JP3183944B2 - Audio coding device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding apparatus,
In particular, the present invention relates to a speech encoding device using an analysis and synthesis method.

[0002]

2. Description of the Related Art A typical example of an analysis / synthesis method conventionally used in a speech coding apparatus is a linear prediction analysis. In this linear prediction analysis, assuming that a time discrete signal obtained by sampling a speech waveform at a fixed time interval is xt (t: integer representing time), a sample value xt at the present time and a past P There is a correlation between these sample values and a linear prediction value of xt is obtained from a linear equation expressed by the following equation (1).

[0003]

(Equation 1)

In the above equation (1), αi is a sample value xt
Is determined so as to minimize the sum of squares of the prediction error et between the and the linear prediction value. Incidentally, there are a covariance method and an autocorrelation method as a solution of αi. On the other hand, there is CELP (Code-Excited Linear Prediction) coding using the above linear prediction analysis.

FIG. 9 is a diagram showing an example of the configuration of an encoder employing the CELP coding. As shown in FIG. 9, a speech input from a speech input section 5 is supplied to a linear prediction analysis section 15 and The prediction coefficient αi is obtained. The linear prediction coefficient αi is scalar-quantized by the linear prediction coefficient quantization unit 16 and supplied to the subsequent linear predictor 17.

The linear predictor 17 also receives the excitation vector bj from the codebook 22 at the same time, calculates the difference between the input speech x and its linear prediction speech, and obtains a prediction error ej.
Is found. Then, the prediction error ej is passed through the auditory weighting filter 18 so that the auditory noise is reduced, and is supplied to the mean square error calculator 19 in the subsequent stage. This average 2
The square error calculator 19 calculates the mean square error of the prediction error e′j , and calculates the minimum mean square error and the excitation vector b at that time.
holds j.

[0007] After performed such processing on all the excitation vectors in the codebook 22, the audio decrypt circuit 21,
The quantized linear prediction coefficient αi and the excitation vector bj are transmitted.

[0008]

However, the above-described linear prediction analysis is a prediction analysis using a linear model, and the prediction error et is solved by the linear least squares method. is there.

Further, in the above linear model, it is assumed that the speech signal in a short section is stationary. However, since the speech signal is not stationary in practice, the linear prediction error does not become so small. Further, in the CELP coding, since only the excitation vector which minimizes the mean square error of the prediction error processed by the auditory weighting filter is searched from the codebook, the prediction error is calculated by the quantization error. Is not so small. SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and it is an object of the present invention to use a learning function of a neural network to prevent precision loss due to digit loss and quantization error.

[0010]

In order to achieve the above object, a speech coding apparatus according to the present invention comprises a linear prediction apparatus for calculating a linear prediction coefficient for an analysis order from input speech sampled at fixed time intervals. Analysis means, a linear prediction coefficient quantization means for quantizing the linear prediction coefficient calculated by the linear prediction analysis means, a signal quantized by the linear prediction coefficient quantization means, and information from a codebook. A linear prediction means for calculating a linear prediction speech by using a filter, a filter means for reducing a feeling of noise with respect to a linear prediction error which is a difference value between the input speech and the linear prediction speech, and an average of two output signals from the filter means. Means for calculating a squared error, a minimum mean squared error, a mean squared error calculating means for holding the excitation vector at that time, and a two-layered neural network comprising an input layer and an output layer A zero-state response which receives the excitation vector and the linear prediction coefficient, calculates a response value only by the excitation vector, and outputs a difference value between the input voice and the response value as teacher data of the hierarchical neural network means. Calculating means for learning and updating by using a linear prediction coefficient of the input speech calculated in advance as an initial value of the synaptic coupling coefficient of the hierarchical neural network means.

[0011]

That is, in the speech coding apparatus of the present invention, when linear prediction coefficients are calculated by the analysis order from the input speech sampled at predetermined time intervals by the linear prediction analysis means, the linear prediction coefficient quantization means , The linear prediction coefficient is quantized, and a linear prediction speech is calculated by the linear prediction means based on the quantized signal and information from the codebook. Then, the filter means reduces the noise sensation for the linear prediction error, which is the difference value between the input speech and the linear prediction speech, and the mean square error calculation means calculates the mean square from the output signal from the filter means. The error is calculated,
The minimum mean square error and the excitation vector at that time are retained. Then, the zero-state response calculating means receives the excitation vector and the linear prediction coefficient, calculates a response value based only on the excitation vector, and calculates the difference between the input voice and the response value of the two-layer hierarchical neural network means. Output as teacher data. The initial value of the synaptic coupling coefficient of the hierarchical neural network means is learned and updated using the linear prediction coefficient of the input speech calculated in advance.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The principle of the present invention will be described below. The linear prediction value can be represented by the above equation (1) from the linear prediction coefficient αi and the past sample value xt-i, and the prediction error et is represented by the equation (2).

[0013]

(Equation 2)

Here, considering a two-layer hierarchical linear neural network 1 as shown in FIG. 2, past sample values xt-i are used as input values to each neuron unit of the input layer 2,
The linear prediction coefficient αi can be regarded as the synaptic coupling coefficient between the input and output layers, and the linear prediction value can be regarded as the output value of each neuron net of the output layer 3.

Therefore, the current sample value xt is used as the teacher signal, and learning of the synaptic coupling coefficient, that is, the linear prediction coefficient αi is performed so as to minimize the sum of squares of the prediction error et.

In the hierarchical linear neural network 1 shown in FIG. 1, when past sample values xt-i are input to the input layer 2 for the order of the linear prediction analysis, the product of the product and the synapse coupling coefficient corresponding to the linear prediction coefficient is obtained. The sum is taken and a linear prediction is obtained. For learning, the error E can be determined as in equation (3).

[0017]

(Equation 3) Then, a method called back propagation learning as in the equation (4) is adopted.

[0018]

(Equation 4)

Next, FIG. 3 shows a feature of a speech in which a nonlinear neuron unit 4 is added between the input and output layers to the hierarchical linear neural network 1 shown in FIG. Is predictable. The nonlinear neuron unit 4 shown in FIG.
The product sum of the input value from the input layer 2 and the synaptic coupling coefficient is converted by a nonlinear function f (x) and output. At this time,
The output value Ytk of the non-linear neuron unit k is expressed as in equation (5).

[0020]

(Equation 5)

Note that f (x) = 1 / (1 + exp (-
x)), and the back propagation learning method is used as described above. P ′ is the number of synaptic connections between the neuron unit of the input layer and the nonlinear neuron unit. Hereinafter, embodiments of the present invention based on the above-described principle will be described. FIG. 4 is a diagram showing the configuration of the first embodiment of the present invention.

As shown in FIG. 1, in the encoder 20, the speech input unit 5 includes a two-layer hierarchical linear neural network 1.
The output layer 3 of the neural network 1 is connected to a synapse coupling coefficient learning unit 8. The input speech 5 further includes a linear prediction coefficient calculation unit 6 for obtaining a linear prediction coefficient for the analysis order from the input speech, a synapse coupling coefficient learning unit 8 for performing a learning operation of a synaptic coupling coefficient by back propagation learning, Each is connected to a prediction error calculator 10 for obtaining a prediction error et.

The linear predictive coefficient calculating section 6 and the synaptic coupling coefficient learning section 8 are connected to a synaptic coupling coefficient setting section 7, and the synaptic coupling coefficient setting section 7 is connected to the hierarchical linear neural network 1. I have. The hierarchical linear neural network 1 is also connected to a synapse coupling coefficient quantization unit 9 that quantizes the synaptic coupling coefficient, and the synapse coupling coefficient quantization unit 9 includes the prediction error calculation unit 10 and It is connected to an audio decoder 21 that synthesizes an audio waveform based on the quantized data of the synapse coupling coefficient and the prediction error of the input audio.

The prediction error calculator 10 is connected to a prediction error quantizer 11 for quantizing the prediction error. The prediction error quantizer 11 is connected to a speech decoder 21.

In such a configuration, when a predetermined number of voices sampled at predetermined time intervals are input to the linear prediction coefficient calculation unit 6 from the voice input unit 5, a covariance technique known in the art is used. The linear prediction coefficient is calculated for the analysis order by the method or the autocorrelation method. Usually, this analysis order P is 1
It is about 0. The result of this calculation is supplied to the synapse coupling coefficient setting unit 7, and is set as an initial value of the synaptic coupling coefficient αi of the hierarchical linear neural network 1.

When the initial value is set, the hierarchical linear neural network 1 is operated while sequentially inputting the input value xt-i by the analysis coefficient P, and the linear prediction value of the current speech waveform is converted to the synapse connection coefficient learning unit. 8 is output.

The synapse coupling coefficient learning unit 8 uses the linear prediction value, the synaptic coupling coefficient αi, the current sample value xt, and the input value xt-i to the input layer 2 to convert the synapse coupling coefficient αi to the back propagation learning. Is updated and learned. Then, the updated synapse coupling coefficient αi is supplied to the synapse coupling coefficient setting unit 7, and is set as a new synaptic coefficient of the hierarchical linear neural network 1.

The above-described back propagation learning is performed until the error E no longer decreases. However, only when the prediction error et is equal to or greater than a certain threshold, learning can be performed until the error falls within the threshold. Thus, the pitch as the sound source information is conventionally extracted from the prediction error, but this processing can be omitted.

Conversely, if the back propagation learning is performed only when the prediction error et is equal to or less than a certain threshold, the prediction error is pulsed, that is, power concentration occurs, and efficient coding can be performed.

Although the pitch component usually remains as a periodic impulse in the prediction error, the error can be effectively removed by threshold processing. Further, since the prediction error is set to be equal to or less than the threshold, the dynamic range is reduced, which contributes to the reduction of the code amount. When the back propagation learning is completed, the synapse coupling coefficient quantization unit 9
In, the synaptic coupling coefficient of the hierarchical linear neural network 1 is read and quantized with a predetermined number of quantization bits.

The prediction error calculator 10 calculates the prediction error et between the predicted value obtained from the quantized synaptic coupling coefficient and the current sample value xt, and the prediction error quantizer 11 calculates the prediction error et. Is quantized. In this way, the quantized data of the synapse coupling coefficient and the prediction error are supplied to the audio decoder 21 to perform the audio synthesis. Next, FIG. 5 is a diagram showing a configuration of a second embodiment of the present invention.

The present embodiment is characterized in that a random number generator 12 is provided in addition to the configuration of the first embodiment, and a nonlinear neuron unit 4 is provided in the hierarchical linear neural network 1 between input and output layers. Having.

In such a configuration, the synapse connection coefficient setting unit 7 receives the initial value of the synapse connection coefficient αi from the linear prediction coefficient calculation unit 6 and simultaneously receives the synapse connection between the input layer and the nonlinear neuron unit from the random number generation unit 12. When a small random number is received as the initial value of the coefficient βik and the initial value of the synaptic coupling coefficient γk between the nonlinear neuron unit and the output layer, the values are set in the hierarchical neural network 1 ′. Then, when the initial value is set, the same processing as in the first embodiment is performed, but the predicted value of the current voice form is expressed by equation (6).

[0034]

(Equation 6) Here, K indicates the number of nonlinear neuron units, J indicates the number of synaptic connections between the input layer and the nonlinear neuron unit, and has a relationship of P ≧ J. In the present embodiment, by adding the nonlinear neuron unit 4, nonlinear prediction of the speech waveform becomes possible, and the prediction error can be further reduced.

In order to prevent the linear prediction coefficient αi from largely changing in the initial stage of learning, αi is fixed in the initial stage of learning, and the coefficients βik, γk
A method is also possible in which only updating is performed, and in the next stage, all synaptic coupling coefficients are learned and updated.

The embodiment in the case where the present invention is applied to the linear prediction analysis has been described above. Next, the embodiment in the case where the invention is applied to the CELP coding using the linear prediction analysis will be described. With reference to FIG.
An outline of a speech encoding device incorporating LP encoding will be described.

