JPH0679239B2 - Speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Industrial field of application) The present invention is used in a voice synthesizing device for synthesizing voice by reading information in a waveform region of a voice waveform from a storage region. The present invention relates to an ADPCM player that converts a code into a normal PCM code.

(Prior Art) As a voice band compression method, there is ADPCM. This method takes the difference between the predicted value calculated at time T _i and the voice signal at T _{i + 1} in the data between adjacent samples of voice (time T _i and time T _{i + 1} ), and encodes it to obtain ADPCM. This is a method for reproducing a sound in a normal PCM code format by obtaining a quantized value of a difference signal by using a code and sequentially adding the values. It is also characterized in that the quantization width required when obtaining the quantized value of the difference signal is changed according to the ADPCM code. The ADPCM method requires a voice sample coded value of 3 to 12 which requires about 12 bits in the normal PCM code format.
It can be expressed by 4 bits, and an information compression degree of 3 to 4 can be obtained.

In addition, the actual waveform of FIG. 4 (a) is different from that of FIG. 4 (b).
There is also a method of performing information compression by repeatedly synthesizing the first waveform as shown in FIG. In general, the transfer function of the vocal tract changes slowly, and is considered to be almost steady for 20 to 30 ms. In the voice sound part, waveforms that are considered to be almost the same for each pitch synchronization appear repeatedly. Therefore, it is possible to substitute some of the succeeding segment waveforms by one segment waveform (each of the periodic waveforms is referred to as segment waveform). The repetition of this segment waveform does not impair the quality of the synthesized speech 3 to 4 times on average throughout the entire speech, and its compression effect is extremely large. However, if the number of repetitions becomes too large, the steadiness of the voice waveform is exceeded and the sound quality deteriorates. In this case, considering the average power of voice,
As shown in FIG. 6C, the average power discontinuity occurs due to the repetition of the waveform in the synthesized voice, although the power of the actual voice waveform changes smoothly.

Therefore, there is a need for a waveform repetition method in which the average power of the waveform changes smoothly. As a matter of course, one repeating voice waveform is encoded, and a method of reproducing waveforms having different powers by repeating the encoded data is required.

As a conventional technique in this regard, for example, Japanese Patent Publication No. 59-147.
It is described in Japanese Patent No. 60. The ADPCM regenerator disclosed in the above publication is capable of regenerating waveforms having different powers by using the same ADPCM code for one ADPCM-encoded segment waveform with a simple circuit configuration.

FIG. 2 shows the ADPCM reproducing device disclosed in the above publication. However, in order to facilitate the following description, the circuit disclosed in the above publication is slightly simplified.

The configuration and operation will be described below.

First, the decoding process of the ADPCM code in FIG. 2 will be described. The result of adding 0.5 as a bias to the input code L _i is multiplied by the quantization step width Δ _i to obtain the differential decoded value q. (Equation (1)).

q _i = (Li + 0.5) · Δ _i (1) The obtained differential decoded value q _i is stored in the register 55. Is added and the result Is stored in the register 55 (equation (2)).

this Is the voice PCM code. Next, the quantization step width Δ _{i + 1} used for decoding the ADPCM data L _{i + 1} next time is calculated, and the decoding process for one ADPCM code L _i is completed. here,
The calculation of the quantization step width Δ _i will be described. The quantization step width Δ _i is obtained from the quantization ROM 63 by referring to the pointer P _i . As shown in Table 1, the quantization ROM 63 stores a value that increases exponentially with respect to the address, and a quantization step width of Δ _i = Δ _min · A ^Pi (3) for the pointer P Is obtained. Therefore, in the block for changing the quantization step width, the quantization step width is changed by moving P _i by the pointer movement amount D _i . That is, P _{i + 1} = P _i + D _i (4) Δ _{i + 1} = Δ _min · A ^{Pi + 1} (5) Here, the pointer movement amount output from the ROM 60 is D _n is ADPCM.
It is obtained by referring to the pointer movement ROM 60 with the code L _i as an address. Table 2 shows the contents of ROM60.
The calculation of the equation (4) is performed by the pointer 61.

The above is the operation of ADPCM reproduction in the circuit of FIG. 2, but since reproduction of the ADPCM code is a sequential calculation, when waveform reproduction is started, Δ _i in equation (1) and equation (2) The initial value of is required.

