JPH0769718B2 - Speech synthesis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention relates to an adaptive differential pulse code modulation method (hereinafter, simply referred to as AD
The present invention relates to speech synthesis using the PCM method), and more particularly, to a waveform segment synthesis method in which one representative waveform segment is synthesized and voices are synthesized by repeatedly using this representative waveform segment.

[Prior Art] Conventionally, known waveform coding methods such as PCM and ADPCM have been adopted as a voice synthesis method.
This is a very effective method in terms of data volume compression, since waveforms can be synthesized with less than half the data volume compared to the CM system. The ADPCM method is a data compression method that uses the strength of the correlation between adjacent samplings of voice, and when encoding and quantizing the difference in amplitude value at consecutive sampling points, the quantization width is set to This is a method of adaptively changing to reduce noise in the signal. As the quantization width used for quantization, the next quantization width is predicted and used based on the current quantization width and the ADPCM code. An example of this prediction calculation is Δn + 1 = Δn × M (Ln), where Ln is the nth
ADPCM code data, Δn represents the size of the quantization width for the nth sampling point, M (Ln) is a prediction coefficient, and the current quantization width Δn is multiplied by the prediction coefficient M (Ln) according to the above formula. Use the value as the quantization width Δn + 1 for the next sampling point. When the voice level is low, the prediction coefficient M (Ln) becomes smaller than 1 and the quantization width Δn +
1 will be smaller. On the contrary, when the voice level increases, the prediction coefficient M (Ln) becomes larger than 1, and as a result, the quantization width Δn
+1 becomes larger. However, if the quantization width becomes too large, noise will increase. Therefore, it is necessary to determine the upper limit value of the quantization width. If the quantization width becomes 0, any prediction coefficient Since it remains 0 even when multiplied, it is necessary to set the lower limit of the quantization width in advance. A process including multiplication with such a prediction coefficient M (Ln) is performed, data of the quantization width Δn is stored in a table in a read-only memory (hereinafter, ROM), and the prediction calculation shown in the above equation is performed. Highly integrated circuit (hereinafter referred to as LSI) is a method that uses the quantization width Δn read from ROM.
Suitable for converting. In the LSI, a table of difference value data is created so that the ROM address can be designated by the ADPCM code and the quantization width pointer.

Increasing the value of the quantization width pointer means multiplying the quantization width by a prediction coefficient larger than 1 if the difference value to be inserted into the ROM according to the quantization width pointer is set to a value that increases at a constant rate. Meaning, the multiplication of the prediction coefficient can be replaced by addition and subtraction for the quantization width pointer.

[Problems to be Solved by the Invention] Generally, a voice waveform is divided into a voice noise of a consonant part, a tone waveform of a vowel part, and a connecting part between them. In particular, a tone waveform of a vowel part has almost the same period and the same shape. Are usually lined up with an envelope while gradually changing their shapes little by little. 2 to 3 consecutive in that
Since the waveforms have almost the same shape, the waveform segment synthesis method compresses the data amount of speech synthesis by selecting one waveform as a representative waveform and repeatedly using it.

When trying to compress the amount of data for speech synthesis by repeating such waveforms in the ADPCM method, the ADPCM method samples the speech at the Nyquist frequency and determines the difference value of the amplitude value of the speech waveform between adjacent sampling points. This is a method of encoding with an appropriate quantization width, and is a method of adaptively changing the quantization width according to the magnitude of the difference value at each sampling point. Therefore, when the representative waveform segment is synthesized by the ADPCM method, the quantization width is not constant within one waveform, and the quantization width at each sampling point depends on the immediately preceding ADPCM data. Also ADPCM
In the method, the sampling frequency is usually 4KHZ to 8KHZ depending on the relationship between sound quality and bit rate. When synthesizing the representative waveform segment using the ADPCM method as described above, the pitch period of the actual original waveform and the pitch period of the synthesized waveform cannot be perfectly matched, and they are synthesized due to the influence of the pitch period error. The waveform does not have zero amplitude at the last sampling point. Even when the pitch period of the original sound and the pitch period of the synthesized waveform completely match, the amplitude value at the last sampling point is the amplitude value at the previous sampling point plus the difference value determined by the quantization width and ADPCM code. Therefore, depending on the value of the quantization width at the immediately previous sampling point, the amplitude value at the last sampling point in one waveform may not be 0.

