JPH04281500A - Voiced sound phoneme recognition method - Google Patents

Abstract

PURPOSE:To improve the phoneme recognition performance by providing a method whereby the characteristics extracting process of voice waveforms becomes the most suitable per recognized phoneme. CONSTITUTION:First, a neurocircuit network which has a learning process is provided. This network outputs impulse string, which has the same period of the basic period of an input phoneme and corresponding to the input phoneme, when voiced sound phoneme waveforms are inputted. The signal value, whereby the phoneme normalizes the size of unknown input voice waveforms by the power over the prescribed time section, is inputted into the neurocircuit network (S20), plural output series peaks of the neurocircuit network are detected by slightly time shifting the added input voice waveforms (S22) and assume the phoneme, which corresponds to the output having the maximum peak value, as the phoneme recognition result (S23).

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voiced phoneme recognition method in speech recognition technology, which performs voiced phoneme recognition processing using a neural network or neural network group that has undergone learning processing.

0002

2. Description of the Related Art Conventionally, there are various methods for recognizing voiced phonemes, and as an example, the method for recognizing voiced phonemes shown in FIG. 2 will be described.

FIG. 2 is a flowchart showing the processing procedure of a conventional voiced phoneme recognition method.

[0004] In this voiced phoneme recognition method, step S1
After performing audio waveform extraction processing in which the audio signal to be recognized is multiplied by a window function to extract the audio signal in a certain time area at a desired time, in step S2, the audio signal is extracted based on the autocorrelation function of the audio waveform. , calculate the linear prediction coefficients. Then, in step S3, pattern matching is performed using the linear prediction coefficient obtained in step S2 as a feature quantity, and the phoneme of the standard pattern having the minimum distance value is output as the phoneme recognition result.

[0005] The contents of each processing step S1 to S3 will be explained below.

First, in step S1, a part of unknown input speech is extracted. Let s(m) be a speech waveform with unknown discrete phonemes in the time domain, and let w(m) be an appropriate window function. Here, m is a discrete time. Now, let n be a desired discrete time in the speech waveform at which the type of phoneme is to be determined. At this time, the audio waveform sn (
m) is determined by the following equation (1). sn (m)=sn (m+n)w(m)
...(1) However, 0≦m≦N-1 N: Desired window function size In step S2, the autocorrelation function Rn (k) of the speech waveform at the desired time n is calculated using the following equation (2). demand. However, k: lag, and the linear prediction coefficient is determined. For example, Durbin (D
According to the recursive method of urbin), linear prediction coefficients αj (where j: j-th prediction coefficient) can be obtained using the following equations (3) to (8). E(0)=R(0)
・
...(3) However, p: is the order of linear prediction and is arbitrary. αi (i) = ki

...(5) αj (i)=αj (i-
1) -ki αi-j (i-1)
...(6) E(i)=(1-
ki2)E(i-1)
...(7) As a calculation procedure for the linear prediction coefficient αj, first, E is calculated using equation (3).
(0) is obtained. Next, equations (4) to (7) are calculated, and αj (i) is determined recursively in the range of 1≦i≦p. Then, from the following equation (8), the linear prediction coefficient α
get j. αj = αj (p)
1 ≦j ≦p
(8) Finally, in step S3, pattern matching with the phoneme standard pattern is performed. In this process, in advance,
Linear prediction coefficients αj are determined for each phoneme for many speech waveforms whose phonemes are known, and their representative values are stored as standard patterns. Here, the average value of the linear prediction coefficients αj for the phoneme φ is calculated in advance as a standard pattern αj φ. When recognizing a phoneme, the distance Dx φ between the linear prediction coefficient αj of the unknown input speech and the standard pattern αj φ of the phoneme φ is
is measured using the following equation (9).

[0007]

[Math 1]

The phoneme φ whose distance Dxφ has the smallest value is output as the voiced phoneme recognition result at a desired time n.

[0009]

[Problems to be Solved by the Invention] However, the conventional voiced phoneme recognition method has the following problems.

[0010] In conventional methods, specific feature extraction processing is performed on speech waveforms, such as linear predictive analysis or bandpass filter bank, and phoneme recognition is performed using the feature quantities obtained by this feature extraction processing. . However, if the feature extraction process is set to a constant feature extraction method regardless of the phoneme, if the optimal feature extraction process is different for each phoneme, the recognition performance for each phoneme will depend on the feature extraction method. A limitation arises due to the fact that

For example, in the conventional voiced phoneme recognition method, feature amounts are obtained for each fixed time unit called a frame, and it is known that the length of this frame greatly influences the phoneme recognition performance. This is because no matter how good the recognition method after feature extraction is, recognition performance will deteriorate due to deterioration of phonological information in the feature extraction method or deterioration of overall performance due to the inappropriateness of the combination of feature extraction method and recognition method. means that Therefore, it has been difficult to obtain highly accurate phoneme recognition results.

[0012] The present invention solves the problem that the prior art had, and if feature extraction processing is determined to be a specific processing method, it is not necessarily possible to apply the optimal feature extraction processing method for voiced phoneme recognition, resulting in poor phoneme recognition performance. The present invention provides a voiced phoneme recognition method that solves the problem of a decrease in voiced phoneme recognition.

[0013]

[Means for Solving the Problems] In order to solve the above problems, the first invention provides a voiced phoneme recognition method for recognizing voiced phonemes, in which the magnitude of the voiced phoneme signal is used as a presentation signal for learning. A signal value normalized by (power) is input over a certain time interval, and when a predetermined position in the input time interval and the epoch point of the input signal match, the output corresponding to the input phoneme takes a large value, At other times, a neural network trained using the error backpropagation method is used so that it takes a small value.

[0014] Then, a signal value obtained by normalizing the magnitude of an input speech waveform whose phoneme is unknown over a certain time interval by power is input to the neural network, and the time of the input speech waveform to be applied is shifted little by little. The peaks of a plurality of output series of the neural network obtained by the above are detected, and a phoneme recognition judgment is performed based on the peak detection results, and the phoneme corresponding to the output that produces the largest peak value is determined as the phoneme recognition result.

The second invention uses a neural network group consisting of a plurality of neural networks in place of the neural network of the first invention.

[0016] After performing learning processing on this neural network group, a signal value obtained by normalizing the magnitude of an input speech waveform whose phoneme is unknown over a certain time interval by power is applied to the neural network group. The peak of each output series of the neural network group obtained by slightly shifting the time of the input speech waveform that is input and added is detected, and a phoneme recognition judgment is performed based on the peak detection result to determine the largest peak value. The phoneme corresponding to the neural network that produces the above is taken as the phoneme recognition result.

[0017]

[Operation] According to the first invention, when a voiced phoneme waveform is input in advance, a learning process is performed so that the output corresponding to the input phoneme outputs an impulse train with the same period as the fundamental period of the input phoneme. Prepare a neural network. Then, when an input speech waveform signal with an unknown phoneme is input to the neural network, the phoneme corresponding to the output that outputs the largest impulse train is taken as the phoneme recognition result.

In this way, processing from waveform input to feature extraction and phoneme recognition is performed as a consistent processing method, and the feature extraction parameters that determine the characteristics of the processing method are optimized using neural networks and learning processing. , the feature extraction process becomes an optimal method for each recognized phoneme, and the phoneme recognition performance can be improved.

According to the second invention, when a voiced phoneme waveform is inputted in advance, a learning process is performed so that the output corresponding to the input phoneme outputs an impulse train having the same period as the fundamental period of the input phoneme. Multiple neural networks are prepared for each recognition target. Then, when an input speech waveform signal whose phoneme is unknown is input to the plurality of neural network groups, the phoneme corresponding to the neural network that outputs the largest impulse train is set as the phoneme recognition result. Compared to , the amount of processing is larger, but the phoneme recognition performance can be further improved. Therefore, the above problem can be solved.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First Embodiment FIG. 3 is a functional block diagram of a voiced phoneme recognition device showing a first embodiment of the present invention.

This voiced phoneme recognition device has a waveform input section 10 that stores input signal waveforms, and a neural network 20 that performs learning and recognition processing is connected to the output side of the waveform input section 10. Neural network 20 includes a number of processing units (i.e., cells).
21, and a recognition evaluation unit 30 is connected to the output side of the recognition evaluation unit 30, which evaluates the output and outputs a recognition result.

The neural network 20 also includes a learning control section 40.
is connected to the learning control section 40, and further connected to the learning control section 40 is a teacher signal input section 50 that inputs a teacher signal during learning. The learning control unit 40 has a function of evaluating the output error of the neural network 20 during learning and changing the connection weight coefficient of the neural network 20.

[0023] This voiced phoneme recognition device can be realized using any method, such as computer software, dedicated digital circuits, analog circuits, photoelectric elements, etc.
They may also be combined. In the embodiment, computer software is used.

FIG. 1 is a flowchart showing the processing procedure of a voiced phoneme recognition method using the apparatus shown in FIG.

The processing procedure of the voiced phoneme recognition method of this embodiment includes the learning processing procedure of the neural network shown in FIG.
It consists of the phoneme recognition processing procedure using the neural network after the learning processing shown in (b). Below, (I)
The learning process and (II) phoneme recognition process will be explained.

(I) Learning process of neural network Figure 1 (a
), the neural network learning process includes an initialization process for neural network learning (step S10), a presentation signal (hereinafter referred to as
Presentation signal input processing (step S11) for inputting a presentation signal (referred to as a presentation signal), forward propagation processing of the neural network (step S12), output error calculation processing of the neural network (step S13), error of the neural network Backpropagation learning process (step S14) and learning end determination process (step S15)
). Below, each step S10
The contents of ~S15 will be explained with reference to FIGS. 3 and 4.

Note that FIG. 4 is a diagram showing an example of an audio waveform used as a presentation signal in the first embodiment. Figure 4
60 is the presentation signal, 61 is the position of the epoch point that gives the peak of the teacher signal, 62 is an example of the time interval of the presentation signal when 0.1 is given as the teacher signal, and 63 is 0.9 is given as the teacher signal. This is an example of a time interval of a presentation signal in the case of FIG.

In the learning process, a discrete signal waveform in the time domain is assumed to be s(m), and a presentation signal is particularly assumed to be as(m). In this embodiment, the presentation signal 60 is a vowel waveform generated by a man sampled at 12 kHz and 12 bits. In advance, an epoch point position 61 (hereinafter referred to as a teacher epoch point) at which a peak of the teacher signal is given to the presentation signal 60 is set by human inspection.

In step S10 of FIG. 1(a), a variable p representing the number of times learning is performed as an initialization process is set to 1.

In step S11, the waveform input unit 10
extracts the presentation signal 60 from the section centered on the teacher epoch point. The time interval of the presentation signal in this case is indicated by reference numeral 63 in FIG. The length of this time interval 63 is set to be the number of samples equal to the number of cells in the input layer of the neural network 20. In this embodiment, 512 samples are used as the presentation signal. In this embodiment, the phoneme φ of the presentation signal is regarded as a number from 0 to 4 corresponding to the phoneme of five vowels, and the presentation signal of the phoneme number φ in this case is expressed as asφ0.9 (m). here,
0≦m≦511.

The waveform input unit 10 receives the presentation signal asφ0.9
The output opj(0) of each cell 21 in the input layer of the neural network 20 is set by power-normalizing (m) using the following equation (10) and adding an offset.

[0032]

[Math 2]

However, opj(0); output C of the j-th cell 21 for the p-th pattern in the q-th layer.
; Positive constant for normalization Here, the presentation signal asφ0
．． 9 (m) is the first pattern, and the input layer is the o-th layer.

In step S12, forward propagation processing of the neural network 20 is performed. Neural network 20 in this embodiment
The structure is a three-layer structure in which the input layer is the o-th layer, the first layer is the intermediate layer, and the second layer is the output layer.
The output of the qth layer is calculated using the following equation (11).

[0035]

[Math 3]

Here, opj(q) is the output of the j-th cell 21 in the q-th layer, when the q-th pattern is presented. Nq is the number of cells 21 in the q-th layer, w ji (q) is the weighting coefficient from the j-th cell 21 in the q-1th layer, which is input to the q-th layer, and θj (q) is the bias of the j-th cell 21 of the q-th layer. In this example, the number of cells in each layer is 512 for N0, 64 for N1, and N2
is 5. wji(q) and θj(q) are set to random small values before learning.

The calculation of equation (11) is performed in order for all q, j, and the output opj(2) of the cell 21 of the output layer, which is the second layer, is calculated in order, and the second
Output opj (2) of cell 21 of the output layer which is the th layer
get.

In step S13, the learning control section 40 performs output error calculation processing. Prior to this, the teacher signal input unit 50 determines 0.9 as the teacher signal of the output cell 21 corresponding to the phoneme φ of the presentation signal as the teacher signal of the output cell 21 for the presentation signal, and outputs 0.9 corresponding to the phoneme φ of the presentation signal. 0.1 is determined as the teacher signal of cell 21. In other words, if the teacher signal for the p-th presentation signal is tpj,
tpφ=0.9
tpj=0.1 (j ≠φ)


...(12).

The learning control section 40 compares the teacher signal from the teacher signal input section 50 and the output of the neural network 20 as follows. If the error inside the cell for the p-th presentation signal in the j-th cell 21 of the q-th layer (error before conversion using a function generally called a sigmoid function) is expressed as δpj (2), The error δpj(2) inside the cell in the output layer is calculated using the following equation (13). δpj (2) = (tpj-op
j(2) ) opj(2) (1-opj(2) )


(13) In step S14, the learning control unit 40 performs error backpropagation learning processing for the neural network 20. When the internal error δpj(q) of each cell 21 in the q-th layer has been calculated, the internal error δpj(q-1) of each cell 21 in the (q-1)th layer can be calculated using the following formula (
14). Furthermore, using the error δpj(q) inside the cell, the (qth
-1) Calculate the correction amount Δp wji (p) of the weighting coefficient wji (q) from the layer to the qth layer using the following equation (15). Δp wji (q)=ηδpj (q) opi
(q-1)

...(15) In order to speed up the learning speed, the following equation (15-1) may be used. Δp wji (q)=αΔp-1 wji(
q)+ηδpj(q)opi(q-1)

・・・・・・
(15-1) In the equation (15-1), α and η are constants that determine the learning speed, and when using the equation (15-1), Δ0 is set in the initialization process step S10.
Let wji(q)=0. Also, the bias θ in the j-th cell 21 of the q-th layer
The correction amount Δp θj (q) for j (q) is also calculated using the following equation (16). Δp θj (q) = ηδpj (q)

...(16) In order to speed up the learning speed, the following equation (1
6-1) may also be used. Δp θj (q)=αΔp−1 θj (q
)+ηδpj(q)...(16-1
) In equation (16-1), α is a constant that determines the learning speed, and when using equation (16-1), Δ0 θj (q)=0 is set in initialization processing step S10.

The calculations of equations (15) and (16) are performed in all output layers and intermediate layers while decreasing the layer number q, and all weighting coefficients wji(q) and biases θj
(q) Correction amount Δp wji(q) and Δ
Find p θj (q). All correction amounts Δp wj
After calculating i(q) and Δp θ j (q),
Using this correction amount, all weighting coefficients wji(q) and biases θj(q) are corrected by the following equation (17). wji(q)=wji(q)+Δp
wji (q)


...(17)
θj (q)=θj (q)+Δ
p θj (q)
The above processing is performed on the presentation signal asφ0.9 (m) with p=1. At this time, as the teacher signal tpj, one according to the above equation (12) is used.

[0041] In step S15, the learning control unit 40
performs learning completion determination processing. As the weighting coefficient wji(q) approaches the optimal value by repeating the learning process, the error δpj(2) within the output cell approaches 0. Determine whether the error δpj(2) inside the output cell has become a value smaller than a sufficiently small value ε, and if the error δpj(2) inside the output cell is large, determine that learning has not finished,
Add 1 to p and return to step S11. If the error δpj(2) inside the output cell is small, all learning processing is completed.

When the process returns to step S11, the section 62 which is not centered at the aforementioned teacher epoch point is used as the presentation signal.
Take. The deviation of the center of the section 62 from the teacher epoch point is random. Presentation signal in this case asφ0.1 (
m) is power normalized using the following equation (18), an offset is added, and the output of each cell 21 in the input layer is set as opj(0).

[0043]

[Math 4]

The teacher signal tpj at this time is set to the same value of 0.1 in all output cells 21, and from step S12 onwards, the same processing as described above is performed.

As shown above, asφ0.9 (m) and a
By repeatedly performing the learning process of alternately presenting sφ0.1(m) for many phonemes φ and their presentation signals, the optimal weighting coefficient can finally be obtained.

(II) Phonological recognition processing using neural network after learning As shown in FIG. 1(b), phoneme recognition processing includes input processing of neural network (step S20), forward direction of neural network It consists of propagation processing (step S21), neural network output peak detection processing (step S22), and phoneme recognition determination processing (step S23). The contents of each step S20 to S23 will be explained below with reference to FIGS. 3 and 5.

FIG. 5 is an explanatory diagram showing the relationship between the input speech signal and the output signal sequence during phoneme recognition in the first embodiment. In FIG. 5, 70 is an input speech waveform to be subjected to phoneme recognition, 71 is a section of one input speech waveform input to the neural network, and 72 is an output of the neural network for the section 71. 73 is the section of the next input speech waveform input to the neural network, 74 is the output of the neural network for this section 73, 75 is the output signal sequence of the neural network obtained by phoneme recognition processing, and 76 is the section of the next input speech waveform input to the neural network. Indicates the time of the peak extracted from the output signal sequence.

In the phoneme recognition process, x(m) is an input speech waveform whose phoneme is unknown and is discrete in the time domain. In this embodiment, a different sound from the presentation signal used in the learning process is used as the input sound waveform. Here, xu
(m).

