JP2008139573A - Vocal quality conversion method, vocal quality conversion program and vocal quality conversion device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a method, a program and a device of vocal quality conversion capable of giving variation to synthesis voice, by converting voice sound to voiceless sound. <P>SOLUTION: The method for converting voice sound to voiceless sound comprises : an input step of inputting an original waveform; a prediction analysis step of predicting a transfer function from the original waveform which is input in the input step; a residual signal extracting step of extracting a residual signal by using output of the original waveform and the prediction analysis step; a white noise generating step of outputting a white noise signal corresponding to power of the residual signal; and a voice synthesizing step of performing voice synthesis based on the output of the white noise generating step and the transfer function which is predicted in the prediction analysis step. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a method, a program, and an apparatus for converting voiced sound to unvoiced sound.

Conventionally, as a technique for converting a whispering voice (unvoiced sound) into a normal voice (voiced sound), “the whispering voice analysis means 12 analyzes the whispering voice in a large amount of learning data and extracts the whispering spectrum information. To do. The normal voice analysis means 13 analyzes normal voice in the learning data and extracts normal spectrum information. The mapping function estimation means 14 estimates and stores a mapping function from a large amount of whisper spectrum information and normal spectrum information pairs. When a whispering voice is input, the input voice analyzing means 11 analyzes the voice and extracts input spectrum information. Then, the spectrum converting means 15 converts the input spectrum information into converted spectrum information using a mapping function. The voice synthesizer 16 synthesizes a normal voice from the converted spectrum information and outputs it. Is proposed (Patent Document 1).
JP 10-254473 A (summary)

However, in the above prior art, although the unvoiced sound can be made voiced, the voiced sound cannot be made unvoiced.
Therefore, there has been a demand for a voice quality conversion method, program, and apparatus that can give variations to synthesized speech by converting voiced sound to unvoiced sound.

The voice quality conversion method according to the present invention includes:
A method for converting voiced sound to unvoiced sound,
An input step for inputting an original sound waveform;
A predictive analysis step of predicting a transfer function from the original sound waveform input in the input step;
A residual signal extraction step for extracting a residual signal using the original sound waveform and the output of the prediction analysis step;
A white noise generation step for outputting a white noise signal;
A speech synthesis step of performing speech synthesis based on the output of the white noise generation step and the transfer function predicted in the prediction analysis step;
It is characterized by having.

  According to the voice quality conversion method of the present invention, voiced sound can be converted into unvoiced sound and output, so that a new variation of “sounds like a whisper” can be given to the synthesized voice output.

Embodiment 1 FIG.
FIG. 1 is a functional block diagram of voice quality conversion apparatus 100 according to Embodiment 1 of the present invention.
The voice quality conversion apparatus 100 includes a waveform storage unit 101, an LPC coefficient calculation unit 102, a residual signal extraction unit 103, a white noise generator 104, a power calculation unit 105, an amplitude conversion unit 106, and a voice synthesis unit 107.
The waveform accumulating unit 101 receives the original sound waveform data and accumulates it as digital data of 16 kHz, 16 bits, for example.
The LPC coefficient calculation unit 102 reads the original sound waveform data from the waveform storage unit 101, performs LPC analysis (linear prediction analysis), and calculates an LPC coefficient. The analysis parameters are, for example, analysis window length = 20 ms, analysis order = 15th order, and the like.
The residual signal extraction unit 103 extracts a residual signal using an LPC coefficient analysis filter.
The white noise generator 104 generates white noise having an arbitrary time length.
The power calculation unit 105 calculates the average amplitude ratio of the two signal series data from the function of calculating the average power (square mean square root) of the given signal series data and the average power of the two signal series data. And having a function.
The amplitude converter 106 performs an amplitude conversion operation using the given signal series data and the amplitude ratio value.
The voice synthesis unit 107 performs voice synthesis using an LPC synthesis filter.

The waveform storage unit 101, LPC coefficient calculation unit 102, residual signal extraction unit 103, white noise generator 104, power calculation unit 105, amplitude conversion unit 106, and speech synthesis unit 107 use hardware such as a circuit device. Alternatively, it may be realized as software including a buffer for waveform data input / output.
When implemented as software, a program that realizes the functions of these units is stored in an HDD (Hard Disk Drive) or the like, and an arithmetic unit such as a microcomputer or CPU reads the program, and the functions of the units according to the instructions of the program Is configured to execute processing corresponding to the above. The processing flow may be configured similarly to the processing flow shown in FIG.
It should be noted that the same applies to the following embodiments.

