyo Parts Ward Name (2)
18) Sony Corporation Representative Director Norio Ohga 4, Agent 6, Number of inventions increased by amendment (1) In the specification, from the end of page 3 to the first line of page 11, "Figure 2...・Easy to misrecognize.'' is corrected as follows. "Figure 2 shows a conceptual diagram of the DP matching process. The input parameters (M = 19) are arranged on the horizontal axis, the standard parameters (N = 12) are arranged on the vertical axis, and the (
M, N) There are M x N points on the grid plane, and one distance corresponds to each point. For example a3 and b
The distance from 5 corresponds to . which is located at the intersection of a straight line extending vertically from a3 and a straight line extending horizontally from b5. In this case, if we take the Chebyshev distance as the distance, the Chebyshev distance d (3°5) between a3 and b5 becomes (In this case, since the dimension corresponding to the frequency axis direction i is omitted, I = 1. be.). Then, as a DP-bus with fixed end points, for the grid point (m, n), the state before connecting to this grid point (m, n) is the left grid point (rn-1゜n), the diagonally lower left side Lattice point (m-1, n-
1) and the lower grid point (m, n-1), the Chebyshev distance l between the starting point, that is, al and bl
Starting from the point ■ indicating id (1, 1), the bus (route)
Select three directions as follows, and take a bus to the end point, that is, point ■, which indicates the Chebyshev distance Md (M, N) between aM and bN.
Find the one that minimizes the sum of the distances of each passing grid point,
The result obtained by dividing the sum of these distances by (M+N-1), which is obtained by subtracting 1 from the sum of the number of input parameters M and the number of standard parameters, is the parameter time series a of the input pattern.
This is the DP matching distance between i, . . . , aM and the parameter time series bl+ . . . bN of the standard pattern. The initial condition and recurrence formula showing such processing are expressed as the initial condition g (1, 1) = d (1, 1) recurrence formula, and from this, the DP matching distance DM (A, B
) is DM (A, B) -g (M, N)/ (M+N-1
) (The reason why g (M, N) is divided by (M+N-1) is to correct the difference in distance value due to the difference in the number of frames N of the standard pattern.) Through such processing, when the number of standard patterns is IJIW, L DP matching distances for the input cover patterns are obtained, and the standard pattern having the minimum distance among the L DP matching distances is determined as the recognition result. As described above, the speech recognition device using P matching processing can deal with variations in speaking speed and differences in word length, that is, perform speech recognition with time axis normalization. However, when speech recognition is performed using P matching processing like this, the constant density part of the speech is largely reflected in the DP matching distance, and it is easy to misrecognize between words that are partially similar. It became clear that. In other words, the acoustic parameter time series Pi (nl) can be thought of as drawing a trajectory in its parameter space.In reality, the parameter of each frame n corresponds to one point in the parameter space, so although it is a point sequence, it is If you connect them with a curve in the series direction, you can think of one trajectory from the starting point to the ending point.For example, two types of words "SAN" and "IAI"
, the standard patterns A' and B' are "S", "A", "N", and "H", respectively, as shown in FIG. Draw a trajectory that passes through each phoneme region of “A” and “I”. If you say "SAN" in the recognition mode, overall there are very few similarities between the standard pattern B' and the input cover turn A, but the standard pattern B' is similar to the input cover turn A.
The "A" part of "SAN" is the "SAN" of standard pattern A'
” is more similar to the “A” part of the “HAI” of the standard pattern B′, and that part (semi-densified part)
may have a large number of points. Here, as shown in FIG. 3, the parameters of the input cover pattern A are generally similar to the parameters of the standard pattern A',
Taking a one-dimensional parameter as an example, a case will be described in which a DP matching process causes erroneous recognition when the parameter is partially similar to the parameter of the standard pattern B'. In this case, the situation shown in FIG. 3, that is, the input cover turn A as 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 standard pattern B' as shown in FIG. 6; 7.6.
