JPH0449956B2 - - Google Patents

Description

[Detailed description of the invention]

(Technical Field) The present invention relates to a speech synthesizer that reads information in the waveform region of phoneme waveforms from a storage area and synthesizes speech, and in particular, performs addition to superimpose a plurality of symmetrical ADPCM-encoded phoneme waveforms and outputs speech. Concerning a speech synthesizer that obtains. (Prior Art) When a human voice is a voiced sound, it is emitted when airflow from the lungs becomes a quasi-periodic impulse flow through the vocal cords and resonates with a hollow resonator called the vocal tract. Figure 1 shows the waveform of a human voice. As shown in Figure 1, the human voice is “pitch”.
Almost the same waveform is repeated every cycle called . This is because, as stated earlier, the human voice is emitted by a quasi-periodic impulse stream, and the pitch period is equal to the interval of this impulse stream. When trying to synthesize such voices, various methods can be used depending on the format in which the voice information is stored. One method is to store one-pitch period waveforms of speech (called phoneme waveforms) for various speech sounds in a storage area, and output speech by connecting these phoneme waveforms according to control information. be. Figure 2 shows a configuration using this method. 1 is a storage area for storing phoneme waveforms of 1 pitch period for various voices, 2 is a storage area for controlling information such as pitch and amplitude multiplication for connecting phoneme waveforms, and 3 is a storage area for storing phoneme waveforms with one pitch period. This is the synthesis section that connects the pieces together. Speech is synthesized by the synthesizer 3 by connecting the voices stored in the storage area 1 using the control information stored in the storage area 2. However, in order to create a highly natural synthesized sound, it is necessary to The pitch and volume of the voice must be appropriately controlled. The loudness of the voice is controlled by multiplying the phonemes in the storage area by a certain amount using amplitude multiplication information in the storage area. Further, the pitch of the voice is controlled by pitch information in the storage area 2. However, the phoneme waveform in storage area 1 has the pitch period length of the voice when the phoneme waveform was extracted, and does not necessarily match the pitch given by the control information. Therefore, if the pitch of the control information is shorter than the speech segment length, the rear end of the speech segment is cut off, and if the pitch of the control information is longer than the speech segment length, the last value of the speech segment is extended for control. It is assumed that the phoneme has the same pitch length as the information. (See Figure 3) However, if the phoneme waveform stored in storage area 1 is cut in the middle, if the phoneme waveform is not sufficiently attenuated, discontinuity will occur between the connected phoneme waveforms and the synthesized sound will be affected. This has the disadvantage of causing a shadow, and conversely, when the phoneme waveform is lengthened, the spectrum of the phoneme waveform is deformed, leading to deterioration of sound quality. (Objective and Summary of the Invention) The object of the present invention is to solve these drawbacks, and to solve the problem of multiple channels that are shifted by pitch period.
In a speech synthesizer that superimposes ADPCM-encoded symmetric phoneme waveforms, by time-multiplexing ADPCM code reproduction, the adder used for the superimposition operation of symmetric phoneme waveforms can also be used as cumulative addition/subtraction means in the ADPCM code regenerator. By doing so, we have realized a speech synthesizer that synthesizes natural, high-quality speech with a simple circuit configuration, and will be described in detail below. (Premise of the invention) As mentioned before, speech is produced by the resonance of impulse flows from the vocal cords, and this can be replaced by an electric circuit. That is, speech is considered to be a superposition of output waveforms (hereinafter referred to as phoneme waveforms) of a resonant filter corresponding to one impulse emitted by an excitation circuit corresponding to the vocal cords. This will be explained using FIG. 4. Reference numeral 10 in FIG. 4 is an impulse train generated by the excitation circuit. Here, the time interval between each impulse is the pitch period interval. 11 is a synthesized audio output waveform obtained by driving a resonant filter with an impulse train generated by an excitation circuit. 12 is a phoneme waveform when the resonant filter is driven by impulse _P1 of the impulse train generated by the excitation circuit. Similarly, 13 to 17 are phoneme waveforms when the resonance filter 4 is driven by impulses _P2 to _P6 .
Since the driving impulse train 10 emitted from the excitation circuit is the addition of impulses P ₁ to P ₆ , according to the superposition theorem, the synthesized speech output waveform 11 is the sum of each phoneme waveform 1.
Obtained by adding 2 to 17. The phoneme waveform 12 shown in FIG. 4 is zero before time point _t1 . The phoneme waveform output after time point t ₁ has the property of attenuating over time and becoming 0 at an infinite time point. In fact, in the case of a speech waveform, the phoneme waveform is considered to be almost 0 when 16 milliseconds have passed from the excitation point (t ₁ ). In the present invention, even symmetrical phoneme segments with values at both ends of approximately 0 are used, but in order to achieve this, (1) the length of each phoneme segment is constant, and (2) the length of each phoneme segment is set to the pitch period as shown in Figure 4. Add the condition that it is set sufficiently long (for example, 16 milliseconds),
It realizes the creation of phoneme pieces. By setting the phoneme segment length to be sufficiently long (for example, 16 milliseconds) in this way, the waveform of the phoneme segment calculated from the inverse Fourier transform of the spectrum envelope will have values at both ends of approximately 0. Therefore, 16 milliseconds from the excitation time point (t ₁ ) is sufficient for the reproduction processing of the phoneme waveform 12. However, the reproduction process of phoneme waveform 13 is performed at the excitation time point.
Starting from t ₂ requires multiple reproduction processing because the reproduction processing of the phoneme waveform 12 has not yet been completed. The required waveform reproduction multiplicity n is the following (1)