Equation (2) in the circuit of FIG. Is given by the output of the quantization ROM 63. That is,
From equation (3) and Table 1 The quantization ROM 63 is read using a pointer value P _INT that satisfies And Similarly, the initial value of the quantization step width Δ _i in equation (1) can be calculated from equation (3) by giving the pointer initial value P _1. Is calculated. Further, the initial value of this pointer is the equation (4) P _i
Is the initial value of. From equations (1), (2), (3), and (6) and (7), the ADPCM code string L ₁ of the segment waveform is
Playback value of L ₂ ... L _i .... Becomes

Next, by using the ADPCM data L ₁ , L _2, ... In equation (6) Pointer decrement value S in addition to the pointer value P _INT give, which on new and P _INT, the initial value of the time (9) becomes, Is the value multiplied by (A ^S ). Similarly, if a constant value S is added to the pointer value P _i in the ADPCM reproducing process,
From equation (8) (Ten) It will be. In FIG. 2, the value S'added to the pointer is added every time the segment waveform is repeated by the adder 81, the register 82, and the adder 83, and S '= S, S "= S' + S, S.
= S ″ + S, the amplitude of the segment waveform is not changed and the amplitude is (A ^S ) times, (A ^2S ) times the first reproduced waveform,
(A ^3S ) times ... It gradually reduces the power discontinuity when the waveform is repeated by a simple circuit.

(Problems to be Solved by the Invention) However, in the circuit configuration of the conventional speech synthesizer described above,
The waveform does not necessarily become a constant multiple depending on the value taken by the pointer increase / decrease value S or the pointer value P _i , and there is a case where the purpose cannot be sufficiently achieved in an application in which amplitude control with high accuracy is required. That is, in order for the relations of equations (9) and (10) to hold, the relation of P _i ′ = P _i + S (11) always holds, and the quantization step width corresponding to P _i ′ in the above equation Must be present in the quantization ROM 63. However, the pointer value actually input to the quantization ROM 63 is limited by the pointer limiter 62, and S
Or, depending on the value of P _i ′ ^± , the condition of Eq. (11) may not be satisfied.

Here, the function of the pointer limiter 62 will be described. As described above, the quantization ROM 62 has the address P _min (=
0) to P _max against Δ _min A ^O ~Δ _min A ^Pmax becomes the quantization step size and is stored, thus it will be appreciated that the range of pointer value O～P _max inputted to quantization ROM63 Must be in However, since the pointer value P _i may not satisfy the above limit depending on the value of the ADPCM code L _i , a circuit for limiting the pointer to a value of O to P _max is required. This is the pointer limiter 62, this kind of ADPC
It is indispensable for M regenerator.

FIG. 3 (a) shows the function of the pointer 61 which is a problem in the conventional method. In the figure, P ′ _i is limited to P _max by the action of the pointer limiter 62 during the time interval Ta-Tb (1
1) Formula does not hold. The P _i are between the interval Tc-Td is despite being limited by P _min, P _'i is also regardless limiter limits (11) is not satisfied. In this example Ta
Equation (11) does not hold between -Tb and Tc-Td, but in actual ADPCM playback, since there is a point where even one point does not satisfy Equation (11) because it is a successive addition process, After that, the relation of equation (8) is broken in all points.

The present invention eliminates the drawback that the similarity of the waveform is broken in the amplitude control described above, and performs the amplitude control with high accuracy.
An object of the present invention is to provide a suitable speech synthesizer even when applied to law synthesis or the like in which one unit is used at various amplitudes.

(Means for Solving Problems) According to the present invention, a first memory storing an amount corresponding to a plurality of predetermined quantization step widths, an address of the first memory, and A pointer for outputting one of the quantities corresponding to the quantization step width from the first memory, a second memory for moving the output of the pointer according to the ADPCM code, and the output of the pointer were used for ADPCM code encoding. Speech synthesizer having a pointer limiter for limiting the quantization step width to a maximum value and a minimum value, and a register capable of storing a reproduced PCM code reproduced based on the quantization step width output from the first memory Target.

According to the present invention, in such a speech synthesizer, first, information including an amplitude increase / decrease value regarding a phoneme waveform is input, and zero is set in the process of reproducing the first phoneme waveform, and the second and subsequent phoneme waveform reproductions are performed. In, a means is provided for accumulating and storing the amplitude increase / decrease value for each process.

