JPS5914760B2 - ADPCM regenerator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION 5 The present invention is used in a speech synthesis device that reads out information in the waveform region of a speech waveform from a storage area and synthesizes speech, in which a differential code, particularly an ADPCM code, is used as a normal PCM code. This invention relates to an ADPCM regenerator that converts .

10-Generally, a speech synthesis device stores sentences, words, or syllables to be output in a storage area, converts this information into audible speech through the synthesis device, and outputs the same as speech.

The audio stored in this storage area uses, for example, a 12-bit digital pattern (PCM code) sampled every 8kHz215 by an AD converter.
The amount of information is 12 bits×Bk-96 bits per second. A lot of research has been done20 on methods to reduce the amount of information in this enormous amount of audio, and Log-PCM, D
Waveform encoding methods such as PCM, ADM, ADPCM and L
There are analysis and synthesis methods such as PC, PARCOR, and LSP.

The present invention relates to a differential encoding method 25 among waveform encoding methods, and particularly relates to an ADPCM encoding method.

FIG. 1 shows a conventional ADPCM regenerator in the differential encoding method. In Fig. 1, 1 is an input terminal;
2 is an adder, 3 is a multiplier, 4 is an adder, 5 is a register,
6 is a table that converts the 30 ADPCM code Ln into a quantization step size movement coefficient Mn and outputs it.

7 is a multiplier, 8 is a transmitter, 9 is a register, 10 is a PC
This is an output terminal for the M code, and is a terminal for connecting to a DA converter 35 for converting a digital audio signal into an analog audio signal.

Let Ln be the ADPCM code of the audio from input terminal 1.

Ln is added by adder 2 by 0.5 for bias:07-.

The result is output by the multiplier 3 to the register 9. is multiplied by Output of register 9 △. is called the quantization step size. If the output of multiplier 3 is Qn, then Qn is A
This is the differential demodulation value reproduced by the DPCM code Ln, and is given by equation (1). The differential demodulated value Qn of the multiplier 3 is the entire output of the register 5.

is added and the result is credited. +1 is stored in register 5.
This father.

, Da. +1 is the voice PCM code. On the other hand, the ADPCM code LO at input terminal 1 is converted by table 6 and outputs a movement coefficient Mn.

Table 6 is a table for obtaining output Mn for ADPCM code Ln. The output Mn of table 6 is multiplier 7
△. is multiplied by Δn+1 to obtain Δn+1. Δ1+1 in equation (3) is set to the minimum value ΔMl by the limiter 8.

, the maximum value ΔNlax. That is, △0
+1 is △Ml. When the limiter 8 becomes smaller, the limiter 8 becomes Δ
Ml. is output, and if it becomes larger than ΔNlO, △
Output Nlax. The output Dn+1 of the limiter 8 becomes the quantization step size at the time of the next sample. Dn
+1 is stored in register 9. As described above, the block diagram of FIG. 1 operates. On the other hand, the present applicant filed a patent application in 1984-1
In 09800, we proposed an ADPCM regenerator that could be realized with a simple circuit configuration using low-speed elements by replacing the multiplication in the third equation with a simple operation on a memory storing a quantization step size.

FIG. 2 shows the improved conventional ADPCM regenerator.

In Fig. 2, 50, 51, 52, 53 are input terminals for the ADPCM code to be sent, 50 is the polarity bit, 5
1 is the input terminal for the most significant bit among the amplitude bits of the ADPCM code, 52 is the input terminal for the second bit, and 53 is the input terminal for the least significant bit.

54 is a serial input of register 58.

55, 56, 57, and 58 are 1-bit registers for storing ADPCM codes.

Registers 56-58 have a shift function. 59 is a read-only memory (hereinafter referred to as ROM), and register 5
The output of register 57 is the topmost address, and the output of register 57 is 2.
The third address, the output of the register 58, is set as the third address (the lowest address), and the pointer movement amount Dn is output.

60 is a 10-bit output pointer and is the output D of ROM59.
The output P is changed by 1.

A pointer limiter 61 limits the output Pn of the n pointer 60 to a specific range and outputs Qn.

A ROM 62 outputs 16-bit data X using the output Qn of the pointer limiter 61 as an address.