As described above, when a typical waveform segment is combined using the conventional ADPCM method as it is, since the amplitude value at the last sampling point of the representative waveform segment cannot be set to 0, such a representative waveform segment is used. If the synthesis of the representative waveform segment is repeated a plurality of times, the amplitude center of the synthesized waveform will change. This is a differential encoding method in which the ADPCM method basically adds the difference value to the amplitude value of each sampling point to obtain the amplitude value of the next sampling point, so if the final amplitude value of one waveform is not 0 This is because the value of the amplitude value becomes an error, and these errors are accumulated by repeating a plurality of times, and the amplitude center of the composite waveform changes. If the amplitude center of the synthesized waveform changes, the D / A converter that converts the waveform data to voice may overflow by the amount of the change in the amplitude center, and if it overflows, a normal voice waveform cannot be synthesized. There are drawbacks.

[Means, Actions and Effects for Solving Problems] According to the present invention, first, respective initial values are set in the first storage means and the second storage means, and then the ADPCM code of the first sampling point is set. When the difference value supplying means receives the data and outputs the difference value based on the ADPCM code data and the initial value of the first storage means, the output waveform amplitude data updating means makes a difference to the initial value of the second storage device. The values are added to form output waveform amplitude data at the first sampling point.
The output waveform amplitude data is stored and output in the second storage means. On the other hand, the quantization width pointer updating means
A new quantization width pointer value is formed based on the ADPCM code data and the initial value of the first storage means, and this is stored in the first storage means. At the subsequent second sampling point, a difference value is output based on the ADPCM code data and the updated value of the quantization width pointer, and the output waveform amplitude data updating means updates the difference value of the output waveform amplitude data. To form output waveform amplitude data at the second sampling point, store this in the second storage means, and output it as output waveform amplitude data at the second sampling point. On the other hand, the value of the quantization width pointer is also updated. When a plurality of output waveform amplitude data are output in this way, the first storage means and the second storage means are reset to their respective initial values, and the subsequent repeated waveforms are output in the same manner as described above. Waveform amplitude data is formed.

Therefore, even when the representative waveform pieces are combined, the output waveform amplitude data at the last sampling point of each representative waveform piece can be set to an initial value, for example, "0" by the initial value setting means, and a plurality of representatives can be set. Even if the waveform pieces are continuously processed, the amplitude center of each combined waveform can be maintained at a constant initial value, and a defect such as overflow can be completely solved.

Furthermore, in one embodiment of the present invention, a set of a plurality of difference values is provided, and since one of these sets is selected and used, even when the same ADPCM code data is repeatedly used, the same amplitude data is used. It is not repeated, and has an advantage that the envelope can be loaded for each waveform.

EXAMPLES Next, the present invention will be described with reference to the drawings.