As shown in the figure, a zero-state response calculating unit 13 is connected to the encoder 20. The zero-state response calculating unit 13 and the voice input unit 5 are connected via a differentiator 14 to a hierarchical neural network. 1 connected. And
The encoder 20 further includes a hierarchical neural network 1
The hierarchical neural network 1 is connected to a decoder 21.

In such a configuration, the optimum excitation vector bj output from the encoder 20 is supplied to the zero-state response calculator 13, where the zero-state response St is calculated and output. Then, the zero-state response St can be expressed as in equation (7) using the linear prediction coefficient αi and the excitation vector bj as in the case of the linear predictor.

[0039]

(Equation 7)

However, this is different from the linear predictor in that the values of the initial state St-i at the time of calculation are all zero. Then, in the differentiator 14, a difference x ′ (= x−S) between the input voice x and the zero-state response S of the excitation vector bj is obtained and supplied to the hierarchical neural network 1.

The hierarchical neural network 1 is a two-layer linear neural network having an input layer 2 and an output layer 3. The input layer 2 and the output layer 3 are connected to each other by synaptic connection. Then, the linear prediction coefficient αi obtained by the encoder 20 is used as the initial value of the synaptic coupling coefficient of the hierarchical neural network 1.

When the past output value xt-i is input to the input layer 2 of the hierarchical neural network 1, for example, an error E
Is calculated by the equation (8), and a back propagation learning method as shown in the above equation (4) is performed so as to minimize the error E.

[0043]

(Equation 8)

In the above equation (8), the first term is a normal output error minimizing term using the output value x 'from the differentiator 14 as teacher data, while the second term is a linear prediction coefficient αi whose quantum is The formula is such that the value becomes smaller if it is closer to any element vim in the conversion table vi. Here, ε is a positive constant close to “0”.

In the back propagation learning, a sequential learning method of updating the synaptic coupling coefficient for each audio signal x't is also possible, but here, collective learning for updating the synaptic coupling coefficient collectively for each analysis frame section T is used. Each time the synaptic coupling coefficient is updated using the method, the linear prediction coefficient αi of the zero-state response calculation unit is updated with the synaptic coupling coefficient αi of the hierarchical neural network 1.

This operation is repeated until the error E becomes sufficiently small, and the synapse coupling coefficient αi is quantized and output as a more optimal linear prediction coefficient. FIG. 1 is a diagram showing a configuration of a third embodiment of the present invention.

As shown in the figure, the speech input unit 5 is connected to a linear prediction analysis unit 15,
Are connected to the linear prediction coefficient quantization unit 16. The linear prediction coefficient quantization unit 16 is connected to a linear predictor 17, and the linear predictor 17 has a codebook 2
2 is connected to a gain adder 23 for giving a gain γ to the excitation vector bj obtained from the step 2.

Further, the speech input unit 5 and the linear predictor 17 are connected to an auditory weighting filter via a differentiator 14a, and the auditory weighting filter 18 is connected to a mean square error calculator 19. I have. The mean square error calculator 19 is connected to the synapse coupling coefficient setting unit 7 and the zero state response calculator 13.

The zero-state response calculating section 13 and the voice input section 5 are connected to the synapse coupling coefficient learning section 8 via a differentiator 14b.
The synapse coupling coefficient learning unit 8 is connected to the synapse coupling coefficient setting unit 7.

The synapse coupling coefficient setting section 7 is connected to the hierarchical neural network 1, and the hierarchical neural network 1 is connected to the synaptic coupling coefficient learning section 8 and the synaptic coupling coefficient quantizing section 9, respectively. ing. The synapse coupling coefficient quantization unit 9 is connected to a speech decoder 21, and the speech decoder 21 is connected to a mean square error calculation unit 19.

In such a configuration, when a predetermined number of input voices sampled at a fixed time interval are input to the linear prediction analysis unit 15, the input voice is linearized by a known technique such as a covariance method or an autocorrelation method. The prediction coefficients are calculated for the order of analysis. Usually, the analysis order P is about 10. The calculation result is supplied to the linear prediction coefficient quantization unit 16, scalar-quantized with reference to a quantization table (not shown), and supplied to the linear predictor 17. Further, at the same time, the excitation vector bj from the codebook 22 is multiplied by γ in the gain adder 23 and supplied to the linear predictor 17 to obtain a linear predicted speech.

Next, the difference value between the input speech and the linear prediction speech,
In other words, the linear prediction error ej is supplied to the auditory weighting filter 18, and the noise perception is reduced based on the human auditory characteristics. The mean square error of the filter output is calculated by a mean square error calculator 19, and the minimum mean square error is calculated.
The multiplication error and the excitation vector γbj at that time are held.

This operation is performed for all the excitation vectors in the code book 22, and the minimum error excitation vector γbj and the linear prediction coefficient αi that are the results are supplied to the zero-state response calculator 13.

Here, a response value based on only the excitation vector γbj, that is, a zero-state response is calculated, and the difference value x ′ between the input speech x and the zero-state response S is used as the teacher data of the hierarchical neural network 1 for synapse coupling coefficient learning. It is supplied to the unit 8.

The mean square error calculator 19 is used as an initial value of the synaptic coupling coefficient of the hierarchical neural network 1.
Are set via the synapse coupling coefficient setting unit 7.

Then, the hierarchical neural network 1
Is performed by the synapse coupling coefficient learning unit 8 while operating based on the above equation (1). The error equation to be minimized by the back propagation learning is, for example, the above equation (8). Defined in This is to minimize the error represented by the following equation (9) while bringing the linear prediction coefficient αi closer to Vim, one of the elements of a linear prediction coefficient quantization table (not shown).

[0057]

(Equation 9)

That is, the scalar quantization of the linear prediction coefficient and the minimization of the output error are simultaneously optimized. The back propagation learning here uses a collective learning method of updating the synapse connection coefficient for each analysis frame section T, and updates the linear prediction coefficient αi of the zero-state response calculation unit 13 for each update.

Further, after repeating the learning of the hierarchical neural network 1 via the synapse coupling coefficient setting unit 7 until the error E becomes sufficiently small, the synaptic coupling coefficient is scalar-quantized by the synaptic coupling coefficient quantization unit 9. ,
Output to the audio decoder 21. The speech decoder 21 receives the optimum excitation vector γbj from the mean square error calculator 19 at the same time and synthesizes the speech. Next, FIG. 7 is a diagram showing a configuration of a fourth embodiment of the present invention.

As shown in the figure, in the present embodiment, the zero-state response calculation unit 13 is eliminated from the configuration of the first embodiment, and the input unit of the excitation vector bjt is added to the hierarchical neural network 1 instead. And the gain γ is initially set as the synaptic coupling coefficient.

In such a configuration, the gain γ of the excitation vector bj sent from the mean square error calculator 19
Are initially set in the hierarchical neural network 1 via the synapse coupling coefficient setting unit 7.

Then, the learning operation is started when the element bjt at the time point t of the excitation vector bj is input to the hierarchical neural network 1. The gain γ is learned so as to approximate a quantization table or a quantization step (not shown), like the linear prediction coefficient α. That is, equation (10) is added to the equation indicating the error E.

[0063]

(Equation 10) Here, Un is one element of the quantization table U of gain γ, and n is the number of elements in the table.

In this way, the speech encoder 21 receives the excitation vector bj from the mean square error calculator 19 and the linear prediction coefficient αi optimized by the synapse coupling coefficient quantizer 9 and the gain γ of the excitation vector, and synthesizes the speech. Do. FIG. 8 is a diagram showing the configuration of the fifth embodiment of the present invention.

The present embodiment is different from the conventional example shown in FIG. 9 in that the zero-state response calculator 13 is provided so as to feed back the quantization error caused by the codebook 22 to the linear prediction analyzer 15. It has features.

In such a configuration, when the optimum excitation vector γbj is obtained by the mean square error calculation section 19, it is sent to the zero-state response calculation section 13 and the optimum excitation vector γbj is obtained.
Is calculated, and the difference value x from the input voice x is calculated.
, A new linear prediction coefficient αi is obtained in the linear prediction analysis unit 15.

It is possible to immediately send the quantized linear prediction coefficient to the speech encoder 21. However, in order to further improve the encoding accuracy, the optimum excitation vector is determined again. Then, the above processing is repeated until the quantized data of the linear prediction coefficient does not change.
In the present embodiment, such an operation makes it possible to optimize both the linear prediction coefficient and the excitation vector. The embodiments of the present invention have been described above. However, it is needless to say that the present invention is not limited to the embodiments, and various modifications and changes can be made.

For example, the hierarchical neural network 1 used in the third and fourth embodiments is a two-layer linear network, but a nonlinear neural network can be added between the input and output layers.

[0069]

According to the present invention, it is possible to prevent a decrease in precision of a linear prediction coefficient due to a digit loss in a conventional numerical calculation by a learning process of a hierarchical neural network. Further, by adding a non-linear neuron unit, non-linear prediction of an unsteady speech waveform can be performed, and the prediction error can be further reduced.

Since the linear prediction coefficients are optimized by learning the hierarchical neural network using the linear prediction error and the quantization error of the input speech, the input speech can be encoded with high efficiency. .

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a speech encoding device according to a third embodiment of the present invention.

FIG. 2 is a diagram showing a configuration of a hierarchical linear neural network 1.

FIG. 3 is a diagram showing a configuration of a hierarchical nonlinear neural network 1 ′.

FIG. 4 is a diagram illustrating a configuration of a speech encoding device according to a first embodiment of the present invention.

FIG. 5 is a diagram illustrating a configuration of a speech encoding device according to a second embodiment of the present invention.

FIG. 6 is a conceptual diagram when the present invention is applied to CELP coding.

FIG. 7 is a diagram illustrating a configuration of a speech encoding device according to a fourth embodiment of the present invention.

FIG. 8 is a diagram illustrating a configuration of a speech encoding device according to a fifth embodiment of the present invention.

FIG. 9 is a diagram illustrating a configuration of a conventional speech encoding device.

[Explanation of symbols]

1 ... Hierarchical neural network, 2 ... Input layer, 3 ...
Output layer, 4 ... Intermediate layer, 5 ... Speech input section, 6 ... Linear prediction coefficient calculation section, 7 ... Synapse connection coefficient setting section, 8 ... Synapse connection coefficient learning section, 9 ... Synapse connection coefficient quantization section, 10
... Prediction error calculator, 11 ... Prediction error quantizer, 12 ... Random number generator, 13 ... Zero state response calculator, 14 ... Differentiator,
15: Linear prediction analysis unit, 16: Linear prediction coefficient quantization unit,
17: Linear predictor, 18: Auditory weighting filter, 19
... mean-square error calculator, 20 ... encoder, 21 ... decoder,
22: code book, 23: gain adder.

Claims (1)

(57) [Claims] 1. A linear prediction analysis means for calculating a linear prediction coefficient by an analysis order from input speech sampled at a fixed time interval, and a linear prediction means for quantizing the linear prediction coefficient calculated by the linear prediction analysis means. Prediction coefficient quantization means; linear prediction means for calculating a linear prediction speech based on a signal quantized by the linear prediction coefficient quantization means and information from a codebook; difference between the input speech and the linear prediction speech Filter means for reducing noise perception of a linear prediction error, which is a value; calculating an average square error from an output signal from the filter means; and calculating an average square error and an average holding an excitation vector at that time Square error calculating means, two-layer hierarchical neural network means comprising an input layer and an output layer, receiving the excitation vector and the linear prediction coefficient, Zero-state response calculating means for calculating a response value by only the input voice and a difference value between the input voice and the response value as teacher data of the hierarchical neural network means, and a synaptic connection of the hierarchical neural network means. A speech encoding apparatus characterized in that learning and updating are performed using a linear prediction coefficient of an input speech calculated in advance as an initial value of a coefficient.