This X is an amount corresponding to the quantization step size, and in this example is a reference value corresponding to the amount of demodulation of the most significant bit of the amplitude bit. 63 is a 16-bit shift register which has the function of storing the output of the ROM 62 and shifting down.

64 is composed of 16 Ex-0R gates, and one input of each Ex-0R gate is commonly connected to register 5.
It is connected to the output of 5.

Therefore, when the output of register 55 is 1, E-0R gate 64 inverts the output X bit of shift register 63 and outputs it to adder 65.

When the output of shift register 55 is O, ExOR gate 64 outputs the output bit of shift register 63 to adder 65 as is. 65 is a 16-bit adder,
The output of Ex-0R gate 64 and the output of register 66 are added.

The result is stored in register 66 only when the output of register 56 is 1. 66 is a 16-bit register.

67 is a 16-bit register.

68 is an output terminal of the register 67.

70 is an initial value input terminal for setting an initial load value in the register 66, and 71 is a pointer initial value input terminal for setting a pointer initial value in the pointer 60.

Figure 3 shows the format of the data sent to the ADPCM regenerator shown in Figure 2 through the buffer memory.The data corresponding to one frame cycle is sequentially transmitted from the beginning by the number of ADPCM codes. A certain phoneme length (for example, 8 bits), the number of repetitions of the entire ADPCM code (for example, 4 bits), and an initial load value (for example, 12
bit), pointer initial value (for example, 6 bits) used as the initial value of the pointer in each repetition unit within one frame period, and ADPCM code L,, L2, L3, . . .
There are data such as LO, Lm (4 bits each), which are sent in this data format every frame period. 1 second
The operation will be explained using the drawings and FIG. Of each data sent in the data format shown in Figure 3, phoneme length, number of repetitions, initial load value, initial pointer value,
are respectively stored in registers (not shown).

Also, ADPCM codes L1, L2, ..., Ln,
. . . LITl is stored in a memory (not shown). Among the data stored in each register, five initial load values are input from the terminal 70 and set in the register 66.

* Also, the pointer initial value (6 bits) is input from the terminal 71 and set in the pointer 60. The pointer 60 performs sign extension to add the upper 4 bits to the initial value of the pointer, and outputs 10 bits. Next, A is read from memory.
The DPCM code is input to input terminals 50, 51, 52, and 53.

4-bit AD from input terminals 50, 51, 52, 53
The PCM codes are stored in register 55, register 56, register 57, and register 58, respectively. Table 1 shows the relationship between each register pattern and Ln. At the same time that these data are stored, the output X of the ROM 62 is stored in the shift register 63. register 55, register 56
, register 57, and register 58, the adder 65 performs the operations shown in Table 1. This calculation corresponds to equation (1). However, the first
In the table, the tree indicates the value of the register 66 before the operation. Also, [.] represents the largest integer that does not exceed the number in it (generally called the Gauss symbol). Next, it is expressed by equation (3). △n+1:4n0Mn゛゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゛゜゜゜゜゛゜゜゛゜゜゛゛(4) will be explained below.

In equation (4), MO and Δo are transformed as follows. Mn=ADn △ =△ ・Nmln ・・・・・・・・・・・・・・・・・・・・・・・・
・・・・・・・・・・・・・・・・・・(5)・APn・
・・・・・・・・・・・・・・・・・・・・・・・・
・(6) However, A1ΔMl.

are positive constants, Dn, . Pn is an integer. Then (4)
According to the formula, △n+1=6min'APn+Dn"゜゜゜゜゛゜゛゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゜゛゜゜゜゜゜゜゜゜゜゜゜゜゜゜゛゜゜゜゜゜゜゜゜ (7), and since Pn.Dn is an integer, Pn+1-Pn+Dn...
・・・・・・・・・・・・・・・・・・(8) is also an integer, and formula (7) is △n+1−△Min′APn
+1 ・・・・・・・・・・・・・・・(9), and △0+1 is also △.

It can be expressed by an expression of the same form as . Also, quantization step size △.

