JPH04254897A - Voice recognition system - Google Patents

Abstract

PURPOSE:To significantly reduce processing amount in the DP matching of input voice and registered voice. CONSTITUTION:In a characteristic quantity extraction pat 1, as the characteristic quantity of input voice, a first characteristic quantity having relatively less information quantity, and a second characteristic quantity having relatively larger information quantity are extracted. In a matching part 4, first, on the basis of both the first characteristic quantity of input voice and the first characteristic quantities of respective registered voice inside a registered-voice memory 3, a matching path of the input voice and the respective registered voice is determined by a dynamic programming, and while significantly restricting the matching path, in accordance with this matching path, tolerances of the input voice and the respective registered voice are obtained by using the second characteristic quantity having relatively larger information quantity. One candidate out of the tolerances of respective registered voice is selected, and this is output as the result of recognition.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method for recognizing input speech by matching the features of the input speech with the features of a standard pattern.

0002

[Prior Art] When trying to recognize input speech by matching the features of the input speech with the features of a standard pattern, it is common to Since the length changes each time and expands/contracts nonlinearly, it is necessary to perform time normalization to nonlinearly expand/contract the time axis so that the same phonemes in the standard pattern of the input voice and the registered voice correspond to each other. For this reason, matching using dynamic programming (DP) (hereinafter referred to as DP matching) is employed in this type of speech recognition method.

However, DP matching usually requires a large amount of processing, and in particular, continuous DP used in the word spotting method, which is suitable for speech recognition in noise, requires an even larger amount of processing.

In order to reduce the amount of processing in DP matching, a speech recognition method has been known, for example, as disclosed in Japanese Patent Laid-Open No. 1-283599. In this speech recognition method, BPF (
The output value of a bandpass filter), the LPC (linear prediction) analysis result, etc. are extracted as the first feature quantity of the input speech, and the short-term energy (power) increase/decrease trend and the formant transition state are extracted as the second feature quantity. Extract. After that, DP
In matching, the distance between these frames is calculated from the first feature amount of the input voice and the first feature amount of the registered voice. DP matching with a limited matching range is performed using information on the temporal correspondence between the patterns of the input speech and the registered speech based on the local similarity with the feature quantities of
The final recognition result is obtained based on the matching value.

[0005]

[Problems to be Solved by the Invention] As described above, in the conventional speech recognition method described above, the input speech recognition method is based on the local similarity between the second feature amount of the input speech and the second feature amount of the registered speech. Although the amount of processing in DP matching can be reduced to some extent by locally restricting the path of DP matching of the first feature using information on the temporal correspondence between patterns with registered speech, the amount of processing in DP matching can be reduced to some extent; Since there is a large amount of information,
Even if the DP matching paths are locally restricted, there is a problem in that the amount of processing cannot be significantly reduced.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech recognition method that can significantly reduce the amount of processing required for DP matching between the feature quantities of input speech and the feature quantities of registered speech.

[0007]

[Means for Solving the Problems] In order to achieve the above object, the present invention provides a voice that recognizes an input voice by matching the feature amount of the input voice with the feature amount of a plurality of registered voices registered in advance. In the recognition method, the feature amount of the input voice and the feature amount of each registered voice are both a first feature amount with a relatively small amount of information and a second feature amount with a relatively large amount of information. Based on the first feature amount of the input voice and the first feature amount of each registered voice, a matching path between the input voice and each registered voice is determined using dynamic programming, and the matching path is determined. the second feature amount of the input voice and the second feature amount of each registered voice that respectively temporally correspond to
The feature is that the likelihood between the input voice and each registered voice is determined using the feature amounts, and the recognition result of the input voice is obtained based on the likelihood. Further, the first feature amount of the input voice and each of the registered voices is a feature amount using a rate of change in the power of the voice. Furthermore, when determining the first feature amount of the input voice, if the power value of the input voice is less than a predetermined lower limit of power, the lower limit of the power is set to the power value of the input voice. It is characterized by being used as a. In addition, the first feature amount of the input voice and the first feature amount of each registered voice
When determining the matching path between the input voice and each registered voice based on the feature value of the input voice, the distance between the first feature value of the input voice and the first feature value of each registered voice is determined, The system is characterized in that candidates are narrowed down according to the value of , and the likelihood is calculated using the second feature amount only for the registered voices of the narrowed down candidates.