  Here, before explaining the detailed configuration and operation of the voice quality conversion device 100, the characteristics of the voiced sound and the unvoiced sound will be described in order to facilitate understanding of the contents of the processing performed by the voice quality conversion device 100.

Voiced sound is sound accompanied by vocal cord vibration. In the voice signal waveform of voiced sound, a periodicity called pitch corresponding to this vocal cord vibration appears.
On the other hand, unvoiced sound is sound that does not involve vocal cord vibration. Since the voiceless sound is not accompanied by vocal cord vibration, the voice signal does not show the periodicity of vibration caused by the voice signal.

Usually, the utterance includes many voiced sounds including vowels. On the other hand, a whisper is a voice that does not include a voiced sound. Therefore, the periodicity due to the pitch does not appear in the whistling voice signal waveform.
For example, even for the same “Konchiwa” voice, the waveform of the voice signal differs greatly depending on the presence or absence of periodicity depending on the pitch between the normal utterance and the whispering voice. However, since both voices are common in that they can be heard by the human ear, they have a common sound quality feature from the viewpoint of phonology. That is, it can be said that normal speech and whispering are similar and non-speech sounds.

The present invention focuses on the above characteristics of voiced sound and unvoiced sound, and removes the periodicity due to the pitch described above from voiced sound, thereby converting a voiced method, program, and apparatus that can convert voiced sound into unvoiced sound. It is what we propose.
By converting a voiced sound to an unvoiced sound, the effect of converting a normal uttered voice to a whispered voice can be obtained in the human ear as a result. As a result, a “whispering voice” can be generated by voice synthesis, so that a new variation of “whispering voice” is created in the synthesized voice, and the range of expression of the synthesized voice is widened.

  The following will return to the detailed description of the voice quality conversion device 100.

FIG. 2 is a block diagram showing a specific configuration of portions corresponding to the LPC coefficient calculation unit 102 to the residual signal extraction unit 103.
The LPC coefficient calculation unit 102 receives data corresponding to an original sound waveform signal having a predetermined time length from the waveform storage unit 101, and uses a linear prediction coefficient (in this case, assuming n = 15) with a predetermined order n. LPC coefficient) α1 to αn are estimated. Obtaining the LPC coefficients α1 to αn corresponds to obtaining the transfer function of the vocal tract on the time axis.
Data corresponding to the residual signal can be obtained from the original sound waveform signal data and the predicted waveform signal data in which the original sound waveform is predicted using the obtained LPC coefficients α1 to αn.
The blocks for which the LPC coefficients α1 to αn are calculated correspond to the LPC coefficient calculation unit 102 in FIG. Further, the part for which the residual signal is obtained from the original sound waveform signal and the predicted waveform signal corresponds to the residual signal extraction unit 103.

  Next, the operation of the voice quality conversion device 100 will be described step by step.

(1) Accumulation of original sound waveform data The waveform accumulation unit 101 receives the original sound waveform data and temporarily accumulates it as digital data of 16 kHz, 16 bits, for example. The accumulated original sound waveform data is output to the LPC coefficient calculation unit 102 and the residual signal extraction unit 103.
The storage granularity may be configured to store all of the original sound waveform data, or to receive data of a predetermined frame length and sequentially output it to the LPC coefficient calculation unit 102 and the residual signal extraction unit 103. It may be configured to repeat until the end of data.
Although not shown in FIG. 1, an input unit for receiving the original sound waveform data is provided as necessary. As the configuration of the input unit, for example, a wired digital or analog interface or a network interface such as a LAN interface can be considered.

(2) Calculation of LPC coefficient The LPC coefficient calculation unit 102 receives original sound waveform data having a predetermined frame length from the waveform storage unit 101, performs LPC analysis using the configuration shown in FIG. αn is calculated.
Predicted waveform signal data in which the original sound waveform is predicted using the obtained LPC coefficients α1 to αn is output to the residual signal extraction unit 103.
Note that the LPC analysis is performed with a predetermined frame length regardless of the granularity of the waveform accumulating unit 101 that accumulates the original sound waveform data, so that it is received from the waveform accumulating unit 101 when performing one LPC analysis. The original sound waveform data may correspond to the frame length necessary for the LPC analysis.
The execution of the LPC analysis may be performed sequentially every time the analysis for the frame length to be analyzed at once is completed, or even if the processing after the subsequent step (3) is completed. Good. Which method is used depends on the calculation speed, the method of implementing the synchronization processing of the audio output buffer, and the like, and therefore an appropriate method may be selected as appropriate.

(3) Extraction of residual signal The residual signal extraction unit 103 extracts residual signal data using the original sound waveform data received from the waveform storage unit 101 and the predicted waveform data output from the LPC coefficient calculation unit 102. .
The obtained residual signal data is output to the power calculation unit 105.

(4) White Noise Output The white noise generator 104 outputs white noise waveform data having an arbitrary time length to the power calculator 105 and the amplitude converter 106.
The output timing may be the time when the residual signal extraction unit 103 outputs the residual signal data to the power calculation unit 105, or the white noise may be continuously output continuously.
The time length of the white noise waveform data will be described in the next step (5).

(5) Calculation of Average Power Ratio The power calculation unit 105 calculates the average power (= Pr) of the residual signal from the residual signal data output from the residual signal extraction unit 103, and the white noise generator 104 outputs The average power (= Pn) of the obtained white noise waveform data is calculated.
The time length of data at the time of calculation is adjusted to the time length when the LPC coefficient calculation unit 102 performs LPC analysis. The calculation method may be a mean square of each sample value, or may be obtained by smoothing using the average value of several previous frames.
Next, the power calculation unit 105 calculates the average amplitude ratio (= Pr / Pn) of the two signal series data from the average power of the two signal series data.
The obtained average amplitude ratio is output to the amplitude converter 106.

(6) White Noise Amplitude Conversion The amplitude conversion unit 106 uses the average amplitude ratio value (= Pr / Pn) received from the power calculation unit 105 and the white noise waveform data received from the white noise generator 104. Performs amplitude conversion calculation. Specifically, the power scale is adjusted by multiplying each sample value of the white noise waveform data by the average amplitude ratio to obtain new noise waveform data.
The obtained noise waveform data subjected to amplitude conversion is output to the speech synthesizer 107.

(7) Speech Synthesis The speech synthesis unit 107 uses an LPC synthesis filter configured using the amplitude waveform-converted noise waveform data received from the amplitude conversion unit 106 and the LPC coefficients α1 to αn calculated by the LPC coefficient calculation unit 102. Voice synthesis.
The synthesized speech waveform data is the final output of the voice quality conversion device 100.

The synthesized speech obtained by the above processing has the above-described periodicity of the pitch component removed by converting the residual signal into white noise. That is, it is heard to the human ear as if the original sound waveform which was originally voiced sound is converted to an unvoiced sound like a whisper.
Since the LPC synthesis filter used by the speech synthesizer 107 is configured using the LPC coefficients α1 to αn calculated by the LPC coefficient calculator 102, the phonological property of the original sound waveform is maintained. If “Konchiwa” is a voiced sound, the synthesized speech after conversion is heard as “Konichiwa” of whispering voice.

In the first embodiment, a method for predicting the transfer function of the vocal tract by LPC analysis is used. However, it is not always necessary to use LPC analysis. For example, a PARCOR (partial autocorrelation) coefficient or an LSP (line spectrum pair) is used. A method using a coefficient may be used.
That is, it is only necessary to maintain the phoneme of the original sound waveform, and the method for predicting the transfer function is not limited to LPC analysis.
The same applies to the following embodiments.

As described above, according to the first embodiment, since voiced sound can be converted into unvoiced sound and output, a new variation of “sounds like a whisper” can be given to the synthesized voice output. it can.
The increase in the variation of synthesized speech leads to making the user interface by speech more familiar to humans, and it can be expected that this will be applied as a man-machine interface in various devices that come in contact with everyday life.

Embodiment 2. FIG.
In the second embodiment of the present invention, a configuration of a voice quality conversion apparatus that can insert a silent part at an arbitrary timing will be described.

FIG. 3 is a functional block diagram of voice quality conversion apparatus 100 according to Embodiment 2 of the present invention.
The voice quality conversion apparatus 100 according to the second embodiment further includes a silence insertion unit 108 on the output side of the speech synthesis unit 107. Other configurations are the same as those described in FIG. 1 of the first embodiment, and thus the same reference numerals are given and description thereof is omitted.

The silence insertion unit 108 receives the synthesized voice waveform data subjected to the voice quality conversion process from the voice synthesis unit 107, receives the mora information and the conversion rule information from the outside of the voice quality conversion device 100, and uses these information to make the synthesized voice. Insert silence and output.
The mora information and the conversion rule information will be described with reference to FIG.

FIG. 4 explains mora information and conversion rule information.
A mora is a segmental unit of a sound having a certain time length in terms of phoneme, and corresponds to a “beat” of a sound. That is, the mora information is time information related to the beat of the generated voice. In general, sound, repellent sound, and long sound are counted as one beat even if no sound is produced.
For example, the original sound “Okiden-chan” ((1) in FIG. 4) is pronounced in seven beats, “O”, “Ki”, “De”, “N”, “Ki”, “Cha”, “N”.
The mora information represents on the time axis whether the beat breaks are present in the pronunciation sound. For example, in the case where the above-mentioned “Oidenki-chan” is pronounced in 0.7 seconds, there is information such that there are six breaks at equal intervals every 0.1 second ((2 in FIG. 4). )).
The silent insertion unit 108 is given this mora information and can know at which timing the beat should be divided.

The conversion rule information represents a rule of how many seconds of silence are inserted in each beat segment represented by the mora information.
For example, as shown in FIG. 4 (3), a 0.06 second silence is inserted at the end of the first and fourth beats, and 0.04 seconds at the end of the second, third, and fifth beats. It is conceivable to set a silence insertion rule for each individual beat, such as inserting a silent part.
Alternatively, a time corresponding to 1 mora is determined in advance, and “a silence part for 1 mora is inserted after completion for all mora.” Or “silence for 2 mora after completion for all mora. It can also be determined on a rule basis, such as “Insert part, but not in the second mora from the end.”

The silence inserting unit 108 receives the synthesized voice waveform data subjected to the voice quality conversion processing from the voice synthesizing unit 107, inserts the silence part into the part determined by the above-described mora information and the conversion rule information, and outputs it.
Furthermore, in order to prevent noise, amplitude conversion processing such as fade-in / fade-out may be added before and after the silent part is inserted. Thereby, even if a silence part is inserted, a smooth converted voice can be obtained.
The output of the silence insertion unit 108 is the final output of the voice quality conversion device 100.

In the second embodiment, the conversion rule is described as being given from the outside of the voice quality conversion apparatus 100. However, a storage unit may be provided inside the voice quality conversion apparatus 100 and stored in advance therein.
Since the mora information is information representing the beat of the original sound waveform, the information cannot be provided unless it is the side that provides the original sound waveform, but it is possible to know the original sound in advance because the content of the original sound is limited. If possible, it may be provided inside the voice quality conversion device 100.
As a method for receiving mora information and conversion rule information from the outside, for example, an operation unit may be provided in the voice quality conversion device 100 to allow the user to directly input the information, or transmission / reception means via a network such as a LAN interface may be used. It may be acquired. Or you may read from the memory | storage means provided outside the voice quality conversion apparatus 100. FIG.
Furthermore, when the original sound is synthesized speech, phoneme duration information and the like are obtained in the process of speech synthesis. In this case, only conversion rule information needs to be supplied from the outside.

  In the second embodiment, the silence insertion unit 108 is provided in the voice quality conversion device 100. However, the voice quality conversion device 100 has the same configuration as that of the first embodiment, and a device corresponding to the silence insertion unit 108 is a voice quality conversion device 100. Even if it is configured to be connected to the output terminal, the same effect as in the second embodiment can be obtained.

As described above, according to the second embodiment, since it is possible to provide an auditory effect as if a voice is isolated by inserting a silent part into a voice that has been silenced, the range of expression of the voice can be increased. More than one.
Further, by applying a process such as fade-in / fade-out to the waveform before and after the silent section is inserted, the sound is not unnaturally interrupted before and after the silent section is inserted, and the human hearing can be heard smoothly and less burdened.

Embodiment 3 FIG.
In the third embodiment of the present invention, a configuration of a voice quality conversion apparatus capable of making an arbitrary part longer sound will be described.
In this case, the longer sound is to convert the voice of “Oidenki-chan” to “Oki-Den-Ki-Chan”.

FIG. 5 is a functional block diagram of voice quality conversion apparatus 100 according to the third embodiment. Only differences from FIG. 1 described in the first embodiment will be described.
Voice quality conversion apparatus 100 according to Embodiment 3 includes LPC coefficient complementing section 109.
The LPC coefficients α1 to αn calculated by the LPC coefficient calculation unit 102 are output to the LPC coefficient complementing unit 109.
The LPC coefficient complementing unit 109 receives the LPC coefficients α1 to αn, the mora information, and the conversion rule information, and performs calculations described later to supplement the LPC coefficients α1 to αn on the time axis.
The white noise generator 104 receives the mora information and the conversion rule information, and outputs white noise signal data corresponding to the length of time after making the sound longer.

  Next, the details of the LPC coefficient α1-αn complementing process performed by the LPC coefficient complementing unit 109 will be described.

The LPC coefficient complementing unit 109 extends the input LPC coefficients α1 to αn in the time axis direction. For example, let L1 coefficients at a certain time t be α1 (t),..., Αn (t).
If t is given at intervals of 5 ms and the duration is 100 ms, the LPC coefficient of the phoneme is expressed by the following equation (Equation 1).
α1 (t),..., αn (t) (Equation 1)
t = 0, 5, 10,..., 100

When this time length is extended to 200 ms, which is twice as long, conversion is performed as follows.
(1) Simple extension First, it is simply extended twice on the time axis. The PC coefficient after stretching is expressed by the following formula (Formula 2).
α1 (t),..., αn (t) (Formula 2)
t = 0, 10, 20,..., 200
(2) In Equation 2, since the LPC coefficient is 10 ms apart and there is no value corresponding to t = 5, 15,..., 195, it is supplemented by the following Equation (Equation 3).
αn (t) = (αn (t−5) + αn (t + 5)) / 2 (Expression 3)
t = 5, 15, 25, ..., 195

With the above complement processing, LPC coefficients at intervals of 5 ms can be obtained even if they are doubled on the time axis.
Specifically, how much time is extended is determined by the mora information and the conversion rule information. That is, the information may be extended so as to correspond to the time length after the original sound is made longer. The same is true for white noise.

The overall operation of voice quality conversion apparatus 100 according to the third embodiment is substantially the same as that described in the first embodiment. The difference is that, as described above, the output of the white noise generator 104 is based on the mora information and the conversion rule information, and the LPC coefficient after the speech synthesis unit 107 is complemented by the LPC coefficient complementing unit 109 is used. This is the point of voice synthesis.
By performing the above processing, a devoiced synthesized speech in which the part defined by the mora information and the conversion rule information is lengthened is obtained.

  The voice quality conversion apparatus 100 according to the third embodiment increases the sound length of the corresponding mora by using the spectrum parameter with the extension complemented on the time axis. For this reason, it is preferable to use a PARCOR coefficient or an LSP coefficient which generally has better complementary characteristics than the LPC coefficient.

  As described above, according to the third embodiment, since the devoiced synthesized voice can be made longer, variations in the tone of the synthesized voice are further expanded.

Embodiment 4 FIG.
In the fourth embodiment of the present invention, a configuration of a voice quality conversion apparatus that determines whether a voiced sound or an unvoiced sound is generated for each frame of the original sound waveform and performs the unvoiced process only on the frame of the voiced sound will be described.

FIG. 6 is a functional block diagram of voice quality conversion apparatus 100 according to the fourth embodiment.
The configuration of voice quality conversion apparatus 100 in FIG. 9 is substantially the same as that in FIG. 1 of the first embodiment, except that voiced / unvoiced determination unit 110 is newly provided between residual signal extraction unit 103 and power calculation unit 105. However, it is mainly different.
The residual signal extraction unit 103 is configured to output both the original sound waveform data and the residual signal waveform data to the voiced / unvoiced determination unit 110.
The voiced / unvoiced determination unit 110 receives the original sound waveform data or the residual signal waveform data from the residual signal extraction unit 103, obtains an autocorrelation function for each frame, and determines whether the frame is voiced or unvoiced. Determine.
If the frame is a voiced sound, the residual signal waveform data received from the residual signal extraction unit 103 is output to the power calculation unit 105. The subsequent processing is the same as in the first embodiment.
If the frame is an unvoiced sound, no further devoicing processing is required, and the original sound waveform data is output to the speech synthesizer 107 as it is. The speech synthesizer 107 outputs the received original sound waveform data as it is.

  In the fourth embodiment, the case where the configuration similar to that of FIG. 1 of the first embodiment is provided has been described. However, even when the configuration of the other embodiments is provided, the residual signal extraction unit 103 and the power calculation unit 105 are configured. A voiced / unvoiced determination unit 110 can be provided between them to perform the same processing.

As described above, according to the fourth embodiment, only the voiced sound frame is subjected to the devoicing conversion process, so that the sound that originally does not need to be converted is further converted to destroy the sound quality. Nothing will happen.
In addition, since conversion processing is performed only when necessary, it is possible to configure a voice quality conversion device using a computing device with high conversion efficiency on machine resources and low processing capacity, and downsizing and cost reduction of the entire device Contribute to Even when configured as software, the conversion process can be executed by a CPU or the like having a low processing capability, so that the same effect can be obtained.

Embodiment 5. FIG.
In the fifth embodiment of the present invention, a configuration of a voice quality conversion apparatus capable of mixing and outputting an original sound waveform and a synthesized waveform will be described.

FIG. 7 is a functional block diagram of voice quality conversion apparatus 100 according to the fifth embodiment.
The configuration of voice quality conversion apparatus 100 in FIG. 10 is substantially the same as that in FIG. 1 of the first embodiment, except that voiced / voiceless mixing unit 111 is newly provided on the output side of speech synthesis unit 107.
The voiced / voiceless mixing unit 111 receives the original sound waveform data from the waveform storage unit 101 and also receives the synthesized speech waveform data from the speech synthesis unit 107. Next, each received data is mixed and output by the following equation (Equation 4). The output of the voiced / voiceless mixing unit 111 is the final output of the voice quality conversion device 100.
Sn = βOn + (1-β) UVn (Formula 4)
Sn: Mixed speech waveform On: Original sound waveform UVn: Synthetic speech waveform (unvoiced sound waveform)
β: Mixing coefficient (0 <= β <= 1)
Since the degree of mixing of the original sound (voiced sound) and the synthesized sound (unvoiced sound) is variously determined by setting the mixing coefficient β, there is an effect that the variation of the synthesized sound is widened.
Note that the value of the mixing coefficient β may be variable. The method of changing the value may be a method that changes randomly with time, or a method in which the user inputs a set value via a network or the like.

In the first to fifth embodiments described above, the source of the original sound waveform may be normal voice data that is uttered, or may be synthetic speech itself.
In the latter case, since duration information is obtained in the process of speech synthesis, mora information can be provided in the process, but in the former case, it is necessary to separately supply the mora information from the outside. As the supply means, as described in the second embodiment, direct input or input via a network can be considered.
When the source of the original sound waveform is synthesized speech, the speech synthesizer and the voice quality conversion device according to the present invention are connected via a network or the like, and the original sound waveform is directly input to the voice quality conversion device. You may comprise as follows. Furthermore, the mora information can be directly supplied from the speech synthesizer to the voice quality conversion device.

The “conversion rule information” described in the second to third embodiments can be automatically generated by the voice quality conversion device 100 itself in addition to direct input or input via a network. For example, the following examples can be considered.
(1) Each beat is expanded N times (for example, N = 2), and only the last two mora are left as they are.
・ "Okidenki-san"->"O ・ ・ ki ・ ・ で ・ ・ ん ・ ・ き ・ ・ さ ん"
・ "Okinenki-chan"->"Okiki-de-nkikiki
(2) Extend only the second beat from the end. The length of the extension is 1 mora.
・ "Okidenki-san"->"Okidenkisan"
・ "Saito-san"->"Saito-san"

Note that the reason why the unvoiced sound is easily deformed is that the unvoiced sound is easily deformed by performing a modification process such as inserting a silent part or making the sound longer. This is because unvoiced sounds are less likely to have discontinuities such as pitch synchronization inherent to voiced sounds.
Thus, the voice quality conversion apparatus according to the present invention also has a secondary effect of facilitating the deformation of the voice.

3 is a functional block diagram of voice quality conversion apparatus 100 according to Embodiment 1. FIG. 3 is a block diagram illustrating a specific configuration of a portion corresponding to an LPC coefficient calculation unit 102 to a residual signal extraction unit 103. FIG. 6 is a functional block diagram of a voice quality conversion device 100 according to Embodiment 2. FIG. This explains the mora information and the conversion rule information. FIG. 9 is a functional block diagram of a voice quality conversion device 100 according to a third embodiment. FIG. 10 is a functional block diagram of a voice quality conversion device 100 according to a fourth embodiment. FIG. 10 is a functional block diagram of a voice quality conversion device 100 according to a fifth embodiment.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 100 Voice quality conversion apparatus, 101 Waveform storage part, 102 LPC coefficient calculation part, 103 Residual signal extraction part, 104 White noise generator, 105 Power calculation part, 106 Amplitude conversion part, 107 Speech synthesizer, 108 Silence insertion part, 109 LPC coefficient complementing unit, 110 voiced / unvoiced determining unit, 111 voiced / unvoiced mixing unit.

Claims (15)

A method for converting voiced sound to unvoiced sound,
An input step for inputting an original sound waveform;
A predictive analysis step of predicting a transfer function from the original sound waveform input in the input step;
A residual signal extraction step for extracting a residual signal using the original sound waveform and the output of the prediction analysis step;
A white noise generation step of outputting a white noise signal corresponding to the power of the residual signal;
A speech synthesis step of performing speech synthesis based on the output of the white noise generation step and the transfer function predicted in the prediction analysis step;
A voice quality conversion method characterized by comprising: Receiving mora information and conversion rule information for inserting silence into the voice using the mora information;
The voice quality conversion method according to claim 1, further comprising a silent insertion step of inserting a silent part into the voice synthesized in the voice synthesis step using these pieces of information. In the silent insertion step,
The voice quality conversion method according to claim 2, wherein fade-in processing or fade-out processing is performed on the sound before and after the silent portion is inserted. Receives mora information and conversion rule information for expanding the voice on the time axis using the mora information,
The transfer function predicted in the prediction analysis step has a prediction complementing step of complementing on the time axis using these pieces of information,
In the white noise generation step,
The mora information and the conversion rule information are received, and a white noise signal generated using the information is output,
In the speech synthesis step,
The voice quality conversion method according to claim 1, wherein speech synthesis is performed based on an output of the white noise generation step and a transfer function complemented by the prediction complement step. For each frame of the original sound waveform or the residual signal, a determination step of determining whether the sound of the frame is voiced sound or unvoiced sound,
In the speech synthesis step,
The voice quality conversion method according to any one of claims 1 to 4, wherein speech synthesis is performed only for frames determined as voiced in the determination step. The voice quality conversion method according to any one of claims 1 to 5, further comprising a mixing step of mixing the output of the speech synthesis step and the original sound waveform at a predetermined ratio. In the input step,
The voice conversion method according to any one of claims 1 to 6, wherein the synthesized voice output by the voice synthesizer is input as the original sound waveform.   A voice quality conversion program that causes a computer to execute the voice quality conversion method according to claim 1. A device that converts voiced sound to unvoiced sound,
An input unit for inputting an original sound waveform;
A predictive analyzer that predicts a transfer function from the original sound waveform input to the input unit;
A residual signal extraction unit that extracts a residual signal using the original sound waveform and the output of the prediction analysis unit;
A white noise generator that outputs a white noise signal corresponding to the power of the residual signal;
A speech synthesizer that synthesizes speech based on the output of the white noise generation unit and the transfer function predicted by the prediction analysis unit;
A voice quality conversion device characterized by comprising: Receiving mora information and conversion rule information for inserting silence into the voice using the mora information;
The voice quality conversion device according to claim 9, further comprising a silence insertion unit that inserts a silence part into the voice synthesized by the voice synthesis unit using these pieces of information. The silent insertion part is
The voice quality conversion device according to claim 10, wherein fade-in processing or fade-out processing is performed on the sound before and after the silent portion is inserted. Receives mora information and conversion rule information for expanding the voice on the time axis using the mora information,
The prediction function is provided with a prediction complementing unit that complements the transfer function predicted by the prediction analysis unit on the time axis using these pieces of information.
The white noise generator is
The mora information and the conversion rule information are received, and a white noise signal generated using the information is output,
The speech synthesizer
10. The voice quality conversion apparatus according to claim 9, wherein speech synthesis is performed based on an output of the white noise generation unit and a transfer function supplemented by the prediction complementing unit. For each frame of the original sound waveform or the residual signal, a determination unit that determines whether the sound of the frame is voiced sound or unvoiced sound,
The voice quality conversion apparatus according to any one of claims 9 to 12, wherein the voice synthesizer performs voice synthesis only on a frame determined by the determination unit as a voiced sound. The voice quality conversion device according to any one of claims 9 to 13, further comprising a mixing unit that mixes the output of the speech synthesizer and the original sound waveform at a predetermined ratio. The input unit is
The voice conversion device according to any one of claims 9 to 14, wherein the synthesized voice output by the voice synthesizer is input as the original sound waveform.

Images