JP2970254B2 - Speech synthesis method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an ADPCM (Adaptive Differe) which is a kind of waveform encoding system.
neutral Pulse Code Modulati
The present invention relates to a method and an apparatus for reducing the difference value data memory by reducing the redundancy of the difference value data even if the difference value data is entirely provided as a read-only memory (ROM).

[0002]

2. Description of the Related Art First, an ADPCM (Adaptive Dif) performed in a conventional speech synthesizer.
Fermental Pulse Code Modu
With reference to the drawings, a description will be given of a table referencing method for performing a combining process without using a multiplier.

FIG. 5 shows an ADP by a conventional table reference method.
FIG. 6 is a block diagram of a CM speech synthesizer, FIG. 6 is an explanatory diagram of a quantization table, FIG. 7 is an explanatory diagram of a 3-bit PCM waveform, FIG.
FIG. 8A is an explanatory diagram of a 4-bit PCM waveform, and FIG.
FIG. 9 is an explanatory diagram of a 5-bit PCM waveform, FIG. 9 is an explanatory diagram of a DPCM system, FIG. 10 is an explanatory diagram of an ADPCM waveform, and FIG.
FIG. 1 is an explanatory diagram of a quantization coefficient k of a bit ADPCM method,
2 is an explanatory diagram of the quantization coefficient k at the time of 2 to 5 bits.

When a voice signal is handled by a digital code,
First, as shown in FIG. 7, sampling is performed at a certain cycle,
A value obtained by quantizing each sampled value and replacing it with a code corresponding to the quantized value is PCM (Pulse C
Ode Modulation). Generally A
The output value of the / D converter is a PCM code. The hatched portion in FIG. 7 indicates an error due to quantization, which is called a quantization error. This encoding method attempts to faithfully attempt to encode an analog signal, as shown in FIG.
As shown in (b), a large number of bits are required and the amount of information increases.

[0005] The DPCM (Dif) described below has been devised to improve the sound quality and reduce the amount of information in the PCM system.
Fermental Pulse Code Modu
lation) method.

In the PCM system, audio bits are encoded by allocating all bits. In this method, the number of code bits must be increased in order to synthesize high quality speech. Therefore, by utilizing the correlation of the audio signal, the difference between the previous audio signal and the PC is calculated as shown in FIG.
The result of the M encoding is the DPCM.

In other words, if the number of bits is the same, higher quality speech can be synthesized by assigning the difference value. Conversely, if the sound quality is the same, the DPCM code bits can be reduced and the information of the audio data can be compressed.

[0008] In the DPCM system, since the PCM is used for the difference value, the quantization width for the quantization is constant, and the amplitude of the audio signal changes suddenly. And appear as noise.

Further, a speech signal which is more efficiently coded by introducing the concept of adaptive coding into the DPCM system is referred to as an ADPCM (Adaptive D) described below.
iferential Pulse Code Mo
(duration) method.

[0010] ADPCM scheme is a method of determining the quantization width delta _i of the difference signal value from the results of the previous sample value. That is, assuming that the quantization width of the previous sample is Δ _i−1 , the quantization width of the next sample is determined by the following equation, and the value of the coefficient k is determined as a function of the amplitude value of the previous sample.

Δ _i = k · Δ _i-1

Here, a simple example is shown. FIG. 11 shows how to give a coefficient k when performing quantization with three bits. When quantizing with three bits, the most significant digit is assigned as a sign bit, so that two bits are responsible for the absolute value of the amplitude. In the case where a code having an amplitude equal to or more than the half value of the value indicated by the arrow in FIG. 11 is encoded in the 2-bit amplitude value, k is set so that the quantization width of the next sample becomes large. Increase the value to greater than 1.

Conversely, if the amplitude value is smaller than the half value, k is made smaller than 1 in order to reduce the quantization value from the previous quantization width. Of course, it is necessary to determine the minimum value and the maximum value of the quantization width so that the quantization width does not become unnecessarily small or vice versa.

Basically, the quantization result based on the adapted quantization width Δ is always absolutely in the vicinity of the center of the allowable level, and the state of overload distortion (approaching the upper limit of the allowable level). State) and granular noise (a state approaching the lower limit of the allowable level). For reference, FIG. 12 shows quantization coefficients at various bits. The speech encoding has been described above. Speech synthesis is realized by decoding by the reverse operation of the above-described encoding. The quantization coefficient is decoded by using the same value as the value used for encoding.