[0008]

[Operation] In the present invention, the matching path between the input voice and each registered voice is determined by dynamic programming using the first feature amount, which has a relatively small amount of information, and the matching path is significantly limited. According to this matching path, the likelihood between the input voice and each registered voice is determined using the second feature having a relatively large amount of information, so that the amount of processing can be reduced.

[0009] As the first feature having a relatively small amount of information, the rate of change in the power of the voice can be used, and in this case, the power value of the input voice is less than the predetermined lower limit of the power. In this case, this lower limit value of power is used as the power value of the input audio.

[0010] Furthermore, when determining the matching path between the input voice and each registered voice, the distance between the first feature of the input voice and the first feature of each registered voice is determined, and the distance is calculated according to the distance value. The amount of processing can be further reduced by narrowing down each registered voice, that is, the candidates, and then calculating the likelihood using the second feature amount.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of one embodiment of the present invention. Referring to FIG. 1, in this embodiment, a feature amount extraction unit 1 extracts a feature amount from an input voice at a fixed frame period (10 to 20 ms), and a voice interval detection unit 2 detects a voice interval of the input voice. , a registered voice memory 3 in which feature quantities of a plurality of registered voices are stored in advance, and a matching unit that performs DP matching between the feature quantities of input speech and the feature quantities of registered speech using dynamic programming (DP). 4 is provided.

[0012] The input voice input to the feature extraction unit 1 is inputted to the A
/D converter, and the feature extraction unit 1 extracts a first feature with a relatively small amount of information and a second feature with a relatively large amount of information as the feature of the input voice. The feature quantity is extracted.

For example, the rate of change in power of the input voice is used as the first feature of the input voice, and in this case, the first feature Px(i) of the input voice is the input voice of each frame. If the power of the voice is e(i),

[0014]

It is extracted as [Equation 1]. Here, i is a frame number.

[0015] Furthermore, as the second feature amount, for example, a known short-time spectrum Sx (i, ω) obtained from a band pass filter (BPF) bank is used.

[0016] The voice section detecting section 2 detects a voice section based on, for example, whether the power e(i) of the input voice exceeds a predetermined threshold.

The registered speech memory 3 also stores, as feature quantities of the registered speech, a first feature quantity Ptk(j) and a second feature quantity Ptk(j) in correspondence with the first and second feature quantities of the input speech. Feature amount S
tk(j, ω) is stored. Here, j is the frame number and k is the registered audio number.

The matching unit 4 extracts the first feature quantity Px(i) of the input speech of the speech interval detected by the speech interval detection unit 2.
DP matching is performed by dynamic programming (DP) using the first feature amount Ptk(j) for each registered voice,
The minimum distance is stored as the distance between Px(i) and Ptk(j), and the matching path that gives that distance is stored, and in this way, the first feature for all registered voices is After determining the amount of distance, registered voices that have a distance greater than a threshold value are excluded from recognition targets, and candidates are narrowed down. The matching unit 4 then calculates the second feature amount Sx of the input speech.
(i, ω) and the second feature Stk (j, ω) of each registered voice of the narrowed down candidates, the distance is calculated using only the matching path stored for each registered voice. ,
This is taken as the distance between the input voice and each registered voice, and if the minimum distance among these distances is less than a predetermined threshold, the category of the registered voice that gave the minimum distance is output as a recognition result. It looks like this.

Next, the speech recognition processing operation in such a configuration will be explained. In this embodiment, the first input audio
As the feature amount, a power variation amount Px(i) for each frame i as shown in Equation 1 is used. As a result, the first feature amount Px of the input voice is

[0020]

It is expressed in the time series of [Equation 2]. Further, the first feature amount Ptk of the registered voice with registered voice number k is:

[0021]

[Math 3] It is expressed in chronological order.

