JPH10232699A - Lpc vocoder - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reproduce an articulate and natural synthesized voice by mixing noise components with a voiced sound excitation signal and adding a secondary pulse and fluctuations. SOLUTION: A pitch extraction and voice decision part 5 extracts the pitch and voice decision value of frame data and outputs the extracted voice decision value to an encoding part 11. Further, when the voice decision value is a voiced sound, the extracted pitch is outputted as a pitch coefficient which approximates the period component of a vocal-chords wave of a frame to the encoding part 11. A mixing coefficient calculation part 6 compares the obtained pitch coefficient with respective band pitches of 3rd to 5th band data when the said voice decision value is the voiced sound. The sum of band voice decision values of band data having differences within a certain value is calculated and outputted as a mixing coefficient to the encoding part 11. A fluctuation decision part 7 extracts a fluctuation decision value for the pitch from the voice decision value, 1st and 2nd band pitches, and the frame data that a shift register 2 holds.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to improving the quality of an LPC vocoder (linear predictive speech analysis / synthesis apparatus) used for highly efficient transmission and recording of speech data.

[0002]

2. Description of the Related Art A vocoder analyzes an input speech to extract a feature parameter characterizing the speech. For example, the vocoder encodes the speech, transmits the decoded speech at a low bit rate, decodes the speech, and reproduces the speech based on the feature parameters. This is a speech analysis / synthesis device that enables highly efficient transmission and recording of original speech by synthesizing, and is based on LPC (Linear Predic
tive coder (vocoder) and the like are integrated into LSIs and widely used.

However, in the conventional vocoder, the synthesized voice is divided into frames of a fixed time length of, for example, 22.5 ms, and characteristic voices are extracted and the voice is re-synthesized in units of frames. There is a problem that it becomes a peculiar artificial thing.

[0004] For this reason, various attempts have been made for the purpose of synthesizing a more natural sound.
Japanese Patent Application No. 7-302945 discloses a "voice signal synthesizing method" (referred to as "prior art"). However, the present invention refers to a feature parameter transmitted in a frame unit with reference to a feature parameter of a previous frame, and a feature parameter between frames. By performing inter-frame interpolation so that changes continuously, more natural speech synthesis is performed.

FIG. 3 is a block diagram showing the configuration of the vocoder according to the prior art. The characteristic parameters extracted from the input voice by the analysis unit 50 of the first vocoder are transmitted via a transmission line 54, and the second The output voice is synthesized based on the characteristic parameters transmitted by the synthesis unit 60 of the vocoder. In general, both the first and second vocoders are provided with an analyzing unit and a synthesizing unit for two-way communication. However, since they are duplicated, FIG. 3 shows the synthesizing unit of the first vocoder and the analyzing unit of the second vocoder. The part is abbreviated.

[0006] The analysis unit 50 includes an analysis processing unit 51 for analyzing an input audio signal, encoders 53-1 to 53-4 for encoding the output of the analysis processing unit 51, and an encoder for each of these encoders. Multiplexer 5 for multiplexing the output and outputting to transmission line 54
2 is provided. The synthesizing section 60 performs a demultiplexer 61 for separating the multiplexed signal input from the transmission line 54, decoders 62-1 to 62-4 for decoding the separated signals, and performs an interpolation process based on the outputs of the respective decoders. An interpolation processing unit 63 to be performed and a synthesis processing unit 64 for synthesizing audio signals based on the output of the interpolation processing unit 63 are provided.

The analysis processing section 51 converts the input audio signal into 8K
After sampling at a sampling frequency of Hz and converting it into a digital signal, it is divided into frames (analysis sections) at a time interval of 22.5 ms, and for each frame, characteristic parameters relating to the signal waveform of the frame, for example, voiced sound in the example of FIG. / Extract unvoiced sound, pitch period, LSP (line spectrum pair) coefficient, and frame power. These characteristic parameters are encoded by encoders 53-1 to 53-4, multiplexed by a multiplexer 52, demultiplexed by a demultiplexer 61 via a transmission path 54, and demultiplexed by a decoder 62-53.
After decoding by 1 to 62-4, it is supplied to the interpolation processing unit 63.