Secondly, there is provided addition means for designating the address of the first memory by the sum of the output of the storage means and the output of the pointer limiter.

Thirdly, in the first memory, the increment or decrement by the adding means with respect to the range of the quantization step width used at the time of ADPCM code encoding is extended to an upper address or a lower address, respectively, and these extensions are performed. The quantization step width corresponding to the generated address is stored in advance.

(Operation) One of the quantities corresponding to the quantization step width output from the pointer is supplied to the adding means via the pointer limiter. At this time, the output of this pointer is controlled by the pointer limiter between the maximum value and the minimum value of the quantization step width used during ADPCM code encoding. On the other hand, to the adder, the amplitude increase / decrease value for each process in the second and subsequent voice waveform reproduction is accumulated and supplied (0 is supplied in the process of reproducing the head phoneme waveform). As a result, the address of the first memory is the sum of the output from the pointer limiter and the amplitude increase / decrease value. That is, the address is expanded above and below the limit value of the pointer limiter. The quantization step width corresponding to this expanded address is stored in advance in the first memory. Then, the PCM code reproduced based on the quantization step width output from the first memory is stored in the register.

Since the limit value of the pointer limiter is shifted by the amplitude increase / decrease value in this way, an appropriate quantization step width can be set, and the above problem is solved.

(Example) Before describing an example of the present invention, the principle of the present invention will be described below.

The above problem is that in the conventional circuit, the relationship of the equation (11) cannot always be established. With respect to this problem, by removing the pointer limiter 62, only the pointer value can always satisfy the expression (11). However, with this method, it is not clear in what range the pointer value changes in various audio data to be ADPCM encoded, and therefore, it is not possible to know how easy the quantization ROM should be, and its capacity will increase considerably. become.

Therefore, the pointer value shown in FIG. 3 (a) is changed to the limiter operation as shown in FIG. 3 (b). That is,
The pointer limit value is vertically shifted by the pointer increase / decrease value S. By doing so, formula (11) is always established and the range of movement of the pointer can be grasped with certainty.
After that, the quantization ROM can be expanded up and down within the range where the amplitude needs to be changed. In real speech, the amplitudes of adjacent segments do not differ so much, so the expansion of the quantization ROM is not so large.

Hereinafter, the present invention will be described in detail based on an embodiment with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention. In the figure, 130 is a control unit for controlling the operation of each unit, 140 is a multiplexer, 141 is a register for storing “phoneme piece length”, 142 is a register for storing “repetition count”, and 143 is an “initial value setting pointer”. Register for storing `` value '', 144 for register for `` pointer initial value '', 145 for register for `` pointer increment / decrement value '', 146 for memory for storing ADPCM code, 147 for A counter that stores the “phoneme piece length” stored in the register 141 and counts down at a predetermined timing, a counter 148 stores the “repetition number” stored in the register 142, and counts down when the count value of the counter 147 becomes 0 or 0. The counter 149 for controlling is a multiplexer for switching the register 143 or the register 144. 155 is a register, 159
Is a pointer movement ROM, 160 is a pointer, 161 is a pointer limiter, 162 is a quantization ROM, 163 is a multiplier, 164 and 165 are adders, 1
Reference numeral 66 is a register, and 167 is an output terminal of the register 167. Further, 181 is an adder, 182 is a register for storing the addition result of the pointer increase / decrease value of the register 145 output from the adder 181 and the output of the register 182 itself, and 183 is a pointer limiter 161.
Is an adder that adds the output of the register 182 and the output of the register 182.

FIG. 5 shows the format of data sent to the voice synthesizer (ADPCM player) shown in FIG. 1 through a buffer memory (not shown in FIG. 1). In order from the beginning, the phoneme piece length that is the number of ADPCM data, the number of ADPCM code string repetitions, the initial value setting pointer value used as a pointer value when reading the initial value of waveform reproduction in each repetition unit within one frame period from the ROM 162, There is a pointer initial value, a pointer increase / decrease value, and ADPCM code L ₁ , L ₂ ... L _m data used as the initial value of the pointer in each repeating unit within one frame period,
This data format is sent every frame period.