Considering a plane consisting of Px and Ptk as shown in FIG. 2, matching path L, that is, Px and P
The correspondence of the time axis with tk is the grid point C= on this plane.
The sequence F of (i, j), that is,

[0023]

It can be expressed as [Equation 4]. Two feature quantities Px(i)
, Ptk(j) is expressed as d(c)=d(i,j), the total distance along F is Dk(F),

002
4]

It can be expressed as [Equation 5], and the smaller this value is, the smaller Px and P
Indicates that the correspondence with tk is good. Here, wl is F
is a positive weight function associated with .

[0025] In the matching unit 4, dynamic programming (DP)
Using , minimize Equation 5 with respect to F under the following constraints. In other words, as the conditions for monotonicity and continuity,

[
0026

[Math 6] and as a boundary condition,

[0027]

[Equation 7] is set, and as a condition for the matching window W (i.e., r is a constant to prevent extreme expansion and contraction),

[0028]

Equation 8 can be simplified by setting wl such that the denominator in Equation 5 is a constant that does not depend on F. for example,

[0029]

[Equation 9] Then, wl is the path along the vertical and horizontal lines of the Go board.
, that is, the urbanization distance,

[0030]

[Math. 10] becomes. At this time, the number 5 is

[0031]

[Formula 11], and the objective function to be minimized becomes additive. In dynamic programming, to perform this minimization,

[0032]

Considering the partial sum of [Equation 12], using the conditions of Equations 6 to 8 and Equation 9,

[0033]

Expressed by the recurrence formula [Equation 13], g (1, 1) = 2d (1, 1), J
= 1, and while changing i within the matching window W, use equation 13.
, then increase j and repeat the same calculation until j=J. Finally, the feature quantity Px(i
) and the feature amount Ptk(j) of the registered speech, the distance Dk after time normalization between the two time series is

[0034]

[Math. 14] It can be found as

The distance Dk obtained in this way and the matching path L given the distance Dk, that is, the equation 13
The history of (i, j) selected by the calculation is stored in a predetermined memory (not shown). For example, the constant r of the matching window W
If you use , there will be a maximum of (2r+1) paths, and
One maximum length is (Imax+2r+1). Here, Imax is the maximum frame length of input audio. Therefore, to store the matching path L, (2r+1)(Imax+2r+1) memories are required. As a mnemonic for matching path MP to memory,
When (a) is selected in Equation 13, g(i,
Add (i, j) to the path of g(i-1, j-1) if (b) is selected, add (i, j) to the path of g(i-1, j-1)
and (i, j), and if (c) is selected, g
(i, j) is added to the path of (i-1, j), and one of them is stored in the memory where the path of g(i-1, j-1) was stored. The final path L is the path of g(I, J), and is stored as shown in FIG. 3, for example. Note that each data may be expressed and stored using only 2-bit information indicating whether i increases or not and whether j increases or not.

[0036] Thereafter, the matching unit 4 excludes, from the recognition targets, registered voices to which a distance greater than a predetermined threshold has been given from among the distances Dk for each registered voice stored in the memory.
Narrow down the candidates. After narrowing down the candidates,
Next, the second feature amount S of only the narrowed down candidates is
tk(j, ω) and the second feature amount Sx(j, ω) of the input speech
) using the path L obtained as above,

[0037]

[Equation 15] Here, d'(ik(n), jk(n)
)teeth,

[0038]

[Formula 16]. Note that ω corresponds to the frequency, and the number of channels E is, for example, 1/3 oc from 250Hz to 6350Hz.
t (generally, the number of channels E is about 8 to 30).

In this way, the second feature amount Sx (i, ω) of the input voice and the second characteristic amount of each registered voice that has been narrowed down are
calculate the distance D'(L) to the feature Stk(j, ω), find the minimum distance among these distances, and if this is less than a predetermined threshold, select the candidate with this minimum distance. category is output as a recognition result.