Since the characteristic parameters supplied as described above are extracted in frame units in the analysis processing unit 51, they are discrete values in frame units, and the output voice synthesized by directly using this is inevitably unique to the vocoder. It becomes an artificial voice. For this reason, in the prior art shown in FIG. 3, an interpolation processing unit 63 is provided, and by interpolating the characteristic parameter value of the frame with reference to the characteristic parameter value of the previous frame so as to continuously change between frames. Processing is performed to obtain a more natural speech synthesis. For example, when the current frame and the previous frame are voiced sounds, the pitch period is interpolated so that the interval change at each pitch is continuously equal, and the line spectrum pair and the frame power of the transmitted frame are calculated. , With reference to the respective values of the previous frame, correction is made so as to change continuously at each pitch of the interpolated cycle, and a line spectrum coefficient and a frame power / 1 pitch are assigned to each pitch.

The synthesizing unit 64 generates one impulse for voiced sound and random noise for unvoiced sound at a power corresponding to the corrected frame power for each pitch period interpolated in this manner. Is passed through a filter circuit having a transfer characteristic corresponding to the value of the interpolated line spectrum pair as an excitation signal, thereby synthesizing voice waveform data at the pitch. As described above, in the prior art shown in FIG. 3, the unnaturalness due to the discontinuous change of the value of the feature parameter at the frame boundary is improved.

[0010]

However, in the above prior art, the unnaturalness due to the discontinuity of the characteristic parameter can be eliminated. However, since the voiced excitation signal is approximated by a single impulse, a high level is required. Degradation of synthesized speech due to spectrum peaks occurring in the frequency band, and unnaturalness of synthesized speech by simply approximating the voiced excitation signal with a fixed pulse train and approximating the unvoiced excitation signal with random noise during speech synthesis Performance is not improved, and frame power extracted at a fixed frame length is assigned to a synthesized frame length that expands and contracts in accordance with the interpolated pitch. .

SUMMARY OF THE INVENTION The present invention has been made to solve such a problem, and a noise component is mixed into a voiced sound excitation signal, and a secondary pulse and fluctuation are added to the signal.
An LPC vocoder with superior reproducibility can be provided by reproducing synthesized speech with excellent naturalness and by adjusting the length of a frame to be analyzed on the premise of pitch interpolation in a synthesis unit when extracting feature parameters. It is intended to be.

[0012]

According to the present invention, an audio signal is divided into frame data having a fixed time length, and each frame data is classified into at least a voiced voice or an unvoiced voice, a power, and a vocal tract used in synthesizing frame data. The characteristic coefficient including the characteristic coefficient of the equivalent filter and the pitch in the case of voiced sound is extracted, and the frame signal corresponding to the characteristic parameter is synthesized to recover the audio signal by synthesizing L.
The PC vocoder is characterized by comprising means for adding a secondary pulse to an excitation signal generated at the time of synthesizing voiced sound frame data.

Means for extracting together a mixing coefficient indicating the degree of mixing of noise components when extracting the characteristic parameters of voiced sound frame data; and adding a noise component to an excitation signal generated at the time of synthesis corresponding to the mixing coefficient. And means for adding

Further, means for extracting a fluctuation judgment value indicating the degree of pitch fluctuation at the time of extracting the feature parameter of voiced sound frame data, and an excitation signal generated at the time of synthesis corresponding to the fluctuation judgment value corresponding to the fluctuation judgment value. It is characterized by comprising means for adding a fluctuation component to the pitch interval.