Next, the algorithm of the table reference method will be described with reference to FIGS. This conventional example shows speech synthesis of the 4-bit ADPCM system.

In FIG. 5, reference numeral 103 denotes an adder used for ADPCM decoding; 105, an adaptive controller; 109, a voice code data ROM for storing voice code data; and 108, a value read from the voice code data ROM 109. Code registers, 110, 111, and 112 are registers used for decoding, 501 is a quantization table that stores difference values used for decoding in advance, and 104 is a quantization width pointer that indicates the quantization table 501.

The table reference method is known as a method that does not require the multiplication process of the quantization coefficient k (for example,
Japanese Patent Application No. 3-115210).

A conventional table reference algorithm will be described. As a basic principle, a value of 1.25 is a method utilizing the fact that the value of the power is approximately close to the quantization coefficient as shown in the following equation. (1.25) ^-1 = 0.8 about 0.9 (1.25) ¹ = 1.25 about 1.2 (1.25) ² = 1.5625 about 1.6 (1.25) ³ = 1.953125 about 2.0 (1.25) ⁴ = 2.4441625 about 2.4

That is, for the maximum difference value data, 1.
The values multiplied by 25 times are stored in the ROM (Re
By writing all the data into a storage medium such as an ad only memory (see FIG. 6), and reading and accumulating the data, voice synthesis is performed.

The operation will be described with reference to FIG. The encoded data read from the audio encoded data ROM 109 is stored in the encoded register 108. The value stored in the encoding register 108 is used as it is as amplitude information in a difference value search of the quantization table 501. Further, the encoded data is input to the adaptive controller 105, and the quantization pointer 104 is controlled by a predetermined adaptive algorithm. When the absolute value of the amplitude information is large, the control is equivalent to multiplying the quantization coefficient k by a predetermined approximate magnification by shifting the pointer to the right by about 2 to 4. Conversely, when the absolute value of the amplitude information is small, it is equivalent to dividing the quantization coefficient k by a predetermined approximate magnification by shifting the pointer by about 2 to 4 to the left.

Quantization width pointer 104 and code register 1
The difference value data of the quantization table 501 indicated by 08 is read out and stored in the register 112. The difference value data stored in the register 112 is added to the content of the register 111 in which the immediately preceding synthesized waveform data is stored, and a new synthesized waveform is stored in the register 110 and output. At the same time as the output, it is also stored in the register 111 for the next synthesis.

Thereafter, new code data is read from the voice code data ROM 109, and the above operation is repeated.

[0023]

As described above, the conventional speech synthesizer using the ADPCM method has all the multiplication results, that is, difference value data, in order to reduce the number of multiplication circuits for multiplying quantization coefficients. Since this has been solved, it has a large-capacity storage medium for storing difference values.

An object of the present invention is to provide a speech synthesizing method and apparatus for reducing memory and hardware resources.

[0025]

Means for Solving the Problems] To achieve the above object, a method speech synthesis according to the present invention, contact with the first quantization width
First difference value data and a first quantization width of 1.25
Second difference value data in a second quantization width corresponding to the double
And a third quantization corresponding to 1.25 times the second quantization width.
Quantization table comprising third difference value data in width
Speech Synthesis Method for Decoding ADPCM Code Using GSM
Wherein the first quantization width, the second quantization width, and the third
The difference value data at the quantization width different from the quantization width is
First difference value data or second difference value data of the conversion table
Data or third difference value data by bit shifting
It is like that.

Further, the speech synthesis apparatus for carrying out the speech synthesis method according to the present invention, the quantization width point from ADPCM code
And the quantization width from the adaptive controller.
A ternary up / down counter that circulates according to the interchange,
Updates the number of ternary up / down counter cycles
And a first difference in a first quantization width.
Separated value data, the second which corresponds to 1.25 times the first quantization width
The second difference value data and the second quantum in the quantization width of
The third in the third quantization width corresponding to 1.25 times the quantization width
Quantization table consisting of differential value data
Quantization table according to count value by down counter
First difference value data, second difference value data,
Of any of the difference value data
Shift by the number of bits according to the number of rounds indicated by the down counter
The shifter that performs the operation and the difference value output from the shifter
The data and the previous synthesized waveform data are added and the synthesized waveform data is added.
And an adder for outputting data .