By the way, in this embodiment, the processing amount of DP matching is the processing amount when calculating the matching path using the first feature amount and the processing amount when calculating the distance D'(L) of the first feature amount. It is the sum of the processing amount.

[0041] When finding a matching path using the first feature, it is first necessary to calculate d(i, j), and d
In calculating (i, j), for each i, j, the power e(i) of the input voice is divided by {e(i)/e(
i-1)} once, its logarithm operation log once, and
The first feature amount Px(
i) and the first feature Ptk(j) of the registered voice is required once, and its absolute value is calculated once, for a total of 4
The processing amount will be 3 times.

[0042] Furthermore, after d(i, j) is calculated in this way, it is necessary to calculate g(i, j), and g(i,
In the calculation of j), for each i and j, the addition operation of (a) of Equation 13 is performed once, the addition operation of (b) of Equation 13 is performed twice, and the addition operation of Equation 13 (b) is performed twice.
3 (c) addition operation is done once, (a), (b), (c)
This requires two comparison operations, resulting in a total of six processing steps.

[0043] Therefore, (4+6) times, that is, 10 times, are required for each i and j, and for each i, [10·(2r+
1)] times, and when calculating the distance of the first feature amount, a total of [10·(2R+1)·I] times of processing are required.

In addition, when calculating the distance D'(L) of the second feature amount, in Equation 16, for each channel, logarithmic operation Log is performed once, subtraction operation is performed once, and absolute value calculation is performed once. Therefore, assuming that the number of channels is E, a processing amount of (3E) times is required at this stage. Furthermore, to add the processing results for each channel, the addition operation is (E-1)
times are required. Therefore, in Equation 16, for each n, the sum (
4E-1) times of processing is required, and in equation 15,
When calculating the distance D'(L) of the second feature amount, the total [
A processing amount of (4E-1)·(I+J)] is required.

[0045] As a result, the total processing amount Q according to this embodiment is:

[0046]

[Math. 17] becomes.

Next, the amount of processing in the speech recognition method of this embodiment will be compared with the amount of processing in the conventional speech recognition method as disclosed in Japanese Patent Laid-Open No. 1-283599. In conventional speech recognition methods, d(i,j) and g(i,j) are respectively

[0048]

[Math. 18]

[0049]

[Formula 19] When calculating d(i,j), one logarithm operation, one subtraction operation, and one absolute value operation are required for each channel, and the number of channels is If it is E, at this stage, a processing amount of (3E) times is required. Furthermore, in order to add the processing results for each channel, the addition operation (E
-1) times are required. As a result, in Equation 18, d(i
, j), the sum (4E-1) for each i, j
The amount of processing is required.

[0050] Furthermore, for the calculation of g(i, j) in Equation 19,
Four addition operations and two comparison operations are required, so each i
, j, a total of six processing times is required.

[0051] As a result, for each i, [(4E+5)・
(2r+1)] times of processing is required, and the total processing amount R is:

[0052]

[Math. 20] becomes.

Now, assuming that the number of channels E is "15", the constant r of the matching window is "20", and I and J are each "60", in the conventional speech recognition method, the total processing amount R required for DP matching is is about 160,000 according to number 20
In contrast, in the speech recognition method of this embodiment, the DP
The total processing amount Q required for matching is 3 according to equation 17.
It only takes 2,000 times, and in this example, it is about 1/1 time compared to the conventional method.
It becomes possible to reduce the processing amount to about 5.

In this way, in this embodiment, the matching path between the input voice and each registered voice is determined by dynamic programming using the first feature amount, which has a relatively small amount of information, and the matching path is greatly improved. , and then calculates the distance between the input voice and each registered voice using the second feature amount, which has a relatively large amount of information, according to this matching path.
When using DP matching, it is also possible to significantly reduce the amount of processing.

Furthermore, in this embodiment, the candidates are narrowed down according to the distance value using the first feature obtained when determining the matching path, so the amount of processing can be further reduced. can.

