JPS5919360B2 - speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech synthesis device that implements a speech synthesis method using a composite sine wave (referred to as CSM synthesis method) using a digital circuit.

The CSM synthesis method was filed in a patent application filed in 1983 by the inventors of this application.
It was proposed in No. 128364, ``Speech Synthesizer,'' and its principle was announced at the Autumn Research Conference of the Acoustical Society of Japan in 1978 (Proceedings of the same conference, 3-2-3, October of the same year).

The most powerful conventional speech synthesis method is the PARCOR speech synthesis method (Patent No. 876024).
Or LSP type speech synthesis method (Patent application 1984-0223)
18 "speech synthesizer").

Both have recently been made into LSIs (semiconductor integrated circuits) and are on the verge of being widely used. However, these methods use linear filters (succursive filters) with feedback loops.
rsive filter), there are inherent problems such as oscillation and instability. For this reason, it is quite difficult to realize these types of speech synthesis using analog circuits, and even when using digital circuits, the calculation precision is more than 10 bits, and the processing accuracy is 9 to 21 bits.
It required several digital multiplications. It was also necessary to pay close attention to the quantization of parameters. As a result, the hardware of this type of speech synthesizer is generally complex and expensive. The CSM synthesis method was invented to overcome the above drawbacks.

CSM synthesis does not require speech synthesis filter calculations, can be realized with relatively simple etware, and can handle parameter quantization relatively coarsely, shortening the calculation word length, and improving the quality of synthesized speech. The present invention provides a speech synthesis device that is sufficiently practical and can be applied to a wide range of applications. However, the implementation method described in the patent application for CSM synthesis uses analog circuits, and subsequent publications on CSM (for example, Electronic Technology Vol. 22, No. 13, Nikkan Kogyo Shimbun, 1980
Although it has been reported that CSM speech synthesis is possible using digital circuits and microprocessors (December 2013), there is no mention of specific implementation methods or calculation methods. The main principle of the CSM synthesis method is to give n (usually about 3 to 6) frequencies {ω1} and amplitudes {a1} and create synthesized speech as the sum of n sine waves. If the principle is carried out as is, n multiplications are required, which increases the scale of the hardware or increases the calculation time.

The purpose of this invention is to significantly reduce the amount of computation required for CSM speech synthesis and to realize a practical speech synthesis device, and is basically characterized by providing a speech synthesis device that does not use multiplication. There is.

This will be explained below using the drawings. The basic principle of the CSM speech synthesis method is as follows.

For an n-order model (n is 3 to 6), n CSM frequencies {ωi}, n CSM amplitudes {Ai}, pitch period T
give. The speech synthesizer uses these given ωI, ai
The following equation for time t = 0, 1, 2..., T-
Calculate for 1.

At time T, t=0 is set, and then t=1, 2, etc.
...continues. The value of Xt at each time t obtained in this manner is used as a sample value to output the synthesized speech.
CSM parameter {ωi; Ai; T} is the frame period,
In other words, the time unit used to analyze and synthesize speech waveforms is the period during which speech is considered to be almost stationary, usually 5 to 40 mS.
Updated and given for each degree. The principle of this invention is to perform the following modifications to the above operations.

In other words, in the frame with the maximum amplitude of the sentence speech to be synthesized, the CS is adjusted such that a1+A2±...Ring r≦1.
Scaling of M amplitude {Ai} is performed j times. Then, {Ai} is expressed for each frame as follows. That is, {Ai} itself has a large spread of distribution, so 2 is grouped out as a common factor as a gain term for each frame at -L.

In this way, the distribution range of {Bi} normalized by a power of 2 becomes smaller.

This value L indicates the approximate amplitude value of the audio signal in that audio frame period. Normalized amplitude {
Bi} is further encoded into several bits through a nonlinear companding law including the μ companding law. {Bi} subjected to nonlinear compression encoding is written as {Bi}. On the other hand, CSM frequency {ω1}
is also quantized and encoded and written as {ωi}.

FIG. 1 shows an embodiment of this invention. Here, n=4 {
Let us consider the case where {ω1} is encoded as 4 bits and 8 bits as {ω1}.

In order to calculate the sample value X at time t, the sinusoidal hemogram storage unit 1 stores the number of possible values of {b1}, that is, 16 types of sinusoidal waveforms with different amplitudes as 256 (=28) points in one period. stored in the section.

In other words, 4096 (=16×256) bytes of sine wave values are stored. This storage unit 1 is usually realized by a semiconductor read-only memory (ROM). The upper 4 bits of the address are used to select 16 levels of amplitude values, and the lower 8 bits are used to select the sine wave angle (0 to 2π) at 256 points. Quantized CSM frequency ω and ω↓, ω
1, ω↓ are angle accumulation registers 31, 32, 3,
, 34 and input ωi (1=1, 2, 3,
The value of 4) is (M) modulo 256 for each sample generated.
Od256) is accumulated. The pitch counter 5 has a period T times one sample generation interval (sample period), and the angle accumulation registers 31 . . . 34 are reset for each period. Given amplitude b:, b:, bλ, b:
One of them is selected by the multiplexer 2, and one output of the corresponding angle accumulation register 3,...34 is selected by the multiplexer 4, and the sine waveform storage section 1 is The storage unit 1 is accessed and used multiplexed in a time-division manner. The four types of sine wave values read out as described above, namely AislnωIt(
The values of 1=1, . . . , 4) are added in an accumulation register (accumulator) 6. The contents of the composite sine wave accumulation register 6 are shifted by the value of L at that time by 0 to 7 bits by the shifter 7. The shifted contents are converted into an analog signal by the DA converter 8 and sent out as a synthesized audio output signal. selection control for multiplexers 2 and 4;
Accumulation control for the accumulation registers 31 and 6, count control for the pitch counter 5, shift control for the shift r, and conversion control for the DA converter 8 are performed by the timing control section 9. The speech synthesis operation is as follows.

If the synthetic voice sampling frequency is 8KHz, it is 125
The following operation is performed during μS. First, angle accumulation register 3P
In the contents, the corresponding input ω1 value is MOd256.
Add . Through multiplexers 2 and 4, the sine waveform storage unit 1 is accessed using a total of 12 bits, including 4 bits of the amplitude designation b1 and 8 bits of the contents of the angle accumulation register 3i, as an address. The resulting sine wave values of the specified amplitude and the specified angle are sequentially displayed in four types (
1=1 to 4), are added on the composite sine wave accumulation register 6, and the result is shifted by the shift amount specified by the shifter 7 (
L bits) to the right. This result is output as an analog signal by the DA converter 8. However, every time the value of the pitch counter (usually a down counter) becomes O, the contents of the angle accumulation register 3i are all set to O, and its output value is also set to O. Although the above is a basic embodiment of the present invention, several measures can be taken to reduce the capacity of the sine waveform storage section 1.

That is, since a sine wave has the following properties, by memorizing 1/4 period of the waveform, the rest can be easily calculated.

In addition, although it has been stated that 16 types of amplitude values are required, when the necessary amplitude values are in a relationship of V to b powers, that is, the relative values of the amplitude values are 1, 1/j', 1/2, 1/2j Ding, 1/4・
..., you can store the sine waveforms of two types of amplitude values, 1 and 1/V, and for other amplitude values, just shift the read result by the necessary amount. In speech synthesis, it is desirable if the amount of information {ωl;b';T} provided for each frame can be reduced.

Generally, each CSM frequency {ωi} has a limited distribution range, so within that range, it is quantized to about 4 bits, and the speech synthesizer is equipped with a decoder that supports the encoding. The amount of information can be compressed. The following modifications of the above embodiment are also possible. /l>-Digital multiplication by digital multiplication of the envelope synchronized with the pitch period on the output side of the l→S knobs 7, or the output side of the DA converter 8 can be multiplied by analog multiplication.

This makes it possible to improve the quality of synthesized speech. Alternatively, in order to reduce the amount of calculation, the synthesized speech waveform for one pitch period calculated as in the above embodiment can be stored in the buffer RAM.
It is also possible to read out the previously stored contents of the buffer RAM repeatedly within the same frame without repeating the waveform generation calculation. In addition, if you prepare two buffer RAMs (two of about 128 bytes long), even when synthesizing a female voice with a short pitch period, you can calculate one pitch period relatively slowly. Write to one buffer, then repeatedly read that buffer at a speed that obtains the desired pitch period, and store the calculation for one pitch period for the next frame in the other buffer. By reading out the signals alternately, it is possible to slow down the speech synthesis calculation speed. 1st
A specific example of the surroundings of the speech synthesizer shown in the figure is shown in Figure 1.
Parts corresponding to those in the figure are given the same reference numerals. A 7-bit pitch period T,3 is stored in the audio data buffer 15 of 8-bit width x 16 bytes at an appropriate time for each audio frame period.
The number L of bit shifts, the amplitude {BWi} of each 4 bits, and the angular frequency {ωi} encoded by I of each 4 bits are sent from, for example, a host computer and stored.

The amplitude read from this pamphlet 15 holes {Bi-
} are respectively stored in the amplitude register 101, and the contents of the amplitude register 101 are selected by the multiplexer 2.

Encoded angular frequency {ω11 read from buffer 15
} are decoded by the decoder 11 and stored as {ωi} in the frequency registers 12i, and these frequency registers 1
The contents of 2i are respectively accumulated in the angle accumulation register 31. The pitch period T and shift number L read from the buffer 15 are pitch register 13 and L register 1, respectively.
It is stored in 4. The content T of the pitch register 13 is preset in a pitch counter 5, and the counter 5 subtracts the synthesized voice sampling clock from the value. Each time the value becomes zero, the register 3i is reset, and T is preset in the counter 5. The shifter 7 is shifted by the content L of the L register 14. In the above explanation, various numbers of bits were determined for convenience, but these numbers of bits are not fixed and may be any number of bits. Generally, the longer the word length is, the higher the quality of the synthesized speech can be. As explained above, the present invention makes it possible to significantly reduce the amount of calculation required in the CSM speech synthesis method.

In other words, according to the present invention, the CSM speech synthesis method, which conventionally required the burden of multiplication operations, can be executed by a simple arithmetic device, and the amount of calculations can be reduced and the process can be performed in a short time. The device of this invention can be realized by a microprocessor. This inventor used an 8-bit general-purpose microprocessor (Z8OA, 4MHz) and succeeded in real-time speech synthesis using only microprocessor software. Furthermore, the present invention can also be realized as a microprogram for a one-chip microcomputer. That is, it is sufficient to have the sine waveform storage unit 1 as an internal memory of a one-chip microcomputer, and to define operational instructions convenient for performing voice synthesis calculations by a microprogram as described above. If it also has a built-in DA converter of about 8 bits, it becomes a microcomputer with an analog synthesized voice output function, which is very effective.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a basic embodiment of this invention.
FIG. 2 is a block diagram showing one of the specific embodiments. 1...Sine waveform storage section, 2...Amplitude b
l multiplexer, 3... Angle accumulation register (
1 to 4), 4... Angle multiplexer, 5...
... Pitch counter, 6 ... Composite sine wave accumulation register, 7 ... L bit shifter, 8 ...
... DA converter, 9 ... Timing control puncture circuit, 10 ... Amplitude Bi value register, 11 ...
・・・ω value decoder, 12 ・・・ω value register, 1
3... Pitch cycle T value register, 14...
...Shift number L value register, 15...Audio data buffer.

Claims (1)

[Claims] 1 Angle values corresponding to n (n is an integer of 2 or more) sine wave components corresponding to the audio to be synthesized for each audio frame period, and amplitude values indicating the normalized amplitude of each of the components. , n angle accumulation registers into which the pitch period and a shift value indicating the approximate amplitude of the audio signal within the frame period are inputted, and which accumulate each of the above-mentioned angle values, respectively, and these angle accumulation registers. and a selection means for sequentially extracting the contents and the corresponding amplitude values as a set, and the contents and amplitude values of the selected angle accumulation register are given as addresses, and the sine wave value corresponding to the angle and amplitude is read out. a sine waveform storage section, a composite sine wave accumulation register that calculates the sum of n sine wave values read from the sine waveform storage section, and a composite sine wave accumulation register that calculates the sum of the n sine wave values read from the sine waveform storage section; and a means for resetting each of the angle accumulation registers for each of the pitch periods.