[0027]

By reducing the redundancy of the quantization table used in the process of decoding the ADPCM code, the storage capacity is reduced and the hardware resources are reduced.

[0028]

Embodiments of the present invention will be described below with reference to the drawings.

(First Embodiment) FIG. 1 is a block diagram showing a first embodiment of the present invention, and FIG. 2 is a diagram showing a quantization table of the first embodiment.

The present embodiment shows 4-bit ADPCM speech synthesis. In FIG. 1, the present invention
A quantization table 101 storing difference value data used in DPCM speech synthesis, a shifter 102, an adder 103 used for accumulating difference value data, and a quantization width pointer 1
04, adaptive controller 105, ternary up / down counter 106, up / down counter 107, ADPC
It comprises a code register 108 for storing the M code, a voice code data ROM 109, a register 110 used for accumulating difference values, a register 111, and a register 112.

First, the configuration of the quantization table shown in FIG. 2 and the principle of the present invention will be described.

The quantization table shown in FIG. 2 stores, in the vertical direction, those having different amplitude levels of difference values corresponding to ADPCM codes. In FIG. 2, the uppermost level stores the positive maximum differential value amplitude in FIG. 2 and is arranged so that the amplitude level of the differential value becomes smaller as going downward. The positive difference value having the smallest amplitude level near the center is set to the value of the amplitude level. The value below the center stores the negative difference amplitude level. The lowermost part stores the amplitude of the maximum negative differential value, and is arranged so that the amplitude level of the differential value becomes smaller as going upward. The negative difference value having the smallest amplitude level near the center is arranged to be the value of the amplitude level.

In the horizontal direction of the quantization table shown in FIG.
25 times, further 1.25 times (approximately 1.6 times from the maximum quantization width), and further 1.25 times (approximately 2 times from the maximum quantization width), such as a power multiplier of 1.25 times It is configured to be done. By the way, when the value is approximately doubled from the viewpoint of a certain appropriate quantization width, the value of the maximum quantization width is called, and if a shift operation is performed as 1-bit digital data, the doubled data is read out. , There is no need to hold all the difference values.

Further, it is also possible to make a half by performing a backward shift operation by making progress and thinking in reverse. Rather, the voice synthesis can be performed more accurately by dividing the original quantization width set. In other words, it means that high quality speech synthesis can be realized if the certain appropriate quantization width set is set by selecting the quantization width of the maximum amplitude set and its vicinity.

Therefore, when 1/25 times is performed three times, about 1/2 is obtained.
Since the quantization width is doubled, the maximum quantization width value and its 1
The virtual quantization table 1 shown in FIG. 2 which has only three types of values, /1.25 times and 1 / (1.25) ² times, is simply implemented by a simple shift operation. The principle is that a corresponding part or a part corresponding to the virtual quantization table 2 can be easily calculated by a shift operation at the same time as reading out the quantization table.

Next, the operation will be described with reference to FIG. It is assumed that the shifter 102, the ternary up / down counter 106, and the up / down counter 107 are shifted from the initial state when performing the voice synthesis processing to indicate the amplitude set of the minimum quantization width.

The coded data read from the voice coded data ROM 109 is stored in the code register 108. The value stored in the encoding register 108 is used as it is as amplitude information in a difference value search of the quantization table 101. Further, the encoded data is input to the adaptive controller 105, and the quantization width pointer 104 is controlled by a predetermined adaptive algorithm. When the absolute value of the amplitude information is large, the control is equivalent to multiplying the quantization coefficient k by a predetermined approximate magnification by shifting the pointer to the right by about 2 to 4. Conversely, when the absolute value of the amplitude information is small, it is equivalent to dividing the quantization coefficient k by a predetermined approximate magnification by shifting the pointer by about 2 to 4 to the left.

The adaptive controller 105 gives up or down control and the amount to the ternary up / down counter 106.

The ternary up / down counter 106 has 0,
Up, such as 1, 2, 0, 1, 2, ..., and conversely, down, such as 2, 1, 0, 2, 1, 0, .... The ternary up / down counter 106 issues an up request to the up / down counter 107 when the value changes from 2 to 0, and requests a down to the up / down counter 107 when the value changes from 0 to 2. The up / down counter 107 performs up / down in response to a request from the ternary up / down counter 106.