6.8.8.8.8.6. 4.4.4. As is clear from the patterns shown in FIGS. 4 to 6, the inlet pattern A is a pattern that is desired to be determined as the standard pattern A'. However, when calculating the DP matching distance of the standard patterns A' and B' with respect to the input cover turn A, it is shown that the input cover turn A is close to the standard pattern B'. That is, the DP of the standard pattern A' for the input cover pattern A
Similar to FIG. 2, as for matching processing, as shown in FIG. 7, the horizontal axis represents the parameter time series of input cover pattern A; 2.4.
6.8.8.8.8.6.4.4゜4, 6.8 are lined up,
Parameter time series of standard pattern A' on vertical axis; 3.5
．． 7. 'It, 9.9.9.7.5°5, 7.9 are arranged, and Chebyshev of the individual parameters of the standard pattern A' for the individual parameters to the input cover pattern corresponding to the intersections in the lattice plane. Keep your distance. Then, the first parameter 2 of the parameter time series of input parameter A and the first parameter 2 of the parameter time series of standard parameter A'
Chebyshev distance d (1, 1
) -1 as the starting point, the 13th parameter 8 of the parameter time series of input pattern A, and the standard pattern A.
The end point is the Chebyshev distance d (13, 12) = 1 with the 12th parameter 9 of the parameter time series of If you are allowed to take a point to the left, a point below, and a point diagonally to the lower left of that arbitrary point (this bus is indicated by a solid arrow), the point on the path will be 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
There are 14 points (13, 11) - d (13, 12), and the sum of their distances is 24, and this DP matching distance DM (A', A') is 1. On the other hand, DP of standard pattern B' for input cover pattern A
The matching process is performed as shown in FIG. 8, similar to the case shown in FIG. 7 above. That is, the individual parameters to the input cover pattern; 2.4.6°8, 8.8.8.6.4.4.
Individual parameters of standard pattern B' for 4.6.8; 7.6.6.8°8, 8.8.6. ．． 4.4.4
If we calculate the Chebyshev distance of , and allow the previous state of any point to be taken as a DP-bus, a point to the left, a point below, and a point diagonally to the lower left of that point (this bus is represented by the solid line arrow). ), and the point on the bus 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 rlL 11) -d (12,11) -
d (13, 11>), the sum of the distances is 15, and this DP matching distance DM (A, B'
) is 0.65. As is clear from the result of flipping the DP-bus 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. ”
(2) Same, page 14, lines 4 to 5, ``Differences in personal characteristics depending on the speaker of the voice'' should be corrected to ``49 differences in the vocal cord sound source depending on the speaker of the voice.'' (3) Same page, line 15 to page 15, line 2 (N=15)
・ ・ ・ (2) (N=15) ・ ・ ・ (3) ” should be corrected as follows. (1 = 15) ・・・ (2) ・ ・ ・ (5) Correct the statement as follows. ... (5)'' (5) Same, page 17, lines 3 to 4, rNAT processing unit (2
I). '' should be corrected as follows. is supplied to the rNAT processing unit (21) (this NAT is
normalization Along Trajec
It is an acronym for tory. )6'' (6) Same, page 19, line 5 [mouth 1 (1) = P t(
O) + 口i(M) = P 1(S) Jru. (7) Same, page 20, lines 5 to 6 rS (111=D (yen (n+1) + yen (nl)
(n=1.=・,N)...(7)" is corrected as follows. rs, fnl-D (Pi (n+1), Pi(n
l) (n=L...,N-1) ・ ・ ・(7)''
(8) Same, page 26, lines 16 and 18, and page 27, lines 16 to 17 “Chebyshev Distance Calculator (25) J
[Chebyshev Distance Calculator (25) J] (9) Same, page 33, lines 16 to 17, ``Chebyshev distance calculator'' is corrected to ``Chebyshev distance calculator''. 00) In the drawings, Figures 3, 7, and 8 are corrected as shown in the attached sheet. that's all

Claims (1)

[Claims] It has an audio signal input section, converts the audio signal from the audio signal input section into an acoustic parameter series, estimates a trajectory in the parameter space from the acoustic parameter series, and recognizes the audio signal based on the trajectory. A speech recognition device characterized by: