JP2537113B2 - Adaptive compression method of vocal tract parameter information in speech coder / decoder - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to voice coding for digitizing a vocal tract signal for transmission or storage, and for converting a transmitted or stored digital signal into an analog signal, such as a telephone, a mobile phone, an automobile. It can be applied to telephone devices such as telephones, voice files, and voice memories.

[0002]

2. Description of the Related Art Conventionally, various techniques have been developed for digitizing an analog voice signal to compress information for efficient transmission or storage.

An example of an encoder and a decoder is shown in FIGS. 4a and 4a.
Shown in b. Assuming that there are two types of sound sources, noise and periodic components, the vocal tract parameters of voice signals, voiced sound / unvoiced sound determination, and periodicity measurement in the case of voiced sound are measured at fixed intervals of the input voice. , Vocal tract parameters, voiced / unvoiced decision results, periodic components are transmitted, and based on the transmitted parameters,
An example is a vocoder that reproduces voice by inputting a signal of that period in the case of voiced sound and random noise in the case of unvoiced sound to a vocal tract filter configured by the measured vocal tract parameters.

An example of the encoder 1 is shown in FIG. 4a. Figure 4a
In the encoder 1, the input voice is digitized by the AD converter 2 in accordance with the sampling clock of the sampling clock generator 3, and this signal string is converted into the vocal tract parameter measuring unit 5 for each section of the frame clock of the frame clock generator 4. The voiced / unvoiced detection unit 6, the pitch period detection unit 7, and the amplitude measurement unit 8 are connected to measure each parameter and perform encoding. Further, as described later in detail, the vocal tract parameter measuring unit 5 can be connected to the vector quantizing unit 9 to perform vector quantization of the vocal tract parameters.

The operation of the encoder 1 will be described. The input voice is digitized by the AD converter 2 into 8 bits according to the sampling clock of the sampling clock generator 3, for example, a clock of 8 kHz. If this digital signal is transmitted as it is, the amount of transmission becomes 8 bit × 8000 = 64 kbps.

Here, for the digitized voice signal, the vocal tract parameter measuring unit 5 makes a tenth-order prediction every 20 msec according to the frame clock of the frame clock generator 4 to obtain a vocal tract parameter, and voice / voice The unvoiced detection unit 6 parameterizes the determination of voiced sound or unvoiced sound with 1 bit, the pitch period is parameterized with the pitch period detection unit 7 with 8 bits, and the input level is measured with the amplitude measurement unit 8. Parameterize with 8 bits.

The vocal tract parameter measuring unit 5 has been devised in various ways, but it can be obtained by, for example, linear prediction. Similarly, various methods have been devised for voiced / unvoiced determination and pitch period. For example, in the former case, the number of zero-crossings of the input speech signal between frames is used, and in the latter case, the self-statement of the input speech signal between frames is performed. It can be obtained from the correlation.

The transmission rate in the case of this example is 10th order vocal tract parameter × 8 bits voiced / unvoiced parameter 1 bit pitch parameter 8 bits amplitude parameter 8 bits, so that (10 × 8 + 1 + 8 + 8) × 1000/20 = 4850 bps. Therefore, the data is compressed to about 1/16 as compared with the case where the voice is transmitted only by the AD conversion.

An example of the decoder 10 is shown in FIG. 4b. FIG.
The decoder 10 shown in FIG. 2b uses a pulse train generation unit 11 that generates a period matching the pitch period parameter sent from the encoder 1, a random noise generation unit 12, and a voiced / unvoiced parameter that is transmitted, If the voice is voiced, the pulse train generated by the pulse train generator 11 is selected. If the voice is unvoiced, the noise generated by the random noise generator 12 is selected.
After the signal selected in 1 is calculated by the vocal tract filter unit 14 controlled by the vocal tract parameter from the encoder 1,
The DA converter 15, which controls the reference signal with the transmitted amplitude parameter, demodulates it into an analog audio signal while restoring the original amplitude.

The operation of the above decoder 10 will be described below. From the encoder 1, as described above, the vocal tract parameter, the pitch period parameter, the voiced / unvoiced parameter,
Four types of parameters of the amplitude parameter are transmitted.

In FIG. 4b, these parameters are the vocal tract filter section 14, the pulse train generating section 11, the random noise generating section 12, the changeover switch 13 and the DA converter 1.
Connected to 5, according to the parameters for each frame,
In the case of voiced sound, a pulse train having the pitch period at that time is input to the vocal tract filter unit 14, while in the case of unvoiced sound, random noise generated by the random noise generation unit 12 is input to the vocal tract filter unit 14. And after the calculation,
The DA converter 15 which controls the reference signal by the amplitude parameter converts it into an analog audio signal. In such a speech coder 1, it is known that it is effective to perform vector quantization in order to lower the bit rate with respect to vocal tract parameters, for example.

FIG. 4c shows an example of the structure of the relevant parts when performing vector quantization of vocal tract parameters. The input speech becomes, for example, a tenth-order parameter as a result of linear prediction in the vocal tract parameter measuring unit 5 in the encoder 1, and this value is distorted by the vector quantizing unit 9 from the distortion calculation with the representative vector. Is selected and the code will be transmitted. In other words, in vector quantization, several vectors are prepared in advance with the combination of tenth-order parameters as one vector, the vector closest to the input combination of tenth-order parameters is calculated, and the number is transmitted. To do. For example, if a 1,248-bit vector is prepared for the tenth-order vocal tract parameter, the transmission amount is compressed to 1/10 as compared with the case where the tenth-order vocal tract parameter is directly quantized. it can.

In the decoder 10, the information-compressed data is transmitted as the address of the vector, so that the vector decoding unit 16 restores it to the tenth order parameter.

[0014]

In the compression means for encoding the vocal tract parameters by vector quantization, an average value is prepared in advance as a vector, and the average value for the vocal tract parameters of the input speech signal is set. Although a vocal tract parameter with a close average value is applied, this method has a problem that the personal voice of the voice is impaired and the sound quality is deteriorated.

[0015]

SUMMARY OF THE INVENTION The present invention has been proposed to solve the above-mentioned problems. Two neural networks are provided for compression of vocal tract parameters, and one neural network has an average weight. For the other neural network, the average value is initially set, but learning is performed using the vocal tract parameter obtained from the input speech as the teacher signal, and the weight is applied in an appropriate interval where the speech is interrupted. Is changed or the neural network is replaced to optimize the speech quality.

[0016]

Therefore, according to the present invention, by preparing a two-circuit neural network, the vocal tract parameters are initially compressed by one neural network while the other neural network performs learning based on the input parameters. If there is a certain amount of difference between the average load value and the trained load, change the load or switch the neural network at an appropriate timing such as voice interruption. ,
The vocal tract parameters can be optimized to improve the voice quality.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the configuration of an embodiment of the present invention. In the encoder 17 shown in FIG. 1a, the input speech is AD
The converter 18 digitizes the signal according to the sampling clock of the sampling clock generator 19, and the signal is converted into the frame clock generator 20.
Vocal tract parameter measuring unit 21, voiced / unvoiced detection unit 22, pitch period detection unit 23,
The output of the vocal tract parameter measuring unit 21 is connected to the amplitude measuring unit 24, and the output of the vocal tract parameter measuring unit 21 is a fixed weight neural network 25.
And the learning type neural network 26.

The load comparing section 27 compares the weight of the fixed weight type neural network 25 with the weight of the learning value of the learning type neural network 26, and outputs a result to the load control section 28 when a certain difference occurs.

In the load control unit 28, the load comparison result by the load comparison unit 27 and the voiced / unvoiced detection unit 22
A signal is sent to the switch (1) 29 and the counter 30 based on the unvoiced determination result and the measurement result of the amplitude measuring unit 24, and the switch (2) 31 receives the result from the counter 30 and outputs the signal from the pitch period detecting unit 23. The output or the weight of the learning neural network 26 is selected and used as the output.

The counter 30 counts the number of outputs of the load controller 28, and when the number of counts exceeds 40, sends a control signal to the switch (2) 31.

The switch (2) 31 is based on the signal from the counter 30 and has a fixed weight type neural network 2.
5 is a switching device that sends either the vocal tract parameter compressed by 5 or the vocal tract parameter compressed by the learning neural network 26 to the decoder as a vocal tract parameter.

The load switching unit 32 receives the load control signal from the load control unit 28 and the output of the counter signal from the counter 30, and creates a signal for switching the load to the decoder 33 (FIG. 1b). .

The operation of the encoder 17 will be described.
The input voice is sent from the sampling clock generator 19 to the AD converter 18 at a frequency of 8 kHz.
Digitized into bits. The digitized audio signal is usually output by the frame clock generator 20.
Vocal tract parameter measurement unit 2 in the msec frame section
1, the voiced / unvoiced detection unit 22, the pitch period detection unit 23, and the amplitude measurement unit 24 measure respective parameters.

Regarding the vocal tract parameters, the weight-fixed neural network 25 further compresses the amount of information with a network having a weight corresponding to the average characteristics of the voice.

Here, the principle of information compression by the neural network will be briefly described. FIG. 3a shows a three-layer hourglass-type neural network. For simplicity, the input layer has four units 1-1 to 1-4, the intermediate layer has two units 2-1 to 2-2, and the output layer has four units. 3-1 to 3-4
4 units will be described.

The information x1 to x4 input to the input layer is connected to the intermediate layer via the load wij, and its output is returned to the original at the output layers 3-1 to 3-4 via the load wik. j is the number of the input layer in which the input layer 1-1 is i = 1 and 1-4i = 4, and i is similarly the intermediate layer 2-1 and j = 1 and 2-2 is j.
= 2. Here, when the neural network is divided into two in the middle layer, the result is as shown in FIG. 3b. That is, the input information is compressed into two units in the intermediate layer, and this signal can be transmitted and the receiving side can restore the original information.

An example of the configuration of the fixed weight neural network 25 of FIG. 1a is shown in FIG. 2a. In FIG. 2a,
α1 to α10 represent tenth-order vocal tract parameters measured by the vocal tract parameter measuring unit 21 of FIG. 1a. The ten parameters are the ten neural network input layers 1-1.
Is input to 1-10. From each input layer, the four intermediate layers 2-1 to 2-4 are connected with weights wij, and are connected to the output of the neural network. That is, FIG.
This means that 10 parameters have been compressed to 4.

The parameters measured by the vocal tract parameter measuring unit 21 are also input to the learning type neural network 26. In the learning type neural network 26,
The input vocal tract parameter is fetched as a teacher signal of, for example, about 1000, and a calculation for obtaining a weight is performed based on this value. If one frame is 20 ms, 1000 frames correspond to a voice signal for about 20 seconds. This is a voice segment that can be expected sufficiently if continuity is not questioned in one conversation.

In this way, after the learned weight is obtained, when a call is made using this device, if there is a difference between the weight provided in the fixed-weight neural network 25 and the learned weight, there is a silent section or a silent section. It is possible to optimize the load by transmitting the load data in the 8-bit portion of the pitch information and switching the load after transmitting all of the data.

Regarding the transmission of the learned weights, the operation of the encoder 17 shown in FIG. 1a will be described. First, the weight comparison unit 27 compares the learned load value and the average load value, and when the comparison value exceeds a certain value, a control signal is applied to the load control unit 28.

The load control unit 28 receives the load control signal from the load control unit 27, the voiced / unvoiced detection unit 22 outputs a voiceless result, or the amplitude detection unit 24 outputs a detection result of a predetermined value or less. Then, the learned load is transmitted to the decoder 33 by the switching operation of the switch (1) 29.

The counter 30 determines whether or not the learned load has been transmitted a fixed number of times, in the case of the above-mentioned example, all 40 have been transmitted. When all the transmissions are completed, the switch (2) 31 is turned on. Control and change the weight of the fixed weight neural network 25 to the learned weight, and
The load on the decoder 33 side is also changed.

As a result, the encoder 17 of FIG.
Up to learning, vocal tract parameters, voiced / unvoiced parameters, pitch period parameters, and amplitude parameters
Parameters of various types will be transmitted, and each information amount will be vocal tract parameter 4 × 2 bits Voiced / unvoiced parameter 1 bit Pitch period parameter 8 bits Amplitude parameter 8 bits, according to the vector quantization shown in the conventional example The control parameter can be set as 2 bits.

The load switching unit 32 receives the load control signal from the load control unit 28 and the output from the counter 30, and initially uses the average load, but transmits using the pitch cycle slot of the adaptive load, and the timing of load switching. Send a signal.

After learning, if there is an effect of learning, the weight of the pitch period parameter is transmitted in the unvoiced sound or the silent section, and the weight is switched by the control signal when all the weights are transmitted. .

FIG. 1b shows an embodiment of the decoder 33.
In the decoder 33 shown in FIG. 1B, the control signal from the encoder 17 is decoded by the decoder 34 for the weight control signal, and this control signal controls the pitch / weight changeover switch 35 to set the pitch period and the adaptive weight. Switch parameters.

The vocal tract parameter is sent to the neural network decoding unit 36, and the weight of the neural network decoding unit 36 is connected to two types of weights, an average weight 37 and an adaptive weight 38. The two loads are selected by the switch (3) 39.

The adaptive load 38 receives the output of the pitch / load changeover switch 35 which is controlled by the load control decoder 34.

The pulse train generator 40 receives the pitch period signal from the pitch / load changeover switch 35.

The switch (4) 41 is a switch controlled by a voiced / unvoiced parameter, and in the case of voiced sound, the pulse train from the pulse train generator 40, and in the case of unvoiced sound, the random signal from the random noise generator 42. Select. Then, the selected signal is input to the vocal tract filter 43, and the output of the vocal tract filter 43 is converted into an analog signal by the DA converter 44 in which the reference signal is controlled by the amplitude parameter, and becomes a voice output.

The decoder 33 also has a clock generator 45 for supplying a clock.

Neural network decoding unit 3 of FIG. 1b
The basic configuration of 6 is shown in FIG. 2b. Here, β1 to β4 are compressed vocal tract parameters transmitted from the encoder 17, and these values are decoded into 10 vocal tract parameters by the weight w. Then, this load w is adaptively controlled by voice input.

The neural network decoding unit 36 includes
There are two loads, an average load 37 and an adaptive load 38. The average load 37 is an average load and the adaptive load 38 is a learned load. The adaptive load 38 is controlled by the control parameter.
In the case of the load update signal, the signal of the pitch parameter slot is received as the load, and the switch (3) 39 is controlled by this control signal to switch the load.

The operation of the decoder 33 shown in FIG. 1b is to decode the characteristics of the signal of the pitch period or the parameter slot of the adaptive weight by the weight control signal sent from the encoder 17, and if it is the pitch period, the pulse train. The pulse train generated by the generation unit 42 is controlled, and if it is a parameter of the adaptive load, the value of the adaptive load 38 is sequentially updated, and when all the updates are completed, the load is updated from the average value to the adaptive value.

The vocal tract parameters are decoded by the neural network decoding unit 36, and the vocal tract filter 43 is created using these parameters. Then, the voiced / unvoiced parameter is used to determine voiced / unvoiced, and if unvoiced, the random noise generated by the random noise generator 42 is sent to the vocal tract filter 43 if the voice train is the pulse train generated by the pulse train generator 40. input.

The vocal tract filter 43 communicates the reference signal to the DA converter 44 controlled by the amplitude parameter, and the DA converter 44 reproduces a voice signal having a size corresponding to the voice amplitude input to the encoder 17. .

A comparison between the adapted vocal tract parameters using the neural network according to the present invention and the conventional vocal tract parameters will be described below. In principle, the vector quantization given as a conventional example can also be adapted by providing a plurality of vector quantizers.

However, considering the amount of transmission in the case of adaptation, the method using the neural network according to the present invention, the transmission in the present invention from the viewpoint of changing the value of the load and the control signal. The amount is 40 × 2 × 8 = 640 bits, while the transmission amount in the conventional vector quantization is 1024 × 10 × 8 = 81920 bits.

Further, as shown in the present invention, considering that the weight data which has been learned is transmitted to the decoder 33,
In the case of changing the weight of the neural network, in the present invention, 640/8 = 80 frames, whereas in the conventional vector quantization, 81920/8 = 10240 frames.

Assuming that the timing of data change is the silent section and the unvoiced section, the silent section averages 30% of the voice,
Since the unvoiced section is considered to be 20%, the weight change in the neural network is 80 × 0.02 × 2 = 3.2 seconds, and the vector change amount of vector quantization is 10240 × 0.02 × 2. = 409.6 seconds, which means that the data can be changed only when the conversation has continued for a considerably long time, and cannot be put to practical use.

In this case, the weight learning is started from the average weight.
It is also possible to set several types of loads in advance, such as a method of selecting one of the types of loads. Further, in the embodiment, one set of neural networks is used for learning, but a method of providing two to five sets of neural networks such as a home telephone and switching the neural networks by the user is also conceivable.

[0052]

As described above, according to the present invention,
Compared to the compression of information by vector quantization shown in the above conventional example, it is possible to transmit individual characteristics by performing optimal code compression according to the characteristics of the voice input to the encoder. It is possible to improve the voice quality.

[Brief description of drawings]

FIG. 1a is a block diagram of a speech coder according to the present invention. 1b is a block diagram of a speech decoder according to the present invention.

2a is an explanatory diagram of a basic configuration of a coding unit of the neural network according to the present invention. FIG. 2b An explanatory diagram of the basic configuration of the decoding unit of the neural network according to the present invention.

FIG. 3a is a principle diagram of code compression in the neural network according to the present invention. 3b is a principle diagram of code compression in the neural network according to the present invention.

FIG. 4a is a block diagram of a conventional encoder. 4b A block diagram of a conventional decoder. 4c is a block diagram of a related part when vector quantization of a conventional vocal tract parameter is performed.

[Explanation of symbols]

 25 fixed weight type neural network 26 learning type neural network 27 load comparison section 28 load control section 29 switch 1 30 counter 31 switch 2 32 load switching section

Claims (1)

(57) [Claims] 1. When compressing vocal tract parameters in speech coding, two sets of neural networks are used, one neural network performs average weighting, and the other neural network inputs parameters to the neural network. Based on the average weighted neural network,
A method for compressing vocal tract parameter information, which has a function of detecting a section that does not affect conversation and switching to a learned neural network with weighting.