FIG. 1 is a circuit configuration diagram of a first embodiment of the present invention.
In the voice synthesis method shown in the figure, ADPCM code data is latched in the L register 101, and the first address of the difference value data table ROM 106 is designated based on the contents of the L register 101. The value of the L register 101 decoded by the decoder 102 and the value of the A register 104 holding the current value of the quantization width pointer are added to determine the value of the quantization width pointer of the next sampling point. , This is A register 104
And the second address of the difference value data table ROM 106 is designated by the contents of the A register 104. A limiter 103 is inserted between the adder and the A register 105 so that the value of the quantization width pointer does not become a value other than the specified value. The ROM output is switched by the upper 2 bits of the address supplied to the difference value data table ROM 106. The contents of the A register 104 containing the value of the quantization width pointer are transferred to the A'register 105 and stored in advance, and when necessary, the A'register 105
The contents of are read out to the A register 104.
The difference value data is output from the difference value data table ROM 106, the contents of the X register 107 containing the previous waveform amplitude value at each sampling point are added to the difference value data, and the next waveform amplitude value is added to the X register. I put it in 107.
The contents of the X register 107 can be set to "0" by a reset signal. A flow chart of voice synthesis is shown in FIG. When voice synthesis is started, first, initial values “0” are set in the X register 107, the A register 104, and AD1 and AD2 (step I), and ADPCM code data is latched in the L register 101 at each sampling point. (L ← DATA in step II), the ROM address is specified by the value of the L register 101 and the value of the A register 104 of the quantization width pointer, and the output of the difference value data table ROM 106 and X
The value in register 107 is added, and the result is added to X register 1
It is stored in 07 (X ← X + (L, A) in step II), and the contents of the X register 107 are output from the D / A converter. At this time L
Value (L) obtained by decoding the contents of the register with the decoder 102
And the value in the A register 104 are added and stored in the A register 104 as the next quantization width pointer (A ← in step II).
A + (L)). The above is the flow when synthesizing voice by the normal ADPCM method. To explain the flow when synthesizing the waveform to be repeated next, reset the X register 107 and AD1 and AD2 to "0" at the beginning of the first repeated waveform. A register 1
After storing the contents of 04 in the A'register 105 (step III), the value X to be set in the X register 107 is calculated for the first sampling point S1. Less than,
Normal calculation processing is repeated for each sampling point (S1 to S8, see FIG. 5) and the value of the X register 107 is output from the D / A converter (step IV). At the beginning of the second repeated waveform, the X register 107 is reset to “0” (X ← 0 in step V), the contents of the A ′ register 105 are read to the A register 104 (A ← A ′ in step V), and the difference is calculated. As the contents of the value data table ROM 106, normal difference value data is selected when the ROM upper address is AD1 = 0, AD2 = 0, and is enlarged 1.2 times as a whole when AD1 = 0, AD2 = 1. The difference value data is selected, and when AD1 = 1 and AD2 = 1, the difference value data that is 0.8 times smaller is selected as a whole.
After setting AD1 and AD2 to appropriate values according to the increase and decrease of the envelope (AD1,2 in step V ← set value), normal arithmetic processing is performed at each sampling point and the value of X register 107 is set to D / Output from A converter (Step V
I). By repeating steps V and VI as many times as necessary, it is possible to output an enveloped repeating waveform using the same ADPCM code data as shown in S9 to S24 of FIG.

In the first embodiment, the A register 104 and the X register 107 respectively serve as the first and second storage means and the difference value data table RO.
M106 is a difference value supply means, L register 101, decoder 10
2. The adder constitutes the quantization width pointer updating means, the output side adder of the difference value data table ROM constitutes the output waveform amplitude data updating means, and the reset signal of the A'register and the X register 107 constitutes the initial value setting means. is doing.

FIG. 2 is a circuit configuration diagram of the second embodiment of the present invention. ADPC
The M code data is latched in the L register 201, and the address of the difference value data table ROM 206 to 208 is designated by the content of 201 of the L register. The value of the quantization width pointer at the next sampling point obtained by adding the value obtained by decoding the contents of the L register 201 with the decoder 202 and the value of the A register 204 holding the current value of the quantization width pointer In the A register 204, and the addresses of the difference value data tables ROM 206 to 208 are designated by the contents of the A register 204. The limiter 203 is inserted so that the value of the quantization width pointer does not exceed the specified value. The difference value data table ROMs 206 to 208 are generally the same as the ROM 2 (207) that stores a normal difference value data group.
ROM3 that stores the difference value data group that is 1.2 times larger
(208) and the ROM1 (206) that stores the difference value data group that is 0.8 times smaller than that of the switch 2
In 09, the output of ROM is selectively switched.

A flow chart of voice synthesis is shown in FIG. 4, but the difference value data table R is used by the switch 209 as compared with the first embodiment.
The flow is exactly the same except that OM206 to 208 are selectively switched according to the increase / decrease of the envelope.
The repeated waveform with the envelope shown in the figure can be output as in the first embodiment.

In the second embodiment, the A register 204 and the X register 210 constitute the first storage means and the second storage means, respectively, and the difference value data tables ROM1 to ROM3 (206 to 208) are the difference value supply means. The L register 201, the decoder 202, and the adder are quantization width pointer updating means, and the difference value data table RO
The output side adders of M1 to ROM3 constitute output waveform amplitude data updating means, and the reset signals of the A'register 205 and X register 210 constitute initial value setting means.

As described above, in the embodiment of the present invention, the value of the quantization width pointer at the first sampling point of the second and subsequent repeated waveforms is set to the value of the first sampling point based on the amplitude 0 as the difference value of the first sampling point of the repeated waveform. By using the same value as the initial value of the repeated waveform and using the ADPCM code data of the representative waveform segment as it is in the ADPCM method, it is possible to repeatedly synthesize the same waveform and increase the envelope in the difference value data table ROM.・ By inserting the difference value data according to the decrease in advance, it is possible to synthesize a repeating waveform whose envelope changes for each waveform, which is very useful for compressing the amount of synthesized data without degrading the sound quality. Is effective for. The above first and second
In the embodiment, three types of difference value data, which are normal difference value data, 1.2 times larger overall difference value data, and 0.8 times smaller difference value data are used. It is obvious that the value data table ROM may be set arbitrarily.

[Brief description of drawings]

FIG. 1 is a circuit configuration diagram showing a first embodiment of the speech synthesis system of the present invention, FIG. 2 is a circuit configuration diagram showing a second embodiment of the speech synthesis system of the present invention, and FIG. 3 is a diagram of FIG. A flow chart for performing voice synthesis in the first embodiment, FIG. 4 is a second flow chart of FIG.
FIG. 5 is a waveform diagram showing sampling points of the synthesized waveforms, which is a flow chart for performing voice synthesis in the embodiment. 101,201 ... L register, 102,202 ... decoder, 103,203 ... limiter, 104,204 ... A register, 105,205 ... A 'register, 106,206,207,208 ... difference value data table ROM, 209 ... switch, 107,210 ... X register.

Claims (2)

[Claims] 1. In a speech synthesis method for forming and outputting output waveform amplitude data for each of a plurality of sampling points, an initial value or a value of a quantization width pointer of a sampling point preceding each sampling point is temporarily stored. A first storage means, a difference value that outputs a difference value based on the ADPCM code data at each sampling point and the initial value or the value of the quantization width pointer at the preceding sampling point stored in the first storage means. Supply means and at each sampling point
Forming a quantization width pointer value for a subsequent sampling point based on the ADPCM code data and the initial value or the quantization width pointer value at the preceding sampling point stored in the first storage means; Quantization width pointer updating means for storing the value of the quantization width pointer for the subsequent sampling point in the first storage means, and temporarily storing the initial value or the output waveform amplitude data at each sampling point, and Second storage means for outputting output waveform amplitude data, and each sampling by adding the difference value to an initial value temporarily stored in the second storage means or output waveform amplitude data at a preceding sampling point. Output waveform amplitude data at the point, and the output waveform amplitude data is stored in the second storage means. Output waveform amplitude data updating means to be stored; and initial value setting means for setting respective initial values in the first storage means and the second storage means each time a plurality of output waveform amplitude data is output. Characteristic speech synthesis method. 2. The difference value supplying means has a plurality of sets of difference values, selects one of the plurality of sets of difference values, and outputs the difference value from the selected set. The speech synthesis system according to claim 1.