Quantization width pointer 104 and code register 1
The difference value data of the quantization table 501 indicated by 08 is read out, and after the shift value processing is performed on the difference value data read out by the shifter 102 according to the instruction of the up / down counter 107 and the ternary up / down counter 106. Stored in register 112. The adder 103 adds the shift-calculated difference value data stored in the register 112 to the content of the register 111 in which the immediately preceding synthesized waveform data is stored, and
The new synthesized waveform is stored in the register 110 and output. At the same time as the output, it is also stored in the register 111 for the next synthesis.

Thereafter, new code data is read from the voice code data ROM 109, and the above operation is repeated.

(Embodiment 2) FIG. 3 is a block diagram showing Embodiment 2 of the present invention, and FIG. 4 is a diagram showing a quantization table of Embodiment 2 of the present invention.

This embodiment shows a 4-bit ADPCM speech synthesis. In the second embodiment, the values of the quantization table in the first embodiment are set to the positive side and the negative side, and two values of the difference value of the same absolute value level are provided only for the difference between positive and negative. In this example, ADPCM speech synthesis is realized with a quantization table capacity which is half that of the first embodiment by adding and subtracting the amplitude level of the difference value by using the sign bit.

Referring to FIG. 3, in this embodiment, a quantization table 301 storing difference value data used in ADPCM speech synthesis, a shifter 102, an adder / subtractor 303 used for accumulating difference value data, a quantization width Pointer 104, adaptive controller 105, ternary up / down counter 106
, An up / down counter 107, a code register 308 for storing an ADPCM code, and a voice data ROM 10
9, a register 110 used for accumulating the difference value, a register 111, and a register 112.

First, the configuration of the quantization table shown in FIG. 4 and the principle of the present invention will be described. The quantization table shown in FIG. 4 stores, in the vertical direction, those having different amplitude levels of difference values corresponding to ADPCM codes.
In FIG. 4, the amplitude level in FIG. 4 is arranged such that the maximum difference value amplitude for both positive / negative values is stored and the amplitude level of the difference value decreases gradually toward the lower side. The lowermost side is set to the positive / negative differential value amplitude level value having the smallest amplitude level.

In the horizontal direction of the quantization table shown in FIG. 4, based on an amplitude set having an appropriate quantization width as a reference.
25 times, further 1.25 times (approximately 1.6 times from the maximum quantization width), and further 1.25 times (approximately 2 times from the maximum quantization width), such as a power multiplier of 1.25 times It is configured to be done. By the way, when the value is approximately doubled from the viewpoint of a certain appropriate quantization width, the value of the maximum quantization width is called, and if a shift operation is performed as 1-bit digital data, the doubled data is read out. , There is no need to hold all the difference values.

Further, it is also possible to make progress and think in the opposite way, and to halve it by a backward shift operation. Rather, the voice synthesis can be performed more accurately by dividing the original quantization width set. In other words, it means that high quality speech synthesis can be realized if the certain appropriate quantization width set is set by selecting the quantization width of the maximum amplitude set and its vicinity.

Therefore, when 1/25 times is performed three times, about 1/2 is obtained.
Since the quantization width is doubled, the maximum quantization width value and its 1
The virtual quantization table 1 shown in FIG. 4 which has only three types of values of /1.25 times and 1 / (1.25) ² times is simply implemented by a simple shift operation and is not actually implemented. The principle is that a corresponding part or a part corresponding to the virtual quantization table 2 can be easily calculated by a shift operation at the same time as reading out the quantization table.

Next, the operation will be described with reference to FIG. It is assumed that the shifter 102, the ternary up / down counter 106, and the up / down counter 107 are shifted from the initial state when performing the voice synthesis processing to indicate the amplitude set of the minimum quantization width.

The coded data read from the voice coded data ROM 109 is stored in the code register 308. The value stored in the encoding register 308 is divided into an amplitude information part and a code part, and the amplitude information part is used as it is for a difference value search of the quantization table 101. further,
The amplitude part of the encoded data is input to the adaptive controller 105, and the quantization width pointer 1 is calculated by a predetermined adaptive algorithm.
04 is performed. When the absolute value of the amplitude information is large, the control is equivalent to multiplying the quantization coefficient k by a predetermined approximate magnification by shifting the pointer to the right by about 2 to 4. Conversely, when the absolute value of the amplitude information is small, it is equivalent to dividing the quantization coefficient k by a predetermined approximate magnification by shifting the pointer by about 2 to 4 to the left.