The maximum value ΔMax and minimum value ΔMll of +1 are also expressed in the same form. △ ・ One △ ・ ・AO・・・・
・・・・・・－・・・−・・・・・・・・・00) Ml
nmln△Max = 8min'APma
From the equation (self), if Pn+1 is in the range of O-Pmax, then Δ in equation (9).

+1 is △Ml. - ΔMO. Therefore, ΔMl for integers ranging from 0 to Pmax.
・If the value of APn is calculated in advance, stored in memory, and Pn is used as a pointer output (address of the storage element), only the calculation of equation (8) can be performed, without the multiplication of equation (3), Quantization step size △. +1 can be determined. Note that pointer 6 is set for each repetition unit of one frame period.
D in equation (8) where the pointer initial value is set to 0
Similarly to the quantization step size movement coefficient MO, O is a value corresponding to the audio compression code Ln, and is herein referred to as the pointer movement amount, and its values are shown in Table 2. The ROM 59 converts the ADPCM code L1 shown in Table 2 into the pointer movement amount Dn, and the pointer 60 executes the calculation of equation (8).

Further, the pointer limiter 61 limits the output of the pointer 60 to a range of O-PmO.

The output Qn of the pointer limiter 61 becomes the address input of the ROM 62. In the ROM62, the quantization step size ΔMln is set using the pointer limiter output Q1 as an address.
Values from AO to ΔMlnAPmax are stored.
However, due to the calculation process, a value four times as large as the above (demodulator for the most significant amplitude bit in the ADPCM code) is actually stored. The contents are shown in Table 3. Next, ROM6
The output of 2 is stored in the shift register 63.

Based on the output bit patterns of registers 55, 56, 57, and 58, the adder 65 performs the operations shown in Table 1. This calculation corresponds to equation (1). This calculation result is stored in register 66 and then output via register 67. In this way, when the processing for one ADPCM code data is completed, the next ADPCM code is input from the memory (not shown), stored in registers 55, 56, 57, and 58, and then processed as described above. Repeat.

During this time, the phoneme length data stored in the register (not shown) is stored in a counter (not shown), and is counted down by one upon completion of processing one ADPCM code.

This count value is constantly monitored, and when it reaches O,
It is determined that all ADPCM codes for one phoneme period have been completed. In addition, the [number of repetitions] data stored in the register (not shown) is stored in a counter (not shown) and is also constantly monitored. Count down the count value by one.

As a result, if the count value of the number of repetitions is not O, new phoneme length data is stored from the register to the counter, the initial load value is stored in the register 66, and the initial value of the pointer to the pointer 60 is stored. The data is stored and the process described above is repeated.

If the count value of the number of repetitions is O, this 1
Since the reading of the phonemes in the frame period is completed, the phoneme data for the next one frame period is read into the register for the length of each phoneme, the number of repetitions, the initial load value, the initial value of the pointer, and the process of reading the ADPCM code into the memory. I do. When speech is synthesized using such an ADPCM regenerator, the waveform of the synthesized sound is the same as shown in Fig. 4b, although the actual speech waveform is as shown in Fig. 4a. The result was that the waveform was repeated a number of times. Generally, the change in the transfer function of the vocal tract is considered to be gradual and constant for 30 msec.

Of course, it depends on individual differences and how fast you can do it. The number of times that the same waveform is repeated for each pitch in a voiced sound is probably about 3 to 5 times for men and 6 to 10 times for women.

However, the number of repetitions without losing the naturalness of the overall flow of words is limited to three times. Same waveform 3
By repeating the process, the audio information compression effect is tripled.

Therefore, the effect of compressing audio information by repeating the same waveform is extremely large. The larger the number of repetitions, the greater the compression effect. However, when the number of repetitions increases, the stationarity of the audio waveform is exceeded and the sound quality deteriorates. Also, considering the average power of the voice in this case, as shown in Figure 4c, even though the power of the actual voice changes smoothly, the synthesized voice repeats the same waveform, so the average power is constant during the repetition, and discontinuity in the average power with the next repetition waveform occurs. Therefore, a waveform repetition method is required in which the average power of the waveform changes smoothly.

Naturally, one voice waveform that repeats is encoded by some method (DPCM, .ADM, .ADPCM, etc.), and there is a method that reproduces waveforms with different power levels by repeating the encoded data. It becomes necessary.