Further, when synthesizing each frame data corresponding to the characteristic parameter, the characteristic parameter indicating the pitch is interpolated so as to smoothly change between frames, and the frame length of each frame data to be synthesized is adjusted accordingly. In the case where the audio signal is adjusted and restored, when extracting the power at the time of extracting the feature parameter, a means for adjusting a data length of a frame to be extracted corresponding to a frame length adjusted at the time of synthesis is provided. And

Further, when synthesizing each frame data corresponding to the characteristic parameter, the characteristic parameter indicating the pitch is interpolated so as to smoothly change between frames, and the frame length of each frame data to be synthesized is also adjusted accordingly. In the LPC vocoder that adjusts and restores the audio signal, in extracting the characteristic coefficients of the vocal tract equivalent filter used in synthesizing the frame data when extracting the feature parameters,
There is provided a means for adjusting the length of the extraction target frame data in accordance with the frame length adjusted at the time of synthesis.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. 1 and 2 are block diagrams showing an embodiment of the LPC vocoder of the present invention. FIG. 1 shows a configuration on the analysis side, and FIG. 2 shows a configuration on the synthesis side. First, extraction of characteristic parameters on the analysis side will be described with reference to FIG.

On the voice analysis side shown in FIG.
An audio input, which is an analog signal in a frequency band of 3 Hz to 3 kHz, is sampled at a sampling frequency of 8 kHz by the A / D converter 1, converted into a digital signal, and buffered in the shift register 2. In the present embodiment, the audio data is divided into frame units each having a reference frame length of 22.5 ms (180 sample periods), and the feature parameters are extracted. The shift register 2 stores the relevant frame to be analyzed and the immediately preceding frame. It holds audio data of a frame (hereinafter, referred to as the frame data and the previous frame data, respectively).

The frame data is divided into band data of a plurality of frequency bands by a filter bank 3 composed of digital filters, each of which is
And analyzed. Filter bank 3 of the present embodiment
Indicates that the frame data held in the shift register 2 is 100 to 800 Hz, 500 to 1000 Hz, 100
The data is divided into first to fifth band data of 0 Hz to 1500 Hz, 1500 Hz to 2000 Hz, and 2000 Hz or more, and output to the voice analysis unit 4. The voice analysis unit 4 performs an AMDM (AverageMagnitude Difference Functio) on each of the first to fifth band data.
The pitch is calculated by the method n), the autocorrelation coefficient and the average low-frequency energy ratio are analyzed, and another calculation of the voice judgment value, that is, the voiced sound “1” and the unvoiced sound “0” is performed.

Then, based on the band pitch and band sound judgment value of each of the five band data, a sound judgment value indicating whether voiced sound is “1” or unvoiced sound “0”, and a sound judgment value, A pitch coefficient indicating the pitch when the judgment value is a voiced sound, a mixing coefficient for determining a mixing ratio when a voiced sound is to be mixed with an unvoiced sound, and a fluctuation judgment value indicating a case where fluctuation is to be added to a pitch interval.
Are extracting two values.

Hereinafter, extraction of the above four characteristic parameters in the present embodiment will be described. First, the pitch extraction / speech determination unit 5 compares the pitches of the first and second band data and the pitch of the previous frame data determined by the pitch extraction unit 5, and among the first and second band pitches,
A pitch closer to the pitch of the previous frame data is determined as the pitch of the frame data, and the band voice determination value of the band data indicating the band pitch is set as the voice determination value of the frame. If the first and second band data have the same pitch and the band audio determination values are different, the first band audio determination value is used as the audio determination value of the frame data. In this way, the pitch extraction / speech determination unit 5 extracts the pitch and the speech determination value of the frame data, outputs the extracted speech determination value to the encoding unit 11, and furthermore, the speech determination value is “1”. In the case of (voiced sound), the extracted pitch is output to the encoding unit 11 as a pitch coefficient approximating the periodic component of the vocal cord wave of the frame.

In the mixing coefficient calculation section 6, when the voice judgment value output from the pitch extraction / voice judgment section 5 is "1",
The respective band pitches of the third, fourth, and fifth band data are compared with the pitch coefficients, and the sum of the band voice determination values (0 to 0) of the band data whose difference is within a certain value.
3) is calculated and output to the encoding unit 11 as a mixing coefficient. For example, the voice determination value of the frame data is “1”, and the pitch coefficient is 40 (5 ms at the number of samples, 8 kHz sampling frequency, 200 Hz pitch frequency).
, The third, fourth, and fifth band pitches are 42, 45, and 80, respectively, and when the band voice determination values are “1”, “1”, and “0”, respectively, If the value is 3, the mixing coefficient is 1, and if the fixed value is 5, the mixing coefficient 2 is obtained.