It is given by the formula, n = 16 milliseconds/minimum pitch period (1). In the case of normal voice, n of about 4 is sufficient. Therefore, the following description will be made assuming that n=4. Next, an ADPCM code with high coding efficiency is used as a code for reproducing the phoneme waveform, and an even symmetric waveform that requires half the amount of information is used as the phoneme waveform. A regenerator for reproducing symmetrical phoneme waveforms from a series of ADPCM codes has been proposed in Japanese Patent Application No. 185489/1983, so a detailed explanation will be omitted. In differential decoding processing such as ADPCM code reproduction processing, if the reproduction processing is terminated after 16 milliseconds (128 sample periods), the final output value is retained. Therefore, the phoneme waveform condition (Condition 1) is 0 before the excitation time point (Condition 2) and is 0 at the infinite time point.In order to satisfy the condition, the reproduction initial value must be 0 and the reproduction final output value. must be 0. Since the present invention deals with even symmetrical waveforms, if the initial reproduction value is set to 0, the final reproduction value becomes 0 and the condition is satisfied. (Embodiment of the invention) FIG. 5 shows an embodiment of the invention. Here, the term "channel" or "ch" is used to number the processes that are time-divisionally processed within one sample period. Each part in FIG. 5 is as follows. 10
0 is the 1st channel input register, 101 is the 2nd channel input register, 102 is the 3rd channel input register, 103 is the 4th channel input register, 104 and 105 are the selectors, 106 is the register that stores the ADPCM code data Ln, and 107 is the register 106. A pointer movement amount memory that outputs a pointer movement amount Dn using a value as an address, 108 an adder/subtractor, 109 a selector, 110 a selector, 111 a first channel pointer register, 112 a second channel pointer register,
113 is the 3rd channel pointer register, 114 is the 3rd channel pointer register
4ch pointer register, 115 is selector, 11
6 is a pointer limiter that limits the pointer value Pn to a specific range and outputs it as a pointer limiter value Pn'; 117 is a quantization memory that stores the quantization step value X; 118 is a shift register;
9 is an adder/subtractor, 120 is a register, 121 is a register for outputting synthesized speech, 122 is an output terminal, 1
35 is the 1st channel down counter, 136 is the 2nd channel down counter, 137 is the 3rd channel down counter,
138 is the 4th channel down counter, 139 is the 1st channel
decoder, 140 is the second channel decoder, 141 is the second channel decoder, and 141 is the second channel decoder;
3ch decoder, 142 is 4th channel decoder, 143
is the BUSY signal of each channel, 144 is the LEFT signal indicating the reproduction of the left half of the symmetrical waveform reproduction, 1
45 is a start signal for waveform reproduction, and 150 is a controller. The controller 150 is
It is connected to each circuit such as the 1st channel input register 100, the 2nd channel input register 101,..., the 3rd channel decoder 141, the 4th channel decoder 142, etc., and controls their operation, but the diagram in FIG. 5 becomes complicated. The connection details are omitted to avoid confusion. Figure 6 shows the data used in the present invention to reproduce an even symmetrical phoneme waveform, and the initial value of the pointer.
DPn, 64 ADPCM codes Ln ₁ , Ln ₂ , Ln ₃ ,...,
Consists of Ln ₆₄ . As stated in Japanese Patent Application No. 56-185489, 64 ADPCM codes are required to reproduce an even symmetrical waveform of 16 milliseconds (128 sampling periods). These 64 ADPCM codes are reproduced in forward and reverse directions to reproduce a phoneme waveform with a period of 128 samples. Further, the initial value data DPn of the pointer is data stored in the pointer at the time of starting reproduction of the symmetrical waveform. Further, FIG. 7 shows the relationship between the playback processing assigned to each channel and the time. In FIG. 7, 200 is a composite output waveform, 201
is the phoneme waveform reproduced by the reproduction process assigned to channel 1, 202 is the activation signal ST1 for channel 1, and 203 is the BUSY signal for channel 1.
BUSY1, 204 are LEFT signals for channel 1
LEFT1, hereinafter 205, 209, 213 are phoneme waveforms reproduced by the reproduction processing assigned to channels 2, 3, 4, 206, 210, 214 are activation signals ST2, ST3,
ST4, 207, 211, 215 are channel 2,
BUSY signals for 3 and 4 BUSY2, BUSY3,
BUSY4, 208, 212, 216 are LEFT signals for channels 2, 3, 4 LEFT2, LEFT3,
This is the LEFT4 signal. Embodiments of the present invention will be described in detail with reference to FIGS. 5 to 7. At time point _t1 , a starting signal ST1 is applied to the controller 150 from the outside, and the reproduction process of the phoneme waveform corresponding to FIG. 412 is started by channel 1. At this time, controller 150 sets down counter 135 to the value 128. This value corresponds to the aforementioned 16 milliseconds. (If the sampling period of the output waveform is 125 microseconds, 16 milliseconds is 128 sampling periods.) When the output of the down counter 135 is other than 0, the decoder 139 outputs "1" as the BUSY1 signal. If the output of the down counter 135 is 65 or more, "1" is output as the LEFT1 signal.
(See FIG. 7) Thereafter, data for waveform reproduction is read from the input register 100 at every sampling period (125 microseconds) and symmetrical waveform reproduction processing using the ADPCM code is performed. Each time one playback process is completed, the downtown counter 135 is decremented by one. 64 pieces
After playing the ADPCM code sequentially from the beginning in the forward direction, that is, after the 65th playback process,
The LEFT1 signal becomes "0", and ADPCM reverse reproduction processing is performed on the 64 ADPCM codes in the reverse direction starting from the last one. When the 128th reproduction process is completed in this way, the down counter 135 becomes 0 and the BUSY1 signal becomes "0". Table 1 shows that the controller 150 selector 1
The ADPCM code Ln stored in the input registers 100, 101, 102, 103 of each channel input through 04 and the values shifted down by the shift register 118 ([X], [1/2X], [1/4 X], [1/8
ADPCM waveform reproduction calculation processing (excluding pointer movement calculation) is shown.

【table】

[Table] Similarly, at time point t ₂ , start signal ST2 is applied externally.
is applied to the controller 150, and channel 2 starts the reproduction process of the phoneme waveform corresponding to FIG. 4, 13. At this time, the value 128 is set in the down counter 136, and the BUSY2 signal becomes "1". After that, the same playback process as for channel 1 is performed, and the down counter 136 becomes 0 after 128 sampling cycles.
The BUSY2 signal becomes "0". Thereafter, similar control is performed on channel ₃ from time point t3, channel ₄ from time point t4, and channel ₁ from time point t5. In this way, reproduction processing is performed using four channels and a synthesized speech output waveform is output at each sampling period, but the reproduction processing of each channel is performed in a time-division manner between one sampling time point. Here, the time relationship of processing within one sampling period is shown in FIG. 8a, and its flowchart is shown in FIGS. 8b and 8c. Processing within one sampling period is divided into five cycles, which are sequentially processed. In addition, most of the circuits required for ADPCM reproduction processing processed in each cycle are shared, and among the components shown in Figure 6, all components that do not have a channel number added to their names are common to each cycle. These are collectively referred to as the "shared part." (Cycle 1) Waveform reproduction processing assigned to channel 1 is performed using the input register 100, pointer register 111, down counter 135, decoder 139, and a common section, and the result is stored in the register 120. (Cycle 2) Waveform reproduction processing assigned to channel 2 is performed using the input register 101, pointer register 112, down counter 136, decoder 140, and a common section, and the result is stored in the register 120. (Cycle 3) Waveform reproduction processing assigned to channel 3 is performed using the input register 102, pointer register 113, down counter 137, decoder 141, and a shared section, and the result is stored in the register 120. (Cycle 4) Waveform reproduction processing assigned to channel 4 is performed using the input register 103, pointer register 114, down counter 138, decoder 142, and a shared section, and the result is stored in the register 120. (Cycle 5) The value of register 120 is stored in register 121 and used as PCM data for outputting synthesized speech. The control in the above five cycles is shown in the flowcharts of FIGS. 8b and 8c. In this way, in this embodiment, symmetric phoneme waveform reproduction processing using the ADPCM code assigned to each channel is performed by time-division processing between the constituent elements for each channel and the commonly used shared section. There is. Sequential addition means (register 120 and adder/subtractor 119) used for waveform reproduction of each channel
It also serves as an addition means for superimposing phoneme waveforms, and has the advantage of being realized by a small circuit configuration. Furthermore, by changing the interval between activation signals of each channel, the pitch period of the synthesized audio output can be easily changed. Furthermore, since each phoneme waveform starts from 0 and ends at 0, it also has the advantage of not causing any deterioration in sound quality such as waveform discontinuity due to addition of each phoneme. (Effects of the Invention) As described above, according to the present invention, it is possible to construct a speech synthesis circuit that can control the pitch period and outputs high-quality synthesized speech with a simple circuit configuration.