Furthermore, in this embodiment, since the rate of change in the power e(i) of the input voice is used as the first feature, even if there is a difference in the level of the input voice, the influence of this can be reduced.

Note that when calculating the rate of change of the power e(i) of the input voice as the first feature, the value of the power e(i) of the input voice is the lower limit value em of the predetermined power.
If in (>0) is not satisfied, it is better to use the lower limit emin as the value of the input audio power e(i),
Thereby, it is possible to reduce the error in the first feature amount when the power e(i) of the input voice is small. Alternatively, in this case, the power may be smoothed over several frames and used as the input audio power e(i).

Further, as the first feature amount, the logarithm value of the power of the input voice normalized by the maximum power may be used, or the value before logarithmic transformation may be used. Alternatively, a zero crossing point may be used.

Furthermore, the short-time spectrum as the first feature in the above example can also be obtained by FFT. Furthermore, other feature quantities such as cepstrum and mel cepstrum obtained by LPC analysis may be used as the second feature quantity.

[0060] Furthermore, the speech section detecting section 2 may detect speech sections using other methods such as a two-threshold method.

[0061]

[Effects of the Invention] As explained above, according to the present invention,
Using the first feature, which has a relatively small amount of information, dynamic programming is used to determine the matching path between the input voice and each registered voice, and after greatly limiting the matching path, relatively Since the second feature having a large amount of information is used to determine the likelihood between the input voice and each registered voice, the amount of processing can be reduced.

Furthermore, by using the rate of change in the power of the voice as the first feature having a relatively small amount of information, even if there is a difference in the level of the input voice, it is possible to prevent the influence of this difference. In addition, in this case, if the power value of the input voice is less than the predetermined lower limit value of the power, the lower limit value of the power is used as the power value of the input voice, so that the power of the input voice is small. It is possible to reduce the error of the first feature amount in the case of the first feature amount.

In addition, when determining the matching path between the input voice and each registered voice, the distance between the first feature of the input voice and the first feature of each registered voice is determined, and the distance is calculated according to the distance value. The amount of processing can be further reduced by narrowing down each registered voice, that is, the candidates, and then calculating the likelihood using the second feature amount.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of one embodiment of the present invention.

FIG. 2 is a diagram for explaining a matching path between input speech and registered speech.

FIG. 3 is a diagram showing specific data of matching paths between input speech and registered speech.

[Explanation of symbols]

1 Feature extraction part 2 Voice section detection unit 3 Registered voice memory 4 Matching part

Claims (4)

[Claims] Claim 1. A speech recognition method that recognizes input speech by matching the feature amount of the input speech with the feature amount of a plurality of registered voices registered in advance, wherein the feature amount of the input speech and each of the registered voices are matched. The feature amounts include a first feature amount that has a relatively small amount of information, and a second feature amount that has a relatively large amount of information, and the first feature amount of the input voice and each of the registrations. A matching path between the input voice and each registered voice is determined using dynamic programming based on the first feature amount of the voice, and a second feature of the input voice that respectively temporally corresponds to the matching path is determined. The method is characterized in that the likelihood of the input voice and each registered voice is determined using the quantity and the second feature of each registered voice, and the recognition result of the input voice is obtained based on the likelihood. A voice recognition method that uses 2. The speech recognition method according to claim 1, wherein the first feature quantity of the input speech and each registered speech is a feature quantity using a rate of change in the power of the speech. 3. When determining the first feature amount of the input voice, if the power value of the input voice is less than a predetermined lower limit value of power, the lower limit value of the power is set to the lower limit value of the input voice. 3. The speech recognition method according to claim 2, wherein the speech recognition method is used as a power value. 4. When determining a matching path between the input voice and each registered voice based on the first feature of the input voice and the first feature of each registered voice, the first feature of the input voice is determined. The distance between the amount and the first feature of each registered voice is determined, the candidates are narrowed down according to the value of the distance, and the second feature is used only for the registered voices of the narrowed down candidates. 4. The speech recognition method according to claim 1, wherein the likelihood is determined by using the method.