The adaptive controller 105 gives up / down control and its amount to the ternary up / down counter 106.

The ternary up / down counter 106 has 0,
Up, such as 1, 2, 0, 1, 2, ..., and conversely, down, such as 2, 1, 0, 2, 1, 0, .... The ternary up / down counter 106 issues an up request to the up / down counter 107 when the value changes from 2 to 0, and requests a down to the up / down counter 107 when the value changes from 0 to 2. The up / down counter 107 performs up / down in response to a request from the ternary up / down counter 106.

The quantization width pointer 104 and the sign register 1
The difference value data of the quantization table 501 indicated by 08 is read out, and after the shift value processing is performed on the difference value data read out by the shifter 102 according to the instruction of the up / down counter 107 and the ternary up / down counter 106. Stored in register 112. The difference value data subjected to the shift operation stored in the register 112 is added and subtracted by the adder / subtractor 303 with the contents of the register 111 in which the immediately preceding synthesized waveform data is stored, and a new synthesized waveform is registered in the register. Stored in 110 and output.

The sign register 308 distinguishes between addition and subtraction.
Are distinguished by the value of the sign part of. At the same time as the output, it is also stored in the register 111 for the next synthesis.

Thereafter, new code data is read from the voice code data ROM 109, and the above operation is repeated.

[0056]

As described above, the present invention provides an ADPC
There is an effect that hardware resources can be reduced by reducing storage capacity by reducing the redundancy of the quantization table used in the process of decoding the M code.

Further, by using the maximum value, the bit precision of the quantization table for storing the difference value can be realized at the best, and there is an effect that high-quality sound can be synthesized.

[Brief description of the drawings]

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

FIG. 2 is a diagram illustrating a quantization table according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing a second embodiment of the present invention.

FIG. 4 is a diagram illustrating a quantization table according to the second embodiment of the present invention.

FIG. 5 is a block diagram of a conventional ADPCM speech synthesizer using a table reference method.

FIG. 6 is an explanatory diagram of a quantization table.

FIG. 7 is an explanatory diagram of a 3-bit PCM waveform.

FIG. 8A is an explanatory diagram of a 4-bit PCM waveform, and FIG.
FIG. 4 is an explanatory diagram of a 5-bit PCM waveform.

FIG. 9 is an explanatory diagram of a DPCM method.

FIG. 10 is an explanatory diagram of an ADPCM waveform.

FIG. 11 is an explanatory diagram of a 3-bit ADPCM quantization coefficient k;

FIG. 12 is an explanatory diagram of a quantization coefficient k at the time of 2 to 5 bits.

[Description of Code] 101 Quantization Table 102 Shifter 103 Adder 104 Quantization Width Pointer 105 Adaptive Controller 106 Binary Up / Down Counter 107 Up / Down Counter 108 Code Register 109 Voice Code Data ROM 110 Register 111 Register 112 Register 301 Quantization Table 303 Adder / Subtractor 501 Quantization Table

Claims (2)

(57) [Claims] 1. A first difference value data in a first quantization width.
Data and a second quantization width 1.25 times the first quantization width.
Second difference value data at a quantization width of
In a third quantization width corresponding to 1.25 times the quantization width
Using a quantization table composed of third difference value data, A
A speech synthesis method for performing a DPCM code decoding operation,
The first quantization width, the second quantization width, and the third
The difference value data at a quantization width different from that of
The first difference value data or the second
Bit of the second difference value data or the third difference value data
A voice synthesizing method characterized by being obtained by shifting . 2. A quantization width pointer is obtained from an ADPCM code.
An adaptive controller to determine and the quantization width from the adaptive controller
A ternary up / down counter that cycles according to the pointer
And count the number of rounds of the ternary up / down counter.
An up-down counter, and a first
1 difference value data, 1.25 times the first quantization width.
Second difference value data in a corresponding second quantization width;
Third quantization corresponding to 1.25 times the second quantization width
Quantization table comprising third difference value data in width
And the count value of the ternary up / down counter
The first difference value data of the quantization table,
Of the second difference value data and the third difference value data
For any of the difference value data,
A shift operation by the number of bits corresponding to the number of rounds indicated by the
Shifter and difference value data output from the shifter
And the previous synthesized waveform data to add the synthesized waveform data
A speech synthesizer comprising: an adder for outputting .