[Brief explanation of the drawing]

Figure 1 is a diagram showing speech waveforms, Figure 2 is a block diagram of a conventional speech synthesizer, Figure 3 is a diagram showing waveforms used when changing pitch, and Figure 4 is a diagram showing the superposition of phoneme waveforms. Explanatory diagram, Fig. 5 is a diagram showing one embodiment of the present invention, Fig. 6 is a diagram showing the data format of the phoneme waveform, and Fig. 7 is a diagram showing the phoneme waveform, ST signal, BUSY
Figure 8a shows the time relationship between the signals and the LEFT signal. Figure 8a shows the time relationship of processing within one sample cycle time. Figures 8b and c show the processing within one sample cycle time. This is a flowchart. 100...1st channel input register, 101...1st channel input register
2ch input register, 102...3rd channel input register, 103...4th channel input register, 104...
Selector, 105... Selector, 106... Register, 107... Pointer movement amount memory, 108
...adder/subtractor, 109...selector, 110...
Selector, 111...1st channel pointer register,
112...2nd channel pointer register, 113...
3rd channel pointer register, 114...4th channel pointer register, 115...Selector, 116...
Pointer limiter, 117...Quantization memory, 1
18...shift register, 119...addition/subtraction unit,
120...Register, 121...Register, 12
2... Output terminal, 135... 1st channel down counter, 136... 2nd channel down counter, 137...
...3rd channel down counter, 138...4th channel down counter, 139...1st channel decoder, 140
...2nd channel decoder, 141...3rd channel decoder,
142...4th channel decoder, 143...BUSY signal, 144...LEFT signal, 145...start signal, 150...controller.

Claims (1)

[Scope of Claims] 1. Each piece of data for phoneme waveform reproduction, including a pointer initial value and a plurality of pieces of ADPCM code data consisting of a plurality of bits, is sequentially input to a reproduction means in correspondence with a plurality of channels, and each In a speech synthesizer that reproduces even symmetrical phoneme waveforms for each channel, adds the phoneme waveforms of each channel shifted by a pitch period using an adding means, and obtains the addition result as a decoded output, the phoneme waveform of each channel is as follows: An even symmetrical phoneme waveform whose phoneme length is set to be sufficiently longer than the pitch period, has a constant time length sufficient for the final value to become 0 due to waveform attenuation, and whose initial and final values are 0. Yes, the reproduction means corresponds to each channel.
a plurality of input registers that store ADPCM code data; a pointer movement amount memory 107 that outputs pointer movement amounts according to each ADPCM code data;
An adder/subtracter 108 stores a pointer value, stores the pointer initial value as an initial value, and adds/subtracts the corresponding pointer movement amount and the previous pointer value in response to input of ADPCM code data. A plurality of pointer registers corresponding to each channel are used to update and store the results of addition and subtraction performed by the circuit 108 as new pointer values. A quantization memory (117) for outputting, a shift register 118 that can receive and store the quantization step value and shift the quantization step value, and means 11 for cumulatively adding and subtracting the output of the shift register 118.
9 and 120, and time-divisionally performs phoneme waveform reproduction processing in each channel within the sampling time, and the cumulative addition/subtraction means 119 and 120 add the phoneme waveforms of each channel shifted by pitch periods. A speech synthesizer, characterized in that the speech synthesizer also serves as the addition means.