FIG. 6 shows a reproduced waveform when the waveform shown in FIG. 5 is reproduced by the circuit shown in FIG.

The operation of the embodiment shown in FIG. 1 will be described below with reference to FIGS. 5 and 6.

First, 5th input from transmission system or audio file
Data of the data format shown in the figure are sequentially stored in a buffer memory (not shown). The data stored in the buffer memory is sequentially fetched for one frame period. The data fetched from the buffer memory is sequentially switched by the multiplexer 140, whereby the "phoneme piece length" is stored in the register 141, the "pointer value for initial value setting" is stored in the register 143,
The register 144 stores "pointer initial value", the register 145 stores "pointer increment / decrement value", and the memory 146 stores ADPCM code. The “phoneme piece length” output from the register 141 is stored in the counter 147, and the “repetition count” is stored in the counter 148.

Initial value of register 166 Is given by the output of ROM 162. That is, the initial value setting pointer value P _INT stored in the register 143 is stored in the pointer 160 via the multiplexer 149, and the pointer limiter 161 and the adder 183 (addition is performed because the register 182 is initially set to 0). Even if the value does not change), the ROM 162 is read out and the value is stored in the register 166. Initial value of phoneme waveform obtained in this way Is output from the output terminal 167 at time point T ₁ .

Then, prior to the waveform reproduction time point T ₂ later, the pointer 16
Set an initial value of 0. For this purpose the multiplexer 149
Is switched to store the “pointer initial value” P ₁ stored in the register 144 in the pointer 160. This value is (4)
It becomes the initial value of P _{i in} the expression and pointer limiter 161,
The ROM 162 is read out via the adder 183, and i in the equation (1) is read.
The quantization step width Δ ₁ at = 1 is determined.

Next, the waveform reproduction at time point T ₂ will be described. The first ADPCM data L ₁ input from the memory 146 is the adder 168.
After the bias 0.5 is added by the multiplier 163, the quantization step width Δ ₁ is multiplied by the multiplier 163 (equation (1)), and the adder 165 calculates the previous time. It is stored in 166 and is output from the output terminal 167 at time point T ₂ . On the other hand, in parallel with this, the ROM 159 is read from the ADPCM data L ₁ by using this value as an address, and the operation of the equation (4) is performed by the pointer 160, which becomes the pointer value at the next T ₃ point. Similarly, after time point T ₃ , the voice waveform shown in FIG. 6 is reproduced by using ADPCM codes L ₂ ... L ₂₉ .

In this way, the code reproduction value at each time point Is obtained. The control unit 130 counts down one phoneme piece length of the counter 147 each time one ADPCM code is fetched and processed. The count value of the counter 147 is constantly monitored, and when it becomes 0, it corresponds to one phoneme cycle.
Judge that the processing of all ADPCM codes is completed. The "repetition number" data stored in the register 142 is also stored in the counter 148 and is constantly monitored. When the count value of the phoneme piece length reaches 0, the number of repetitions of the counter 148 is counted down by one. .

As a result, when the count value of the number of repetitions is not 0, the phoneme piece length data is newly stored in the counter 141 and the multiplexer 149 is registered in the counter 147.
Switching to the 143 side, the initial value setting pointer value P _INT stored in the register 143 is stored in the pointer 160. At this time,
The pointer increase / decrease value S stored in the register 145 is added to the adder 181.
To the value stored in the register 182 (0 at the end of the first phoneme cycle), and the addition result is stored in the register 182 as the pointer increase / decrease value S ′.

However, if the output of the pointer limiter 161 has already reached the limiter limit value P _min or P _max , the address underflow is caused by using the same quantization ROM as the conventional ADPCM regenerator shown in FIG. Or overflow occurs. Therefore, in the quantization ROM 162 in this embodiment, the limiter limit value is vertically expanded and expanded as shown in FIG. 7, and the corresponding quantization step width is stored. Therefore, in the embodiment shown in FIG. 1, the initial value of the pointer 160 and the pointer limiter 161 are set in order to correct the deviation of the quantization step width corresponding to the address caused by the address expansion.
An offset for the lower extension of the address is added to the limit value of. As a result, equations (9) and (10) always hold,
Therefore, although the shape of the waveform does not change when the phonemes are repeated, it is possible to output a waveform having a constant amplitude. Here, how much the quantization ROM should be expanded becomes a problem, but the size of expansion, that is, the amplitude increase / decrease value S
The size of the cumulative result S'of corresponds to the ratio of the average powers of the adjacent frames. Considering that the change in the amplitude is relatively small in the section where the phoneme shapes can be considered to be almost the same, the amplitude is reduced to 1/2. It is sufficient to expand the quantization step width in FIG. 7 by about 2 times and expand the upper and lower steps by about 8 steps. From the viewpoint of the entire circuit, a very small increase can provide a sufficient effect.

The first embodiment of the invention has been described above. In the first embodiment, the speech information is compressed by repeating the pieces, but the rule synthesis for the purpose of synthesizing an arbitrary word is performed.
This method is a more effective means when using the ADPCM method. FIG. 8 shows the structure of the quantization ROM in the second embodiment. That is, in law synthesis, small speech units such as syllables, phonemes, and phonemes are connected to form a continuous speech. Therefore, the amplitude of the voice unit used must be easily controlled to various values depending on the situation in which it is used.

Also in this case, the circuit shown in FIG. 1 can be used as it is as a synthesizing unit for law synthesis. In this case, however, the method of controlling the amplitude of the segment data that is the source of the synthesis is normalized and the amplitude is controlled as the maximum amplitude only in the direction of decreasing the amplitude. It is advantageous to think. Therefore, the quantization ROM 162 expands the required quantization step width only in the lower direction as shown in FIG.
Step, it is enough to expand about 1/20 times).

(Effects of the Invention) As described above, according to the present invention, a waveform in which the amplitude value of the same ADPCM data is multiplied by a constant value can be output better than that of the conventional configuration, and therefore, by repeating the same waveform, It is possible to further reduce the discontinuity of the average power of the voice-synthesized voice in which the information compression is performed.

[Brief description of drawings]

FIG. 1 is a block diagram of the first embodiment of the present invention, FIG. 2 is a block diagram of a conventional ADPCM regenerator, and FIG.
FIG. 3B is a diagram showing an example of the function of the pointer in the ADPCM reproducing device shown in FIG. 3, FIG. 3B is a diagram showing a limiter operation in the present invention, and FIG. 4A is a diagram showing an example of an actual voice waveform. FIG. 4 (b) is a diagram showing the waveform of a synthesized sound in which the same waveform is repeated three times, FIG. 4 (c) is a diagram showing the average power of each of FIGS. 4 (a) and (b), and FIG. 6 is a diagram showing a data format in the first embodiment of the present invention, FIGS. 6 (a) and 6 (b) are diagrams showing an example of reproduced waveforms according to the first embodiment of the present invention, and FIG. 7 is a diagram showing the present invention. ROM 162 in the first embodiment
FIG. 8 and FIG. 8 are views showing the contents of the ROM 162 in the second embodiment of the present invention. 130 ... Control unit, 140 ... Multiplexer, 141-145 ... Register, 146 ... Memory, 147,148 ... Counter, 149 ... Multiplexer, 155 ... Register, 159 ... ROM, 160 ... Pointer, 161 ... … Pointer limiter, 162 …… ROM, 163,165,181,183 …… Adder, 166,182 …… Register.

Claims (1)

[Claims] 1. A first memory storing an amount corresponding to a plurality of predetermined quantization step widths, and an address of the first memory is designated to change the quantization step width from the first memory. A pointer that outputs one of the corresponding amounts, a second memory that moves the output of the pointer according to the ADPCM code, and a maximum and minimum value of the quantization step width used when the output of the pointer is used for ADPCM code encoding. In a voice synthesizer having a pointer limiter for limiting between values, and a register capable of storing a reproduced PCM code reproduced based on the quantization step width output from the first memory, The information including the input is input, zero is set in the process of reproducing the first phoneme waveform, and means for accumulating and storing the amplitude increase / decrease value for each process in the second and subsequent phoneme waveform reproduction, Addition means for designating the address of the first memory by the sum of the output of the storage means and the output of the pointer limiter is provided, and the first memory has a range of the quantization step width used at the time of the ADPCM code encoding. On the other hand, a speech synthesizer characterized in that the increment or decrement by the adding means is expanded to an upper address or a lower address, respectively, and a quantization step width corresponding to these expanded addresses is stored in advance.