In step S20, as an input to the neural network 20, the input speech waveform xu (m) is power-normalized using the following equation (19), an offset is added, and the output ouj(0) of each cell 21 of the input layer is do.

[0050]

[Math 5]

[0051] However, ouj(q); Output C of the j-th cell 21 for the input speech waveform centered at time u in the q-th layer; Positive constant for normalization In step S21, the output of the neural network 20 is Perform forward propagation processing. This process is performed in the same way as the forward propagation process in the learning process by replacing p in equation (11) with u. Through this process, the output ouj(2) is obtained from the output layer cell 21. moreover,
The input audio waveform is taken from a time interval centered at time u+1, and similar processing is performed. By repeating such processing, a series of outputs ouj(2) corresponding to each phoneme at time u is obtained. An example of this output series is shown at 76 in FIG.

When the center of the time interval from which the input speech waveform is extracted coincides with the epoch point of a certain phoneme in the input speech waveform, a large peak occurs in the output series corresponding to that phoneme. By detecting this peak and comparing the magnitude of the peak of each output series, the phoneme of the input speech waveform can be determined.

In step S22, peak detection processing is performed on the output series. The output is the following equations (20) and (21)
A discrete time vdj that satisfies the condition is detected as the epoch point time of the phoneme. and, where p is the threshold for detecting the peak,
In this embodiment, a constant of 0.5 is used. d is the number given to the detected peak, and j is the type of phoneme corresponding to the output.

Finally, in step S23, the magnitudes of the peaks of the output series at the epoch points of each detected phoneme are compared and used as a phoneme recognition result. In this example, the absolute magnitudes of the peaks are simply averaged over a desired time interval (here, from time Ts to Te) and compared, and the output having the largest average value is selected. The corresponding phoneme is output as the phoneme recognition result. Expressing this in a formula, first, the evaluation value Rj of each phoneme is calculated using the following formula (22). However, Nj is the number of peak times vdj included from time Ts to Te, and the evaluation value Rj is compared over all phonemes, and the phoneme corresponding to j that gives the maximum evaluation value is output as the phoneme recognition result.

The method of this phoneme recognition result judgment process is arbitrary, and various statistical distance measures may be used.

As described above, in this first embodiment,
It has the following advantages:

(i) Since feature extraction of voiced sounds and phoneme recognition are performed by the neural network 20 that integrates them, a process that combines optimal feature extraction and phoneme recognition is automatically created through learning processing. High performance phonological recognition output can be obtained.

(ii) The basic period of speech and the phoneme recognition result can be obtained simultaneously, and recognition results with high temporal resolution can be obtained. Therefore, when the method of this embodiment is used for continuous speech recognition, for example, there is no need to perform phoneme recognition and fundamental frequency extraction separately, and it is possible to obtain phoneme information and fundamental period wavenumber in a single process. can.

Second Embodiment FIG. 7 is a functional block diagram of a voiced phoneme recognition device showing a second embodiment of the present invention.

In this voiced phoneme recognition device, a neural network group 20A is provided in place of the neural network 20 in FIG. Control unit 40A and teacher signal input unit 5
0A is provided. The neural network group 20A has a plurality of neural networks 20 provided for each phoneme to be recognized, and each neural network 20 is composed of a large number of processing units (i.e., cells) 21. .

FIG. 6 is a flowchart showing the processing procedure of the voiced phoneme recognition method using the apparatus shown in FIG. The processing procedure of the voiced phoneme recognition method of this embodiment is almost the same as in FIG.
The learning processing procedure of the neural network group shown in FIG. 6b(a) and the phoneme recognition processing procedure of the neural network group after the learning processing shown in FIG. 6(b) are configured. Below, the (I
A) Learning processing and (IIA) phonological recognition processing will be explained.

(IA) Learning process for neural network group In this learning process, instead of the neural network shown in FIG. 1, initialization processing for the neural network group (step S10A), presentation signal input processing (step S11A), Direction propagation processing (step S
12A), output error calculation processing (step S13A), error backpropagation learning processing (step S14A), and learning completion determination processing (step S15A) are executed in substantially the same manner as in FIG. 1(a).

That is, similar to step S10 in FIG. 1(a), after initialization processing is performed in step S10A, the waveform input section 10A in FIG. 7 performs presentation signal input processing in step S11A.

[0064] In this step S11A, the waveform input section 10A
The presentation signal 60 in FIG. 4 is extracted from a time interval 63 centered on the teacher epoch point. The length of this time interval 63 is set to be the number of samples equal to the number of cells in the input layer of each neural network 20 in the neural network group 20A.

The structure of the number of layers of each neural network 20 may be different for each phoneme as in the first embodiment, but in this embodiment, the structure of each neural network 20 is assumed to be similar to each other. ,
Samples at 512 points, which is equal to the number of cells in the input layer, are used as presentation signals. In this embodiment, the phoneme φ of the presentation signal is regarded as a number from 0 to 4 corresponding to the phoneme of five vowels, and the presentation signal of the phoneme number φ in this case is expressed as asφ0.9 (m).

The waveform input unit 10A inputs the presentation signal asφ0
．． 9 (m) is power normalized using equation (10) and an offset is added to the output opj(β
, 0). Here, opj(β, q
) is the j-th cell 2 for the p-th pattern in the q-th layer of the neural network 20 corresponding to the β-th phoneme.
Represents the output of 1. In this embodiment, as in the first embodiment, the presentation signal asφ0.9 (m) is the first pattern, and the input layer is the 0th layer.

In step S12A, each neural network 20 in the neural network group 20A is processed almost similarly to step S12.
Performs forward propagation processing. The structure of the neural network 20 in this embodiment is a three-layer structure similar to the first embodiment,
The output of the qth layer of the neural network 20 corresponding to the βth phoneme is calculated using the following equation (11A).

[0068]

[Math 6]

In equation (11A), opj (β, q)
is the output of the jth cell 21 in the qth layer of the neural network 20 corresponding to the βth phoneme, and is the output when the pth pattern is presented. w ji (β
, q) is the q-th neural network 20 corresponding to the β-th phoneme.
The weighting coefficient from the j-th cell 21 of the q-1th layer, which is the input to the q-th layer, and θj (β, q) is the q-th layer of the neural network 20 corresponding to the β-th phoneme. is the bias of the j-th cell 21 of . In this embodiment, the structure of each neural network 20 is the same, and the number of cells in each layer is N0 = 5.
12, N1 is 64, and N2 is 1.

The structure of each neural network 20, such as the number of layers and the number of cells, may differ for each phoneme. w ji (β, q) and θ
j (β, q) is set to a small random value before learning. Calculate equation (11A) for all β, q, j
are calculated in order to obtain the output opj (β, 2) of the cell 21 of the output layer, which is the second layer of each neural network 20.

[0071] In step S13A, the learning control section 40A performs an output error calculation process in substantially the same way as step S13. Prior to this, the teacher signal input unit 50A determines 0.9 as the teacher signal of the output cell 21 of the output cell 21 for the presentation signal, and sets 0.9 as the teacher signal of the output cell 21 of the neural network 20 corresponding to the phoneme φ of the presentation signal. 0.1 is determined as the teacher signal of the output cell 21 of the neural network 20 corresponding to the phoneme. That is, for the p-th presentation signal (phoneme φ), the teacher signal of the output cell 21 of the neural network 20 corresponding to the phoneme β is set as tp (β).

The learning control section 40A compares the teacher signal from the teacher signal input section 50A and the output of the neural network group 20A as follows, almost in the same way as in the first embodiment. j of the q-th layer of the neural network 20 corresponding to the β-th phoneme
If the cell internal error for the pth presentation signal in the cell 21 is expressed as δpj (β, q), then the cell internal error δpj (β, 2) in the output layer can be expressed as the following equation (13A):
Calculate with. In step S14A, the learning control unit 40A performs error backpropagation learning processing for the neural network group 20A, almost similarly to the first embodiment. When the internal error δpj (β, q) of each cell 21 in the qth layer of each neural network 20 has been calculated, the internal error δpj (β, q) of each cell 21 in the (q-1)th layer is calculated as , q-1) are calculated using the following equation (14A). Furthermore, using the error δpj (β, q) inside the cell, the weighting coefficient wji (β,
q) correction amount Δp wji (β, q) is calculated using the following formula (15A
) to calculate. Δp wji (β, q) = ηδpj
(β, q) opi (β, q-1)

...(
15A) Similar to the first embodiment, the following equation may be used to speed up learning. Δp wji (β, q) = αΔp−1 wji
(β, q) +ηδpj (β, q) opi (β,
q-1)

...(15A-1) In equation (15A-1), α and η are constants that determine the learning speed,
When formula (15A-1) is used, Δ0 wji (β, q) is set to 0 in initialization processing step S10A. Furthermore, the bias θj (β,
The correction amount Δp θj (β, q) for q) is also calculated using the following formula (
16A). Δp θj (β, q) = ηδ
pj (β, q)

(16A) Similar to the first embodiment, in order to speed up learning, the following formula (
16A-1) may also be used. Δp θj (β, q) = αΔp−1 θ
j (β, q) + ηδpj (β, q)

・
...(16A-1) However, α is a constant that determines the learning speed, and when using equation (16A-1), Δ0 θj (β, q)=
Set it to 0.

The calculations of equations (15-A) and (16-A) are performed for all neural networks 20, output layers, and intermediate layers while subtracting the layer number q, and all weighting coefficients wj
Correction amounts Δ pwji (β, q) and Δ pθj (β, q) for i (β, q) and bias θj (β, q)
). All correction amounts Δp wji (β, q)
After calculating Δpθj (β, q), all weighting coefficients wji (β, q) and biases θj (β, q) are modified using the following equation (17A) using this modification amount. wji (β, q) = wji (β
, q)+Δp wji (β, q)
θj (β, q) = θj (β, q) + Δp θ
j (β, q)

(17A) The above processing is performed on the presentation signal asφ0.9 (m) with p=1. At this time, as the teacher signal tp (β), one according to the above-mentioned formula (12A) is used.

In step S15A, the learning control section 40A performs a learning completion determination process, almost in the same way as in the first embodiment. By repeating the learning process, the weighting coefficient wji(β
, q) approaches the optimal value, the error inside the output cell δ
pj (β2) approaches 0. It is determined whether the errors δpj(β2) inside the cells of the outputs of all the neural networks 20 have become smaller than a sufficiently small value ε, and if the errors δpjβ(2) inside the cells of the outputs are large, the learning is not completed. It is determined that the process has ended, 1 is added to p, and the process returns to step S11A. Cell internal error δp of all neural network 20 outputs
If j(β2) is smaller than ε, all learning processing ends.

When the process returns to step S11A, the section 62 in FIG. 4, which is not centered at the aforementioned teacher epoch point, is taken as the presentation signal. The deviation of the center of the section 62 from the teacher epoch point is random. Presentation signal asφ0 in this case
．． 1 (m) is power-normalized using equation (18) and an offset is added to obtain the output opj (β0) of each cell 21 in the input layer.
shall be. The teacher signal tp (β) at this time is set to the same 0.1 for all output cells 21 of all neural networks 20, and the same processing as described above is performed from step S12A onwards.

As described above, almost similarly to the first embodiment, the learning process of alternately presenting asφ0.9 (m) and asφ0.1 (m) is applied to many phonemes φ and their presentation signals. By repeating this process, the optimal weighting coefficient can finally be obtained.

(IIA) Phonological recognition processing using the neural network group after learning In this phoneme recognition processing, input processing to the neural network group (step S
20A), forward propagation processing (step S21A), output peak detection processing (step S22A), and phoneme recognition determination processing (step S23A) are executed in substantially the same manner as in 1(b).

That is, in step S20A, the input speech waveform xu (m) is power-normalized as an input to all neural networks 20 using equation (19), and an offset is added to each of the input layers in each neural network 20. Let the output of the cell 21 be ouj (β, 0). Here, ouj (β, 0) indicates the output of the j-th cell 21 for the input speech waveform centered at time u in the q-th layer of the neural network 20 corresponding to the β-th phoneme.

In step S21A, forward propagation processing of the neural network group 20A is performed in substantially the same manner as in the first embodiment. This process is performed in the same way as the forward propagation process in the learning process by replacing p in equation (11A) with u. Through this processing, output ouj (β, 2) is obtained from the cells 21 of the output layer of each neural network 20. Furthermore, the input audio waveform is taken from a time interval centered at time u+1, and similar processing is performed. By repeating such processing, a sequence of outputs ouj (β, 2) of the neural network 20 corresponding to each phoneme at time u is obtained.

An example of this output series is as indicated by the reference numeral 76 in FIG. When the center of the time interval from which the input speech waveform is extracted matches the epoch point of a certain phoneme in the input speech waveform, the neural network 20 corresponding to the phoneme in the output series
A large peak occurs in the output series. By detecting this peak and comparing the magnitudes of the peaks of the output series of each neural network 20, the phoneme of the input speech waveform can be determined.

In step S22A, peak detection processing is performed on the output series of each neural network 20, almost similarly to the first embodiment. The output is the following formula (20A) and (21A)
A discrete time vd β that satisfies the condition is detected as the time of the epoch point of the phoneme.

[0082] Here, p is a threshold value for detecting a peak, and the first
A constant of 0.5 is used as in the embodiment. d is a number given to the detected peak, and β is the type of phoneme corresponding to the neural network 20.

[0083] In step S23A, the magnitudes of the peaks of the output series of each neural network at the epoch points of each detected phoneme are compared and used as a phoneme recognition result. In this example,
As in the first embodiment, the absolute magnitudes of the peaks are simply averaged and compared over a desired time interval (times Ts to Te), and the output having the largest average value of the peaks is matched. The resulting phoneme is output as the phoneme recognition result. To express this in a formula, first, the evaluation value Rβ of each phoneme is calculated by the following formula (
22A).

[0084]

[Math 7]

However, Nβ is the number of peak times vd β of the neural network 20 corresponding to the phoneme β, which are included from time Ts to Te. Then, the evaluation values Rβ are compared over all phonemes, and the phoneme corresponding to β that gives the maximum evaluation value is output as the phoneme recognition result.

The method of this phoneme recognition determination process is arbitrary as in the first embodiment, and various statistical distance measures may be used.

As described above, in this second embodiment, feature extraction of voiced sounds and phoneme recognition are carried out by the neural network 20 dedicated to each phoneme, which has the advantage of the first embodiment (
i) Recognition results with higher performance and time resolution can be obtained than in (ii).

[0088]

[Effects of the Invention] As explained in detail above, according to the first invention, feature extraction of voiced sounds and phoneme recognition are performed by a neural network that integrates them. A process that combines phoneme recognition is automatically created, resulting in high-performance phoneme recognition output. Furthermore, the basic period of speech and the phoneme recognition results can be obtained simultaneously, and recognition results with high temporal resolution can be obtained.

Therefore, when the present invention is used for continuous speech recognition, for example, there is no need to perform phoneme recognition and fundamental frequency extraction separately, and phoneme information and fundamental frequency can be obtained by a single process.

According to the second invention, feature extraction of voiced sounds and phoneme recognition are performed by a neural circuit dedicated to each phoneme, which is integrated, so phoneme recognition is achieved with higher performance and temporal resolution than the first invention. Get results.

[Brief explanation of the drawing]

FIG. 1 is a flowchart of a voiced phoneme recognition method showing a first embodiment of the present invention.

FIG. 2 is a flowchart of a conventional voiced phoneme recognition method.

FIG. 3 is a functional block diagram of a voiced phoneme recognition device in the first embodiment of the present invention.

FIG. 4 is an audio waveform diagram used as an instruction signal in the first embodiment of the present invention.

FIG. 5 is a diagram showing the relationship between an input audio signal and an output signal sequence in the first embodiment of the present invention.

FIG. 6 is a flowchart of a voiced phoneme recognition method showing a second embodiment of the present invention.

FIG. 7 is a functional block diagram of a voiced phoneme recognition device in a second embodiment of the present invention.

[Explanation of symbols]

10,10A waveform input section 20
Neural network 20A
Neural network group 21
Cells 30, 30A Recognition evaluation units 40, 40A Learning control unit 50
, 50A Teacher signal input section S10, S1
0A Initialization processing

Claims (2)

[Claims] Claim 1: A signal value obtained by normalizing the magnitude of a voiced phoneme signal over a certain time interval by power is input as a presentation signal for learning, and a predetermined position in the input time interval and an epoch point of the input signal are input. The output corresponding to the input phoneme takes a large value when the input phoneme matches, and takes a small value otherwise. The peaks of a plurality of output sequences of the neural network obtained by inputting a signal value whose magnitude is normalized by power over a certain time interval into the neural network and shifting the time of the input audio waveform to be applied. A voiced phoneme recognition method, characterized in that the voiced phoneme recognition method is characterized in that a phoneme recognition result is determined based on the peak detection result, and a phoneme corresponding to the output that produces the largest peak value is determined as the phoneme recognition result. 2. As a presentation signal for learning, a signal value obtained by normalizing the magnitude of a voiced phoneme signal over a certain time interval by power is input, and a predetermined position in the input time interval and an epoch point of the input signal are input. match and the input signal's phoneme is of the set phoneme type, the output will take a large value, and at other times it will take a small value.
Using a neural network group consisting of a plurality of neural networks subjected to learning processing using the error backpropagation method, the signal value obtained by normalizing the magnitude of an input speech waveform with unknown phoneme over a certain time interval using electric power is calculated as described above. The peak of each output sequence of the neural network group obtained by inputting the input speech waveform to the neural network group and moving the time of the added input speech waveform is detected, and the phoneme recognition judgment is performed based on the peak detection result to determine the most A voiced phoneme recognition method characterized in that a phoneme corresponding to the neural network that produces a large peak value is taken as a phoneme recognition result.