The fluctuation determining section 7 has a pitch fluctuation based on the voice determination value output from the pitch extracting / voice determining section 5, the first and second band pitches, and the frame data stored in the shift register 2. A fluctuation determination value indicating “1” or “0” is extracted. That is, when the audio determination value is “1” and the difference between the first and second band pitches is a fixed value, for example, 7 or more, and when the audio determination value is “0” and the frame power of the frame data is a certain value or more. In the case of, the fluctuation determination value is set to “1” as a voiced sound with fluctuation. The latter is an analysis value often seen at the time of rising of the voice. In this case, the voice determination value is corrected to “1” together with the fluctuation determination value, and the mixing coefficient is calculated again.

The above is a description of the characteristic parameters, especially the mixing coefficients,
This is an embodiment of extracting a fluctuation determination value, and the present invention is not limited to this embodiment. For example, the sampling frequency in the A / D converter 1 can be set to an appropriate value in accordance with the band of the input sound, and the reference frame length can be set to an appropriate value in consideration of the coding efficiency and the quality of the synthesized sound. In this embodiment, the frame data is divided into five bands for analysis. For example, the frame data is divided into two bands, and the lower band voice determination value and the lower band pitch are determined by the voice determination value of the frame data. , A pitch coefficient, and a mixing coefficient is, for example, a value three times as large as the band voice determination value when the band pitch of the high band is near the pitch coefficient, and the voice determination value is “0”;
In addition, the fluctuation determination value may be set to “1” only when the frame power is equal to or more than a certain value. Furthermore, regarding the method of calculating the band pitch and the method of determining voiced / unvoiced sound,
The present invention is not limited to the above-described embodiment, and other known appropriate methods can be applied.

Next, extraction of a linear spectrum pair coefficient and frame power, which are other characteristic parameters of the frame data, will be described. In the present embodiment, the analysis window adjusting unit 8 adjusts the frame length (analysis window length) to be analyzed on the premise of the inter-frame interpolation of the pitch on the synthesis side prior to the extraction of the feature parameter. It is designed to perform faithful output speech synthesis.
Hereinafter, adjustment of the frame length in the analysis window adjustment unit 8 will be described.

First, when the frame data is a voiced sound, that is, when the voice judgment value output from the pitch extraction / voice judgment unit 5 is "1", or the fluctuation judgment unit 7 corrects the voice judgment value to "1". Is described. First, in the same manner as the inter-frame interpolation of the pitch on the synthesizing side, interpolation of each pitch interval is performed so that each pitch interval of the frame continuously changes from the pitch interval of the previous frame. Specifically, a value obtained by adding the number of remaining samples (described later) at the time of determining the pre-analysis window (corrected frame) to the reference frame length is defined as the extended frame length, and the final pitch interval of the pre-analysis window is defined as Ppst (number of samples). ), When the pitch coefficient determined by the pitch extraction / voice determination unit 5 is Pref, an integer value of the extended frame length / (average of Ppst and Pref) is set as the provisional pitch number Ntp of the frame data,
For example, each pitch interval is equal difference δ = (Pref−Ppst)
/ Ntp, linearly interpolates the pitch interval of the frame into Ppst + δ, Ppst + 2δ,... In order.

Further, when the fluctuation judgment value of the fluctuation judgment unit 7 is "1" (there is fluctuation), a constant formula, for example, 0.75 + R (R Is a value of 0 to 0.5) and the fluctuation is taken into account. The pitch intervals thus obtained are integrated in order, the number of pitches giving the maximum value whose integrated value does not exceed the extended frame length is set as the analysis pitch number of the frame, and the maximum value of this integrated value is used as the Let it be the analysis window length. Using the analysis window length, the sample data for the analysis window length following the next sample value of the previous analysis window data is extracted from the previous frame data and the frame data held in the shift register 2. To be analyzed. Further, the difference between the extended frame length and the analysis window length, that is, the number of samples not analyzed in the frame data is added to the reference frame length as the remaining sample number, and the next extended frame length is calculated. . When the voice judgment value of the frame data is unvoiced sound "0", the analysis window data is obtained by adding the number of remaining samples to the frame data length. That is, in the case of a continuous unvoiced sound, the number of remaining samples is 0, and the analysis window data matches the frame data.

The correction of the analysis window length in the analysis window adjustment unit 8 and the specification of the analysis window data to be analyzed have been described above. However, the present invention is not limited to this. The analysis window length is similarly corrected by interpolating each pitch interval so that the pitch changes smoothly between three consecutive frames by storing the frame data and the frame data before and after, and specifying the analysis window data. Alternatively, the synthesis side may perform interpolation of each pitch interval in the same manner.

The line spectrum versus coefficient analysis unit 9 calculates a line spectrum versus coefficient for the speech waveform of the analysis window data specified in this way, and as a characteristic parameter for specifying a vocal tract equivalent filter at the time of speech synthesis. Encoding unit 1
Output to 1. In this embodiment, the PA
An RCOR (partial autocorrelation) coefficient is calculated and corrected so that the vocal tract equivalent filter does not diverge. In addition, the gain calculation unit 10 stores the analysis window data in the first half and the second half.
RMS (Root Mean Squa
re) Calculate the value and output it to the encoding unit 11 as a feature parameter for specifying the frame power. Note that the extraction of the feature parameter for specifying the vocal tract equivalent filter is not limited to the line spectrum pair analysis, but may be performed by another method such as cepstrum analysis.
The line spectrum pair coefficient has a feature that the inter-frame interpolation in the combining unit described later is easy.

The encoding unit 11 encodes the pitch coefficient, the frame power, the line spectrum pair coefficient, the fluctuation judgment value, the speech judgment value, and the mixing coefficient, which are analyzed and extracted as described above, as characteristic parameters of each frame data. And transmit it to the compositing side.

Next, one embodiment of the combining side will be described with reference to the block diagram of FIG. The encoded data transmitted from the analysis side is decoded by the decoding unit 21 into characteristic parameters for each frame data, that is, pitch coefficients, frame powers, line spectrum pair coefficients, fluctuation judgment values, speech judgment values, and mixed coefficients. You.

First, the speech synthesis when the speech judgment value is "1" (voiced sound) will be described. First, the pitch coefficient is input to the pitch interpolation unit 22, and the pitch interval is interpolated so as to change continuously between frames by the same method as the analysis window adjustment unit 8 in FIG. That is, in the present embodiment, the extension frame calculation unit 25 adds the number of remaining samples at the time of the previous synthesis window calculation to the reference frame length (22.5 ms = 180 samples in this embodiment) to calculate the extension frame length. The pitch interpolation unit 22 calculates the extended frame length, the final pitch interval Ppst of the previous synthesis window, and the pitch coefficient P of the frame data.
From ref, the number of provisional pitches Ntp = 2 × extended frame length /
(Ppst + Pref), pitch interval increase / decrease value δ = (Pr
ef−Ppst) / Ntp, and the pitch interval of the frame is sequentially set to Ppst + δ, Ppst + 2δ,.
Linear interpolation.

Next, when the fluctuation judgment value is "1" in the fluctuation adjusting section 23, 0.75 + R (R is 0
Multiplied by a random number of .about.0.5) to take into account fluctuations. The synthetic window calculating section 24 sequentially accumulates the pitch intervals calculated in this way, and sets the number of pitches that gives the maximum value whose integrated value does not exceed the extended frame length as the synthetic pitch number of the frame. Is the synthetic window length. Further, a value obtained by subtracting the synthesis window length from the extended frame length is output to the extended frame calculation unit 25 as a new number of remaining samples.

On the other hand, the first power interpolation unit 26 interpolates and distributes the first half value and the second half value of the frame power output from the decoding unit 21 to each pitch interval of the synthesis window, and assigns the former value to each pitch interval. The power value to be calculated is calculated so as to be continuous between frames and in the first and second half of the frame. That is, for example, when the synthesis window has one pitch, the average value of the former half value and the latter half value of the frame power is set as the power of the pitch. If the number of pitches of the synthesis window is even,
With reference to the power of the final pitch of the pre-synthesis window, the first half value is allocated to each first half pitch, and the second half value is allocated to each second half pitch after linear interpolation in the same manner as the pitch interval. When the synthesis window is composed of odd pitches, the first half and second half pitch numbers are determined so that the first half pitch number is one less than the second half pitch number, and the first half of the frame power is set to each pitch in the same manner as in the case of an even number. Distribute the value and the latter half value. This is because the larger the power in the first half, the less the unnaturalness.

According to the present invention, more natural speech synthesis is performed by adding a certain noise to the synthesis window corresponding to the mixing coefficient analyzed by the analysis unit to the excitation signal of the voiced sound. The noise generation unit 27 sets a noise value obtained from a random number to each sample value at each pitch interval of the synthesis window, and applies the noise value to a high-pass filter 2 having a characteristic of 1-bz ^-1.
By filtering with 8, the noise signal at each pitch interval of the synthesis window is set.

In the pulse generator 29, as the excitation pulse of the synthesized voice, the power value interpolated for each pitch calculated by the first power interpolator 26 at each pitch interval calculated by the synthesis window calculator 24, A pulse having an amplitude corresponding to the value obtained by multiplying the mixed coefficient decoded by the decoding unit 21 is generated one by one. For example, in this embodiment, the mixing coefficient is set to M (0 to 3), a value obtained by multiplying the power value by 0.6 + 0.1 M is set as the amplitude value of the excitation pulse, and this amplitude value is set as the last sample value of each pitch interval. This is set, filtered by a low-pass filter 30 having a characteristic of 1 + az ⁻¹ , and superimposed on a noise signal output by a high-pass filter 28 to obtain an excitation signal for each pitch.

A secondary pulse is added to the excitation signal to enhance the clarity and naturalness of the synthesized sound. Therefore, a secondary pulse is added to the excitation signal for each pitch by the secondary pulse adding unit 31. In the present embodiment, for example, the secondary pulse is added by passing through a 20th-order FIR filter, but other appropriate means may be used. Further, the excitation signal to which the secondary pulse has been added is normalized by the power normalizing section 32 so that the power is distributed by the first power interpolating section 26 at each pitch interval. As described above, the excitation signal for each pitch is obtained for the voiced sound synthesis window.

Next, a description will be given of the generation of the excitation signal of the frame when the speech judgment value output from the decoding unit 21 is "0" (unvoiced sound). When the voice determination value is “0”, the subframe division unit 33 sets the synthesis window as four subframes regardless of the pitch coefficient. Each subframe length is basically a reference frame length, that is, 45 samples obtained by dividing 180 samples into four in this embodiment. However, the pre-synthesis window is a voiced sound, and If there is, the number of remaining samples is added to the last subframe length to clear the number of remaining samples of the extended frame calculation unit 25. This is for the same reason as increasing the number of pitches in the latter half of the odd-numbered voiced sound synthesis window.

The second power interpolation unit 34 includes a decoding unit 21
, The first half value of the frame power of the combined window output to the first two subframes and the second half value to the second half subframe are interpolated and distributed so that the power value changes continuously between the subframes. The second noise generator 35 sets white noise using random numbers for each subframe based on the power value. In the present embodiment, the white noise is converted by the shaping filter 36 into a 20th-order FIR similar to that of the secondary pulse adding unit 31.
It is shaped by a filter and output as an excitation signal for each subframe of the unvoiced sound synthesis window. The excitation signal thus obtained is supplied to the secondary pulse adding unit 31 in the voiced sound synthesis window.
In the unvoiced sound synthesis window, the data is output from the shaping filter 36 for each subframe, and is sequentially input to the synthesis filter 37 having a transfer characteristic equivalent to the vocal tract.

The line spectrum pair interpolating unit 38 interpolates the line spectrum pair coefficient of each frame output from the decoding unit 21 at every pitch interval or every subframe so as to change continuously between the synthesis windows. The parameter is converted into a parameter corresponding to the synthesis filter 37 by the α parameter conversion unit 39, and the transfer characteristic is defined for each pitch or each subframe. In the present embodiment, the linear spectrum pair coefficient of the frame data is linearly interpolated so as to change continuously for each pitch in the voiced sound synthesis window and for each subframe in the unvoiced sound synthesis window, similarly to the power interpolation unit 26. Later, linear prediction parameter (α parameter) for each pitch or subframe
By setting the characteristic value of the synthesis filter 37 composed of a voice synthesis filter by the linear prediction method (LPC) for each pitch and each subframe in synchronization with the sequentially input excitation signal, the voice data is converted. Combined. The output of the synthesis filter 37 is converted into an analog signal by the D / A converter 40 and output as a synthesized voice.

The LPC vocoder of the present embodiment is constructed as described above. By mixing a voiced excitation signal with a noise component and adding a secondary pulse and fluctuation, it is possible to obtain a clear and natural high quality signal. Synthesized speech can be played. In addition, when extracting the characteristic parameters, the synthesis target speech can be reproduced with higher fidelity by adjusting the length of the frame to be analyzed on the premise of pitch interpolation in the synthesizer.

In the present embodiment, mixing of noise components in the voiced sound excitation signal, addition of secondary pulses, addition of fluctuations, extraction of vocal tract equivalent filter coefficients on the premise of pitch interpolation during synthesis, and frame extraction Although power is extracted, the present invention is not limited to this, and these units may be configured alone or in combination as needed.

[0043]

As described above, according to the LPC vocoder of the present invention, a noise component is mixed with a voiced sound excitation signal, and a secondary pulse and fluctuation are added, so that a clear and natural sound is obtained. High quality synthesized speech can be reproduced. In addition, in extracting the characteristic parameters, by adjusting the length of the frame to be analyzed on the assumption that the pitch is interpolated in the synthesizing unit, there is an effect that a synthesized voice with excellent reproducibility can be reproduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a configuration on an analysis side of an LPC vocoder of the present invention.

FIG. 2 is a block diagram showing an embodiment of a configuration on the synthesis side of the LPC vocoder of the present invention.

FIG. 3 is a block diagram illustrating an example of a configuration of a conventional LPC vocoder.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 A / D converter 2 shift register 3 filter bank 4 voice analysis unit 5 pitch extraction / voice determination unit 6 mixing coefficient calculation unit 7 fluctuation determination unit 8 analysis window adjustment unit 9 line spectrum vs. coefficient analysis unit 10 gain calculation unit 11 encoding Unit 21 decoding unit 22 pitch interpolation unit 23 fluctuation adjustment unit 24 synthesis window calculation unit 25 extended frame calculation unit 26 first power interpolation unit 27 first noise generation unit 28 high-pass filter 29 pulse generation unit 30 low-pass filter 31 secondary pulse addition Unit 32 power normalization unit 33 subframe division unit 34 second power interpolation unit 35 second noise generation unit 36 shaping filter 37 synthesis filter 38 line spectrum pair interpolation unit 39 α parameter conversion unit 40 D / A converter 50 analysis unit 51 analysis Processing unit 52 Multiplexer 53-1 to 53-4 Code Vessel 54 transmission line 60 combining unit 61 demultiplexer 62-1 to 62-4 decoder 63 interpolation processing section 64 synthesis processing unit

Claims (9)