The present invention is characterized in that for one ADPCM encoded waveform, waveforms with different powers (amplitudes) are reproduced by using the same ADPCM code.
The purpose is to reduce power discontinuity caused by repeated processing of a synthesized sound waveform that is reproduced by repeating PCM codes.
This will be explained in detail below.

FIG. 5 shows a first embodiment of the ADPCM regenerator of the present invention.

In FIG. 5, 130 is a control unit that controls the operation of each part, 140 is a multiplexer, 141 is a register that stores the "phoneme length", 142 is a register that stores the "number of repetitions", and 143 is a register for "initial value setting". 144 is a register for storing ``pointer initial value,'' 145 is ``pointer increase/decrease value.''
146 is a memory for storing the ADPCM code, 147 is a counter that stores the "phoneme length" stored in the register 141 and counts down at a predetermined timing, 148 is a register 142
The [number of repetitions] stored in the counter 14 is stored.
A counter that counts down when the count value of 7 becomes O and 3, and 149 is the register 143 or register 144.
This is a multiplexer that performs switching.

Each block from 150 to 168 corresponds to 50 to 6 in FIG.
Since the functions are exactly the same as those of each program in No. 8, the explanation will be omitted and only the names will be described.

150, 151, 152, 153 are ADPCM code input terminals, 154 is a serial input terminal, 155, 156,
157 and 158 are 1-bit registers each, 159 is a read-only memory (hereinafter referred to as ROM), 160 is a 10-bit output pointer, 161 is a pointer limiter, 1
62 is ROM, 163 is shift register, 164 is EX
-0R gate, 165 is adder, 166 is register, 1
67 is a register, and 168 is an output terminal of the register 167.

181 is an adder; 182 is a pointer increase/decrease value of the register 145 output from the adder 181 and the register 182;
A register 183 that stores the result of addition with its own output is an adder that adds the output of the pointer 160 and the output of the register 182. FIG. 6 shows the format of data sent through the buffer memory to the ADPCM regenerator shown in FIG. (for example, 8-bit data), the number of repetitions of all ADPCM codes (for example, 3-bit data), and the initial value of waveform reproduction for each repetition unit within one frame period.
An initial value setting pointer value (for example, 6-bit data) used as a pointer value when reading from the Ml62, a pointer initial value (for example, 6-bit data) used as the initial value of the pointer in each repetition unit within one frame period,
Pointer increase/decrease value (e.g. 3-bit data) and ADP
CM code L1, L2,..., Lnl...
- There is data of Lnl (4 bits each), which is sent in this data format every frame period.

FIG. 7 shows the reproduced waveform when the data shown in FIG. 6 is used to reproduce the waveform using the circuit shown in FIG. 6th below
The operation of the first embodiment shown in FIG. 5 will be described with reference to FIGS.

First, the 6th input is input from the transmission system or audio file.
Data in the data format shown in the figure is sequentially stored in a buffer memory (not shown).

The data stored in the buffer memory is sequentially fetched one frame period at a time. By sequentially switching the multiplexer 140, the data taken in from this buffer memory is stored in the register 141 as [phoneme length] and in the register 1.
42 contains the [number of rotations], the register 143 contains the [pointer value for initial value setting], the register 144 contains the [pointer initial value], the register 145 contains the [pointer increase/decrease value], and the memory 146 contains the ADPCM code. Each is stored. The counter 147 receives the output from the register 141.
The phoneme segment length] is stored, and the counter 148 stores the "number of repetitions." When reproducing a waveform using an ADPCM code, the value at the previous time point is as shown in equation (2).

Since the waveform reproduction value +1 is calculated based on , the initial value of the register 166 is required when starting waveform reproduction. The initial value of register 166 is given by the output of ROM162. In other words, from equation (6) and Table 3, the illusion △M
in. APINT. 4. ．． 、． ．． ．． ．． ．． ．． ．． ．． ．． ．． ,6
．． 6. ．． 6.02) is used to read out the ROM162, and the value X1 is stored in the register 166 as an initial value.

The pointer value PINT that provides this initial value is in the register 14.
This is the initial value setting pointer value stored in the register 183, which is stored in the pointer 160 via the magtiplexer 149, and after sign extension is performed by the pointer 160 to add the upper 4 bits, the adder 183 (register 182 is Initially it is set to O, so the value does not change even if added).
The ROM 162 is read out via the pointer limiter 161. In this way, the initial value of the phoneme waveform stored in the register 166 is transferred to the register 167, and the time point T1
is output from the output terminal 168.

On the other hand, as shown in equation (8), when determining the pointer 160, the pointer value is calculated based on the pointer Pn at the previous time point, so the initial value of the pointer 160 when starting waveform reproduction is required. For this purpose, the multiplexer 149 is switched to cause the [pointer initial value] stored in the register 144 to be stored in the pointer 160. The pointer 160 performs sign extension to add the upper 4 bits to the initial value of the pointer, resulting in an output of 10 bits. After time point T2, the audio waveform shown in FIG. 7 is reproduced using ADPCM codes L1 to L29 according to equations (2) and (8).

That is, the ADPCM codes stored in the memory 146 are taken into registers 155 to 158 one by one, and the
The calculations shown in Table 1 are performed based on the value of the M code.

The result of the operation is sent to the register 167 via the register 166.
and output from the output terminal 168. In this way, the code reproduction values at each time point are
··,father. . is obtained. The control section counts down the phoneme length of the counter 147 by one every time one ADPCM code is taken in and processed. This counter 147
The count value of is constantly monitored, and when it reaches O, it is determined that the processing of all ADPCM codes for one phoneme period has been completed. The [number of repetitions] data stored in the register 142 is stored in a counter 148 and is also constantly monitored, and when the count value of the aforementioned phoneme segment length reaches 0, the number of repetitions in the counter 148 is counted down by one. . As a result, the count value of the number of repetitions is O
If not, the phoneme length data is newly stored from the register 141 to the counter 147, the multiplexer 149 is switched to the register 143 side, and the initial value setting pointer value PINT stored in the register 143 is transferred to the pointer 160. Store. At this time, the 3-bit pointer increase/decrease value S stored in the register 145 is added to the adder 181.
is added to the value stored in the register 182 (which is 0 at the end of the first phoneme period), and the addition result is stored in the register 182 as a pointer increase/decrease value S'. The pointer increase/decrease value S' output from the register 182 is subjected to positive or negative sign extension in order to match the number of bits with the output from the pointer 160. Next, in the adder 183, the initial value setting pointer value PlNT output from the pointer 160 and the increase/decrease value g stored in the register 182 are added, and the addition result PlNT+
After limiting the range of S' with pointer limiter 161, R
The contents of OM162 are read out and stored in the register 166 as the initial value 1 of the next phoneme period (here, the second phoneme period). here? According to the formula (self), AS' times the initial value of the first phoneme period becomes the initial value of the second phoneme period. This initial value? is transferred from register 166 to register 167, and output from register 167 at time point T1. Subsequently, the multiplexer 149 is switched to the register 144 side, and the pointer initial value P1 stored in the register 144 is stored in the pointer 160.

This pointer initial value P1 is added to the increase/decrease value d output from the register 182 by an adder 183. This addition result P1+S' is transferred to the ROM1 via a pointer limiter 161.
62 address inputs. Based on this, ROMl6
The output X' of 2 is as follows. Since this is compared with the output x of the ROM 162 at the time point T2 during the first phoneme period, the following equation is obtained.

Also, comparing the quantization width Δ1 at time points T2 and T4 is as follows.

Therefore, the differential demodulation value q1 at each time point
and ql becomes.

Therefore, the time point T2? Code reproduction value father,,
Father 4 becomes.

In this way, the initial value in each phoneme period after the second phoneme period is set by using the initial value setting pointer value PlNT in the register 18.
By shifting the pointer increase/decrease value g stored in 2, the quantization width, differential demodulation value, and waveform reproduction value during ADPCM processing are multiplied by AS'.

Also, the second time point T in each phoneme period after the second phoneme period
The code reproduction value at each point after '2', dinner and dinner...
is similarly multiplied by AS'' for each phoneme period. By repeating this process, it is possible to output a waveform whose amplitude has been doubled by a certain amount, although the shape of the waveform does not change. Since a waveform with varying amplitude is reproduced, the power discontinuity that occurs when the waveform is output repeatedly is significantly reduced, and the quality of the synthesized sound is improved.

Although the first embodiment is an example applied to the ADPCM system, the present invention can also be applied to a DPCM system differential encoder.

A case where this DPCM method is applied is shown in FIG. 8 as a second embodiment of the present invention. DPCMTf shown in this FIG.
+The configuration of the generator is shown in FIG. 5 with the ROM 159 omitted. The difference between the ADPCM method and the DPCM method lies in the operation of the pointer 160.

In the ADPCM method, the pointer 160 changes for each process depending on the ADPCM code Ln input from the input terminals 150 to 153, but in the DPCM method, the pointer 16
0 is fixed. Therefore, the ROM159 is unnecessary in the DPCM system. Since other operations are similar to those of the ADPCM method, it is considered that there is no difference between the ADPCM method and the DPCM method in terms of the basic operation of the present invention. Furthermore, although it has been mentioned that in the DPCM method, the pointer 160 does not change depending on the DPCM code, a method is also considered in which the pointer value (initial value) input from the register 144 is set to a different value each time one phoneme waveform is reproduced. Alternatively, a method of using the same value in the reproduction of all phoneme waveforms is also conceivable. The second embodiment of the present invention can be applied to both. Since the present invention can output a waveform whose amplitude value is multiplied by a certain value from the same ADPCM code data, the average power of synthesized speech by a speech synthesizer that compresses speech information by repeating the same waveform. discontinuities can be reduced.

[Brief explanation of the drawing]

Figure 1 is a block diagram of a conventional ADPCM regenerator, and Figure 2 is a block diagram of a conventional ADPCM regenerator proposed in Japanese Patent Application No. 109800/1980.
CM regenerator block diagram, Figure 3 is the AD shown in Figure 2.
Figure 4a shows an example of an actual voice waveform; Figure 4b shows a synthesized sound waveform that repeats the same waveform three times; Figure 4c is the fourth
FIG. 5 is a block diagram of an ADPCM regenerator according to the first embodiment of the present invention, and FIG. 6 is a diagram showing the average power of FIG. It is. A diagram showing a data format of an ADPCM regenerator, FIG. 7 is a diagram showing an example of a reproduced waveform according to the first embodiment of the present invention,
FIG. 8 is a block diagram of a DPCM regenerator according to a second embodiment of the present invention. 130...Control unit, 140
...Multiplexer, 141, 142, 143
, 144, 145... register, 146...
...Memory, 147,148...Counter,
149...Multiplexer, 150'151,
152, 153... Sign input terminal, 154...
... Serial input terminal, 155, 156, 157,
158...1 bit register, 159...
...ROM, 160...Pointer, 161...
...Pointer limit, 162...ROM
ll63...Shift register, 164...
...EX-0R gate, 165...addition circuit,
166, 167...Register, 168...
...output terminal, 181, 183...addition circuit,
182...Register.

Claims (1)

[Claims] 1 a first memory 162 storing quantities corresponding to a plurality of predetermined quantization step sizes;
A pointer 160 that specifies a memory address and outputs one of the amounts corresponding to the quantization step size from the first memory, and a pointer 160 that stores a pointer movement amount and moves the pointer output according to the ADPCM code. a second memory 159 and a register 16 capable of storing reproduced PCM codes;
6 and the previous playback PC stored in the register 166
means 162 for reproducing the next PCM code based on the M code, the output of the first memory 162, and the ADPCM code;
165, and inputs information including amplitude increase/decrease values regarding phoneme waveforms, zero is set in the first phoneme waveform reproduction process of each frame period, and after the second phoneme waveform reproduction process, Accumulation storage means 181, 182 for accumulating and storing amplitude increase/decrease values for each process
and an addition means 183 for specifying the address of the first memory 162 by the sum of the outputs of the accumulation storage means 181 and 182 and the output of the pointer 160, and the PCM code initial value of each phoneme period is determined by the addition. An ADPCM regenerator characterized in that the addition result of the initial value setting data in the means 183 and the outputs of the accumulation storage means 181 and 182 is obtained from the first memory 162 as an address input and is set in the register 166. .