[Claims] An audio signal is divided into frame data having a predetermined time length, and at least a voiced sound,
By extracting the unvoiced sound, the power, the characteristic coefficient of the vocal tract equivalent filter used in synthesizing the frame data, and the characteristic parameter including the pitch in the case of voiced sound, and synthesizing each frame data corresponding to this characteristic parameter An LPC vocoder for restoring the audio signal, comprising: means for adding a secondary pulse to an excitation signal generated at the time of synthesizing voiced frame data.
PC vocoder. 2. An audio signal is divided into frame data having a predetermined time length, and at least a voiced sound,
By extracting the unvoiced sound, the power, the characteristic coefficient of the vocal tract equivalent filter used in synthesizing the frame data, and the characteristic parameter including the pitch in the case of voiced sound, and synthesizing each frame data corresponding to this characteristic parameter An LPC vocoder for restoring the audio signal, wherein at the time of extracting the characteristic parameters of the voiced sound frame data, a means for additionally extracting a mixing coefficient indicating a degree of mixing of noise components; Means for adding a noise component to the excitation signal to be generated.
C vocoder. 3. An audio signal is divided into frame data having a predetermined time length, and at least a voiced sound,
By extracting the unvoiced sound, the power, the characteristic coefficient of the vocal tract equivalent filter used in synthesizing the frame data, and the characteristic parameter including the pitch in the case of voiced sound, and synthesizing each frame data corresponding to this characteristic parameter An LPC vocoder for restoring the audio signal, wherein at the time of extracting the characteristic parameter of the voiced frame data, means for extracting together a fluctuation determination value indicating the degree of pitch fluctuation; and synthesizing in response to the fluctuation determination value. Means for adding a fluctuation component to the pitch interval of the excitation signal generated at the time. 4. An audio signal is divided into frame data having a predetermined time length, and at least a voiced sound,
Unvoiced sound, power, characteristic parameters including the pitch of the characteristic coefficient of the vocal tract equivalent filter used when synthesizing frame data and voiced sound, when synthesizing each frame data corresponding to this characteristic parameter, LP for interpolating the characteristic parameter indicating the pitch so as to smoothly change between frames, adjusting the frame length of each frame data to be synthesized in accordance with the interpolation, and restoring the audio signal
An LPC vocoder, comprising: a C vocoder, comprising means for adjusting the data length of a frame to be extracted in accordance with a frame length adjusted at the time of synthesis when extracting the power when extracting the feature parameter. 5. An audio signal is divided into frame data having a fixed time length, and at least a voiced sound,
Unvoiced sound, power, characteristic parameters including the pitch of the characteristic coefficient of the vocal tract equivalent filter used when synthesizing frame data and voiced sound, when synthesizing each frame data corresponding to this characteristic parameter, LP for interpolating the characteristic parameter indicating the pitch so as to smoothly change between frames, adjusting the frame length of each frame data to be synthesized in accordance with the interpolation, and restoring the audio signal
In the C vocoder, there is provided a means for adjusting the extraction target frame data length corresponding to the frame length adjusted at the time of synthesizing the characteristic coefficient of the vocal tract equivalent filter used at the time of synthesizing the frame data at the time of extracting the feature parameter. An LPC vocoder characterized in that: 6. The mixing coefficient is obtained by dividing the frame data to be extracted into a plurality of band data, and comparing and extracting respective pitches, voiced sounds, and unvoiced sounds obtained by analyzing the respective band data. The LPC vocoder according to claim 2, characterized in that: 7. The plurality of band data are five band data, and the extraction target frame is extracted from each of pitches, voiced sounds, and unvoiced sounds obtained by analyzing band data of two lower bands. It discriminates between voiced sound and unvoiced sound of the data, and when it is determined that the extraction target frame data is a voiced sound, each of the five band data obtained by analyzing the band data of the three higher bands is obtained. Pitch,
7. The LPC vocoder according to claim 6, wherein the mixing coefficient is extracted from voiced sound and unvoiced sound. 8. The fluctuation determination value is calculated by dividing the frame data to be extracted into a plurality of band data, and comparing and judging respective pitches obtained by analyzing each band data. The LPC vocoder according to claim 3. 9. The fluctuation determination value is calculated by determining that the extraction target frame data has a characteristic of unvoiced sound and that the frame power is a voiced sound with fluctuation when the frame power is equal to or more than a predetermined value. The LPC vocoder according to claim 3, characterized in that: