JPH10240295A - Voice synthesizing method and voice synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a high-quality voice synthesis by making the center of gravity of data a pitch mark to enable the setting of the pitch whose rolling is small with a relatively simple processing. SOLUTION: This device is provided with a text analyzing part 102, a word dictionary 102, a parameter generating part 103, a voice signal input part 104, an element piece producing part 105, an element dictionary 106, a windows hanging part 107 and a synthetic voice part 108. The element producing part 105 detects the minimum value of a voice signal and adds an offset so that the minimum value of the voice signal become non-negative. Windows having a prescribed length are prepared and the windows are hanged to addition data while shifting centers of the window of the prescribed length and the time base coordinate of the center of the windows where gives the maximum of areas is detected by calculating the areas of the window hanged data at the respective center positions of the windows. Then, the voice signal is segmented by being centered around the time base coordinate and window hanging superpositions of the segmented voice signals are performed by shifting them by an amount equivalent to a pitch cycle by making the time base coordinate being the center of the segmented voice signals the center of the superpositions.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis method and a speech synthesis apparatus for synthesizing an arbitrary speech according to rules, and more particularly to a speech synthesis method and a speech synthesis apparatus for obtaining a synthesized speech by connecting speech waveforms.

[0002]

2. Description of the Related Art In a speech synthesis system, a speech waveform itself is recorded, and a speech waveform is created by combining the speech waveforms, and a parameter representing speech characteristics is analyzed and recorded. Some use. further,
There are various voice synthesis methods such as recording and editing, voice unit editing, parameter editing, and rule synthesis, depending on the combination of the methods of the control unit and the waveform forming unit.

[0003] Among them, the editing and synthesis of speech units (mainly speech waveforms of one pitch period) can be controlled in pitch and can be extended to the output of an arbitrary word depending on the selection of speech units. In the parameter editing / synthesis, a sound source parameter and a speech transfer characteristic (spectrum) parameter are encoded and recorded for each unit speech, and speech synthesis is performed in a parameter time series. The rule synthesis of speech is to convert a sequence represented by discrete symbols such as characters and phonetic symbols into continuous speech.

[0004] Text-to-speech synthesis converts text (text data) into speech, and usually does not correspond one-to-one with speech notation. Therefore, language processing such as morphological analysis and syntax analysis is required to convert the input text into a sequence of phonetic symbols and to automatically generate prosodic features.

Conventionally, a text-to-speech conversion for converting a text sentence into a voice is performed by a text analysis unit including a text analysis unit, a parameter generation unit, and a speech synthesis unit. By referring to the morphological analysis, determining the reading, accent and intonation, outputting phonetic symbols with prosodic symbols (intermediate language), the parameter generation unit sets the pitch frequency pattern, phoneme duration, etc., and the speech synthesis unit Then, speech synthesis processing is performed.

In the speech synthesis unit, a linear prediction method or the like has been used before. However, since deterioration due to separately treating vocal tract information and sound source information which are originally correlated cannot be avoided, vocal tract information has recently been used. A method of obtaining high-quality synthesized sound with little deterioration without separating the sound source information and the original sound waveform as it is has come to be used.

As a method of using a speech waveform as it is, for example, a method described in the literature “FJ Charpentier, MGStella,” Diphon
e synthesis using an overlap-add technique for spe
ech waveforms concatenation ", Proc.Int.Conf.ASSP, 20
15-2018, Tokyo, 1986 ". As shown in this document, a pitch mark (reference point) is added to a speech waveform in advance, and the sound waveform is cut out at the center, and the pitch mark position is shifted while shifting the synthesized pitch cycle according to the synthesized pitch cycle during synthesis. PSOLA (pitch-synchronous overlap add me)
thod).

FIG. 15 shows a conventional PS cited from the above document.
It is a schematic diagram which shows the speech synthesis method of superimposing while changing the pitch of OLA.

[0009] This is compared with the time of analysis (when a segment is created).
This is an example of a case where the pitch period is increased (the pitch is decreased) during synthesis.

In PSOLA, since the pitch can be changed, it has been widely used as a speech synthesizer in text-to-speech conversion.

It is necessary to add a pitch mark for each pitch. Until now, the position of the pitch mark,
The following (1) to (5) are proposed.

(1) Peak of audio waveform For example, a method described in Japanese Patent Application Laid-Open No. 4-372999 (NTT Data Communication Co., Ltd.).

(2) Peak of short-time power For example, Kawai, Higuchi, Shimizu, Yamamoto, "Study of speech synthesis system with waveform unit connection", IEICE Technical Report SP93-9 (1)
993-05).

(3) Peak after pitch filter A method described in, for example, JP-A-7-72897 (Nippon Telegraph and Telephone Corporation).

(4) 15% delay point of impulse drive point For example, Literature Arai, Nishimura, Yoshida, Minowa, "Examination of Pitch Waveform Extraction Position", IEICE Technical Report, SP95-8 (1995-0)
5) The method described in the above.

(5) Glottal Closing Point For example, the literature Sakamoto, Saito, Suzuki, Hashimoto, Kobayashi, "On a Japanese text-to-speech synthesis system using a waveform superposition method", IEICE Technical Report S $ 95-6 (1995-05) Method.

Since the energy of the peak of the audio waveform in the above (1) is concentrated at the local peak position of the audio waveform, it is considered to be suitable for storing the spectrum of the cut-out waveform.

The above-mentioned short-time power peak (2) is also used for the same reason.

The peak after the pitch filter of the above (3) is
It is the peak of the driving waveform of the vocal cord of one pitch, and according to the above-mentioned literature, it is considered that the pitch interval is well represented.

At the 15% delay point of the impulse driving point in the above (4), Arai et al. Report that the spectral distortion is minimized.

The glottal closure point in the above (5) is considered to be the same as the impulse drive point (excitation point of one pitch waveform). Sakamoto and colleagues proposed a method to extract glottal closure points stably.
Dynamic Wavelet transform is used.

[0022]

However, the conventional speech synthesizing method using the pitch mark position has the following problems.

In the peak of the voice waveform of the above (1), a high frequency (white noise) component becomes large in a voiced sound before and after an unvoiced consonant, or in a voiced sound including a plosive sound or an affricate sound, and a unit at the time of synthesis ( The pitch mark fluctuates for each frame), resulting in a loose sound with poor connection.

Since the peak value of the short-time power of the above (2) is evaluated equally on the maximum value and the minimum value, a pitch fluctuation may occur depending on the speaker.

According to the experiment conducted by the present inventors, the peak after the pitch filter of the above (3) has a deviation from the peak of the waveform, and the fluctuation of the pitch due to the deviation from the peak position of the waveform is large. The peak was better.

The 15% delay point of the impulse drive point in the above (4) has a large amount of processing and causes a delay in processing.

The glottal closure point (5) requires a large amount of processing to be extracted.

As described above, it is difficult to set the pitch mark position by a simple method in any of the above-mentioned voice synthesizing methods.

An object of the present invention is to provide a method of setting a pitch mark with relatively simple processing and a small fluctuation, and to provide a voice synthesizing method and a voice synthesizing apparatus capable of synthesizing high quality voice.

[0030]

SUMMARY OF THE INVENTION A speech synthesis method according to the present invention is a speech synthesis method for obtaining a synthesized speech by connecting or overlapping speech waveforms. Add the offset so that the minimum value is non-negative, calculate the addition data, prepare a window of a predetermined length, apply a window to the addition data while moving the center of the predetermined length window, and at each center position of the window Calculates the area of the windowing data of, detects the time axis coordinate of the center of the window that gives the maximum area, cuts out the audio signal by centering around the time axis coordinate, superimposes the extracted audio signal and the time axis coordinate Is characterized in that it is superimposed on a window while being shifted by a pitch period.

A voice synthesizing method according to the present invention is a voice synthesizing method for obtaining a synthesized voice by connecting and overlapping voice waveforms, wherein a pitch period is detected from a voice signal and a minimum value of the voice signal is detected. Then, add an offset so that the minimum value is non-negative to the audio signal to calculate addition data, prepare a window having a pitch period length, apply a window to the addition data while moving the center of the window, and apply a window to each center position of the window. Calculates the area of the windowing data of, detects the time axis coordinate of the center of the window that gives the maximum area, cuts out the audio signal by centering around the time axis coordinate, superimposes the extracted audio signal and the time axis coordinate Is characterized in that it is superimposed on a window while being shifted by a pitch period.

The voice synthesizing method according to the present invention is a voice synthesizing method for obtaining a synthesized voice by connecting and overlapping voice waveforms, wherein a pitch period is detected from a voice signal and a minimum value of the voice signal is detected. Calculate the addition data by adding an offset so that the minimum value is non-negative to the audio signal, prepare a window having a pitch period length, move the center of the window, and apply a window to the data obtained by raising the addition data to a predetermined power. Calculate the area of the windowing data at each center position of the window, detect the time axis coordinate of the center of the window giving the maximum area, center the time axis coordinate, cut out the audio signal, and extract the cutout audio signal. In addition, windowed superimposition is performed while shifting by a pitch period with the time axis coordinate being the center of superimposition.

In the speech synthesis method according to the present invention, the detection of the pitch period may be an accurate detection of the pitch period obtained by the cepstrum analysis method.
Any of the square, the fourth, or the eighth power may be used.

Further, in the voice synthesizing apparatus according to the present invention, the minimum value of the voice signal is detected in the voice synthesizing apparatus which obtains a synthesized voice by connecting and overlapping voice waveforms. Means for calculating addition data by adding an offset so as to be non-negative; preparing a window of a predetermined length, applying a window to the addition data while moving the center of the window of a predetermined length, and setting a window at each center position of the window Means for calculating the area of the multiplication data, detecting the time axis coordinate of the center of the window that gives the maximum area, and extracting the audio signal by centering around the time axis coordinate, superimposing the extracted audio signal on the time axis coordinate And a means for windowing and superimposing while shifting by a pitch period.

A voice synthesizing apparatus according to the present invention is a voice synthesizing apparatus for obtaining a synthesized voice by connecting or superimposing voice waveforms to each other, wherein a means for detecting a pitch period from a voice signal and a minimum value of the voice signal are provided. Means for detecting and adding an offset so that the minimum value is non-negative to the audio signal to calculate addition data, and preparing a window having a pitch period length, applying a window to the addition data while moving the center of the window, and applying a window to the window. Means for calculating the area of the windowing data at each center position, detecting the time axis coordinate of the center of the window that gives the maximum area, and extracting the audio signal by centering around the time axis coordinate, and extracting the extracted audio signal Means for windowing and superimposing while shifting by the pitch cycle with the time axis coordinate being the center of superimposition.

A voice synthesizing apparatus according to the present invention is a voice synthesizing apparatus for obtaining a synthesized voice by mutually connecting or superposing voice waveforms, wherein a means for detecting a pitch period from a voice signal and a minimum value of the voice signal are provided. Means for detecting and adding an offset so that the minimum value is non-negative to the audio signal to calculate addition data, and preparing a window having a pitch period length, moving the center of the window, and moving the addition data to a predetermined power. Means for calculating the area of the windowing data at each center position of the window, detecting the time-axis coordinate of the center of the window that gives the maximum area, and centering the time-axis coordinate to generate an audio signal. Means for clipping and clipping the clipped audio signal by windowing while shifting the pitch axis by using the time axis coordinate as the center of superposition.

[0037] The speech synthesizer according to the present invention may determine the window length of the window based on the pitch period obtained by the cepstrum analysis method.

[0038]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The speech synthesis method and speech synthesis device according to the present invention can be applied to a speech synthesis method using text data as input.

FIG. 1 is a block diagram showing a configuration of a speech synthesis method and a speech synthesis apparatus according to the first embodiment of the present invention. The speech synthesis method according to the present embodiment is effective for all speech synthesis methods that use text data as input.

In FIG. 1, reference numeral 101 denotes a text analysis unit;
102 is a word dictionary, 103 is a parameter generator, 104
Is an audio signal input unit, 105 is a unit creation unit, 106 is a unit dictionary, 107 is a windowing unit, and 108 is a synthesized speech unit.

The text analysis unit 101 inputs a sentence mixed with kanji and kana, performs morphological analysis with reference to the word dictionary 102, determines reading, accent, intonation,
Outputs phonetic symbols with prosody (intermediate language). Accent and intonation are most closely related to the temporal pattern of pitch frequency, which not only gives a natural, easy-to-hear tone,
It plays a role in showing a group of words and phrases to make it easier to understand sentence speech.

The word dictionary storage unit 102 stores, for example, an RO
M and RAM, and stores a word dictionary and a word search table that defines the types of subsequent words that can be grammatically linked. The parameter generation unit 103 sets a pitch frequency pattern, a phoneme duration, and the like.

The audio signal input unit 104 is connected to the RS-23.
It is composed of a communication port such as 2C, an FDD, and an internal buffer for storing data, and text data for speech synthesis is input through the communication port and the FDD.

The segment generating section 105 performs a process of setting the center of gravity of the data as a pitch mark, and is a main part of the present voice synthesizing method, and will be described later in detail with reference to FIGS.

The unit dictionary 106 is created by the unit creating unit 105 after inputting a speech signal.

The speech synthesis unit 108 is used to
Is selected, and a time window having a length of Tp1 is set on the selected segment so that the pitch mark becomes the center.
7 and windowed by $ SOLA (pitch-synchronous
Speech synthesis using Overlap Add method).

Here, the time window length Tp1 is controlled as in equation (1), where Tpa is the pitch period at the time of analysis and Tps is the pitch period at the time of synthesis.

Tp1 = C0 × min {Tpa, Tps} Expression (1) where C0 is a value of about 2.0, and X represents multiplication.

The operation of the speech synthesis method configured as described above will be described below.

First, text data such as a text file in a personal computer communication or a text file in a floppy disk (FD) is input via a communication port such as RS232C or an FDD, temporarily stored in an internal buffer, and exceeds a certain amount. Thus, the text is sent to the text analysis unit 101 for each unit (for example, for each sentence).

In the text analysis unit 101, a ROM or RA
M is read while collating the word dictionary of the word dictionary 102 composed of M with its text data, and generates a phonological prosodic symbol in which information such as accent, intonation, and pause is described as a character string, and this is output to the parameter generation unit 103. send.

The synthesis parameter generation unit 103 determines the position of the speech unit data stored in the speech unit data storage unit 106 and the duration of each phoneme (repetition of the speech unit to be used) based on the phoneme prosody symbol string. ), The pitch of the voice (the repetition interval of the speech unit to be used), and the strength of the voice (the magnification of the speech unit to be used), and a synthesis parameter composed of these pieces of information is generated and the synthesized speech unit 108 is generated. Send to

The synthesized voice section 108 generates voice waveform data while reading the pitch mark position output from the windowing section 107 based on the generated synthesized parameters, and converts this into a D / A converter (not shown). ).

Next, the operation of the segment generating unit 105 will be described with reference to the flowchart shown in FIG.

FIG. 2 is a flowchart showing the operation of the segment creating unit 105, and FIG. 3 is a flowchart showing the details of the center-of-gravity point detection processing of FIG. In the figure, ST indicates each step of the flow.

It is assumed that the audio signal is input from the disk or the like by the audio signal input unit 104.

First, in step ST101, the audio signal data is divided into sections called analysis frames. In the present embodiment, the length of one frame is 32 ms, and the next frame is shifted by 8 ms. Here, it is assumed that the total number of frames is N.

Next, in step ST102, a frame number i to be processed is initialized.

Next, a window length for detecting the center of gravity is determined in step ST103, and the center of gravity is detected in step ST104.

Step ST103 and step ST10
Reference numeral 4 denotes a core part of the present speech synthesis method. In this step, a pitch mark is detected. See Figure 3 for details.
Will be described later.

In the present embodiment, first, a range for searching for a pitch mark is given to the data of the i-th frame, and data used for searching for the pitch mark (in the search range and half the window length from the start point / end point of the search range). A minimum value is detected for the audio signal (up to the time axis coordinates), and an offset is added so that this minimum value becomes zero. In the present embodiment, the minimum value detection processing is performed for each frame as the minimum value. However, the minimum value of a long time length or the minimum value that can be easily taken by audio data (2's complement format in 16-bit binary number) Then -3
2768) may be used.

A window of a fixed length is applied to the data to which the offset has been added, and the area of the windowing data at the center position of each window is calculated while shifting the center of the window by one point, and the maximum of the area is given. Detect the time axis coordinates of the center of the window,
This is called the center of gravity of the data of the i-th frame, and is used as the pitch mark of the data of the i-th frame.

When the pitch mark of the data of the i-th frame is detected, the audio data before and after the pitch mark is cut out in step ST105, and the centering is performed so that the pitch mark is located at the center. In the present embodiment, the cut-out length is set to 12 ms, which is a length of the longest pitch period for males with a margin, by a preliminary experiment.

Next, in step ST106, segment writing is performed as a segment in the i-th frame, and step S106 is executed.
The audio data extracted in T107 is sequentially written as a segment dictionary on a storage medium such as a disk.

In step ST108, the total number of frames N
Is greater than the frame number i (i <N), it is determined whether all frames have been completed. If i <N, it is determined that the processing has not been completed, and the frame number to be processed in step 109 is determined. Update (i = i + 1) step ST10
Proceed to 3 to continue the subsequent processing. Step ST
If it is determined in step 108 that the processing of all the frames has been completed, the closing operation of the disc (not shown) and the like are performed, and the operation of the segment creating unit 105 is terminated.

FIG. 3 is a flowchart showing the process of detecting the center of gravity of FIG.
It is a flowchart which shows the detailed operation of T104).

First, the data data [k], (k = k) of the i-th frame of the audio signal divided in step ST201
1, 2, ...). Next, step ST202
To set the window length, and the fixed-length window data window
[K], (k = −window length / 2−window length / 2).

Next, the center-of-gravity point detection processing (step ST1)
04) As an operation, a pitch mark search range (start point / end point) is set in step ST203.

In step ST204, step ST2
Step ST202 from the start point of the search range set in 03
Step ST from the point before the window length set by
Step ST20 from the end point of the search range set in 203
The minimum value [i] of the audio signal between the point after half the window length set in step 2 and the next point is detected.

In step ST205, step ST2
An offset is added to the data in the frame so that the minimum value detected in 04 becomes 0. This offset addition is represented by equation (2).

Off_Add_Data (k) = Data [k] + minimum [i] Expression (2) FIG. 4 is a diagram showing a minimum value detection range and a window function. Hit the value.

Next, j is the time axis coordinate which is the center of the pitch mark search window, and in step ST206, the time axis coordinate of the pitch mark search start point set in step ST203 is substituted as the initial value of j. In steps ST207 to ST210, the window length data is multiplied by the fixed-length window from the search start point, and the area S (j) of the windowed data is calculated.

That is, in step ST207, k is set to -window length / 2, and the area S (j) is set to 0. In step ST208, the data is windowed, and the area is calculated. The area S (j) of the windowing data is calculated by Expression (3).

S (j) = S (j) + Off_Add_Data (k) × window [k] Expression (3) In step ST209, k is larger than the window length / 2 (k
> Window length / 2), and k is equal to or less than window length / 2 (k ≦
In the case of (window length / 2), it is determined that the calculation of the window length has not been completed, and k is incremented by 1 (k = k) in step ST210.
+1) and the process returns to step ST208.

K is larger than window length / 2 (k> window length / 2)
In the case of, it is determined that the calculation of the window length has been completed, and step ST2 is performed.
At 11, it is determined whether or not j is larger than the search end point (j> search end point). If j is equal to or smaller than the search end point (j ≦ search end point), the area S (j) of the windowing data has not been calculated. In step ST212, j is incremented by 1 (j
= J + 1), and returns to step ST207.

As described above, the center of the window is set at step ST21.
The area S (j) of the windowing data is calculated while shifting one point at a time by 2 in step ST211.
Is larger than the search end point (j> search end point), it is determined that the center of the window has reached the position of the search end point, and the process proceeds to step ST213 to terminate the calculation of the area S (j) of the windowing data.

Finally, in step ST213, the time axis coordinate argmax [S of the center of the window giving the maximum area is set.
(J)], and this is called the center of gravity of the window, and the center of gravity point detection processing is terminated as a pitch mark.

According to this flow, when a pitch mark of the data of the i-th frame is detected, step ST in FIG.
The process proceeds to 105 to perform a segment creation process by extracting audio data before and after the pitch mark.

FIG. 5 is a diagram showing the transition of the area of the windowing data at each window length of 64, 96, and 128 points. FIG. 6 is a graph showing the transition of the area of the windowing data by detecting the center of gravity of the waveform having a large high-frequency component. FIG.

FIG. 5 shows the waveform a (/ a /) (// indicates a syllable boundary symbol) and the stationary part of the uttered waveform, and FIG. 6 shows the waveform user (/ uza /) of the uttered waveform. It is shown as a schematic diagram of z / part.

The solid lines in FIGS. 5 and 6 show the waveform data to which an offset has been added. The horizontal axis is the time axis, and the vertical axis is the magnitude of the waveform signal. 0 is the minimum value of the waveform (minimum [i]). The dashed lines in FIGS. 5 and 6 indicate the area of the windowing data when the center of the window is at the time point on the horizontal axis, and the point on the time axis that gives the maximum, that is, *, according to the present speech synthesis method. This is the pitch mark position (the center of gravity of the data).

As described above, the text-to-speech synthesis method and the text-to-speech synthesizing apparatus according to the first embodiment include the text analysis unit 1.
01, word dictionary 102, parameter generation unit 103, voice signal input unit 104, unit creation unit 105, unit dictionary 10
6, a windowing unit 107 and a synthesized voice unit 108. The unit generating unit 105 detects the minimum value of the voice signal and adds an offset to the voice signal so that the minimum value is non-negative. ), Prepare a window of a predetermined length, apply a window to the added data while moving the center of the predetermined length window, calculate the area of the windowing data at each center position of the window, and give the maximum area. Means (mainly step ST208) for detecting the time axis coordinate of the center of the audio signal, extracting the audio signal by centering around the time axis coordinate, and setting the time axis coordinate of the center of the extracted audio signal as the center of superimposition, and setting the pitch cycle Since means (mainly steps ST105 to ST109) for windowing and superimposing while shifting each other is provided, the center of gravity of the data is used as a pitch mark, thereby simplifying the operation. Management in can be set pitch mark allows a high quality speech synthesis.

In a waveform (a waveform in FIG. 6) in which the power of a high frequency component (random noise) seen in a voiced sound including a plosive or affricate or a voiced sound immediately before an unvoiced consonant is considerably large. In addition, pitch marks could be stably extracted, and in voice synthesis listening, poor connection (scratching) between phonemes could be eliminated as compared to a pitch mark-added segment due to a peak.

Further, since the center of gravity of the waveform is used as the pitch mark, the deformation of the pitch waveform is small, the decrease of the formant Q is suppressed, and a brighter sound is obtained.

FIG. 7 is a block diagram showing a configuration of a voice synthesizing method and a voice synthesizing apparatus according to a second embodiment of the present invention. FIG. 1 is used in describing the speech synthesis method according to the present embodiment.
The same components as those of the speech synthesis method and the speech synthesis device shown in FIG.

In FIG. 7, reference numeral 101 denotes a text analysis unit;
102 is a word dictionary, 103 is a parameter generator, 104
Denotes a speech signal input unit, 200 denotes a unit creation unit, 106 denotes a unit dictionary, and 108 denotes a synthesized speech unit. The unit creation unit 200 includes a pitch mark calculation unit 210 and a windowing unit 211.

The segment generating section 200 performs a process of setting the center of gravity of the data as a pitch mark, and is a main part of the present voice synthesizing method, and will be described later in detail with reference to FIGS.

After inputting the speech signal, the segment dictionary 106 is created by the segment creating section 200.

The speech synthesizing unit 108
Select the segment in the ΡSOLA (pitch-synchronous
Speech synthesis using Overlap Add method).

The operation of the speech synthesis method configured as described above will be described below.

In the second embodiment, when detecting the center of gravity of data, a highly reliable pitch cycle is obtained in advance by cepstrum analysis, and the pitch cycle is used as a window function at the time of extracting the center of gravity. This is different from the first embodiment.

FIG. 8 is a flowchart showing the operation of the segment generating section 200. FIG. 9 is a flowchart showing the details of the pitch period detection processing of FIG. 8, and FIG. 10 shows the details of the center-of-gravity point detection processing of FIG. It is a flowchart. The same processing steps as those in the segment creation processing shown in FIGS. 2 and 3 are denoted by the same step numbers.

It is assumed that the audio signal is input from a disk or the like by the audio signal input unit 104.

First, in step ST101, the audio signal data is divided into sections called analysis frames. In the present embodiment, the length of one frame is 32 ms, and the next frame is shifted by 8 ms. Here, it is assumed that the total number of frames is N.

Next, in step ST102, the frame number i to be processed is initialized.

Next, in step ST301, the pitch period Tp of the sound in the i-th frame is detected. As a pitch period detection method, a peak interval of a waveform or the like can be considered as a simple method. In the present embodiment, the cepstrum method is used to calculate the pitch period more precisely. A pitch cycle detection method using the cepstrum method will be described later with reference to FIG.

Next, in step ST302, a window length for detecting the center of gravity is determined based on the pitch period obtained by the cepstrum method, and the center of gravity is detected in step ST303.

The pitch period obtained by the cepstrum method is used as the window length of the window used for detecting the center of gravity, and the window data is created. As a result of the preliminary experiment by the present inventors, the Hanning window having a large attenuation in the side lobe is used as the window to be used. It was confirmed that a pitch mark was required stably on a Blackman window or the like.

Step ST302 and step ST30
Reference numeral 3 denotes a core part of the present speech synthesis method, and in this step, a pitch mark is detected. See Figure 1 for details.
0 will be described later.

In the present embodiment, similarly to the first embodiment, first, a range for searching for a pitch mark is given to the data of the i-th frame, and data used for searching for a pitch mark (in the search range and in the search range). The minimum value is detected for the audio signal from the start point / end point to the time axis coordinate of half the window length), and an offset is added so that the minimum value becomes zero. A window of a fixed length is applied to the data to which the offset is added, and the area of the windowing data at the center position of each window is calculated while shifting the center of the window by one point, and the center of the window giving the maximum of the area is calculated. , And this is referred to as the center of gravity of the data of the i-th frame, and is used as the pitch mark of the data of the i-th frame.

When a pitch mark of the data of the i-th frame is detected, audio data before and after the pitch mark is cut out in step ST304, and the centering is performed so that the pitch mark is located at the center. In the present embodiment, the cut-out length is set to 12 ms, which is a length of the longest pitch period for males with a margin, by a preliminary experiment.

Next, in step ST306, step S306 is executed.
Windowing is performed based on the window data C0 from T305, and the windowed data is written in a segment dictionary in step ST307, and the audio data cut out in step ST107 is sequentially written as a unit dictionary on a storage medium such as a disk. The windowing in step ST306 corresponds to the operation of the windowing unit 107 (FIG. 1) of the first embodiment.

As described above, in the present embodiment, step S
Writing to the unit dictionary is performed in T307, but since the accurate pitch period of the i-th frame has already been obtained in step ST301, the windowed one is written in the unit dictionary in step ST306. Thereby, at the time of speech synthesis, the multiplication of windowing for each pitch, which is required in the first embodiment, becomes unnecessary, and only superposition is performed, and the processing amount of the speech synthesis processing amount is greatly increased. Can be reduced.
For this reason, in an apparatus using the present speech synthesis method, D
Without using an arithmetic processor such as SΡ
It can be realized by PU.

In step ST108, the total number of frames N
Is greater than the frame number i (i <N), it is determined whether all frames have been completed. If i <N, it is determined that the processing has not been completed, and the frame number to be processed in step 109 is determined. Update (i = i + 1) step ST30
Proceed to 1 to continue the subsequent processing. Step ST
If it is determined in step 108 that the processing of all the frames has been completed, a disc closing process or the like (not shown) is performed, and the operation of the segment generating unit 200 ends.

FIG. 9 is a flowchart showing the detailed operation of the pitch cycle detection process (step ST301) of FIG. 8. As the pitch cycle detection method, a cepstrum method that enables accurate calculation of the pitch cycle is used.

First, a time waveform is input in step ST401, and windowing is performed in step ST402.

Next, a discrete Fourier transform (DFT) is performed on the time-waveform subjected to windowing in step ST403, and the square root of the sum of squares of the real part and the imaginary part is logarithmically converted in step ST404.

Next, inverse Fourier transform (IDFT) is performed in step ST405, and cepstrum components are obtained by inverse Fourier transform in step ST406, and this flow ends.

That is, the cepstrum method converts a convolution operation into an additive operation, and a voiced sound signal of speech is obtained by convolving a sound source component with vocal tract information.
Suitable for separation of both.

When the input signal is a voiced voice, if the pitch period is T0, the sound source component appears near T0.
The vocal tract component appears as a component in a short-time region.

The pitch period obtained by the above method is used as the window length of the window used for detecting the center of gravity, and window data is created. In a preliminary experiment by the present inventors, it was confirmed that a pitch mark can be stably obtained in a Hanning window, a Blackman window, or the like having a large attenuation in a side lobe as a window to be used.

FIG. 10 is a flowchart showing the detailed operation of the center-of-gravity point detection process (step ST303) of FIG.

First, the data data [k], (k = k) of the i-th frame of the audio signal divided in step ST201
1, 2, ...).

Next, in step ST501, the pitch period is detected, and the detected pitch period is set as the window length, and the window data window [k], (k = −window length / 2−2)
(Window length / 2) is created.

Next, the center-of-gravity point detection process (step S
As T303), a pitch mark search range (start point / end point) is set in step ST203.

At step ST502, at step ST2
Step ST501 from the start point of the search range set in 03
Step ST from the point before the window length set by
Step ST50 from the end point of the search range set in 203
The minimum value [minimum [i]] of the audio signal up to half the window length set in 1 is detected.

In step ST205, step ST5
An offset is added to the data in the frame so that the minimum value detected in 02 becomes 0. This offset addition is represented by the above equation (2).

Next, the time axis coordinate which is the center of the pitch mark search window is set as j, and in step ST206, the time axis coordinate of the pitch mark search start point set in step ST203 is substituted as the initial value of j, and step S206 is executed.
From T207 to step ST210, the fixed length window is applied to the data of the window length from the search start point, and the area S (j) of the windowed data is calculated.

That is, in step ST207, k is set to −window length / 2, and the area S (j) is set to 0. In step ST503, a pitch length window is applied to the data, and the area S (j) of the windowed data is calculated according to the equation (3). ) Is calculated.

In step ST209, it is determined whether or not k is larger than window length / 2 (k> window length / 2), and k is determined to be window length /
If it is not more than 2 (k ≦ window length / 2), it is determined that the calculation of the window length has not been completed, and k is incremented by 1 (k = k + 1) in step ST210, and the process returns to step ST503.

K is larger than window length / 2 (k> window length / 2)
In the case of, it is determined that the calculation of the window length has been completed, and step ST2 is performed.
In step 11, it is determined whether or not j is larger than the search end point (j> search end point). When j is equal to or smaller than the search end point (j ≦ search end point), the area S (j) of the windowing data has not been calculated. In step ST212, j is incremented by 1 (j
= J + 1), and returns to step ST207.

As described above, the center of the window is set at step ST21.
The area S (j) of the windowing data is calculated while shifting one point at a time by 2 in step ST211.
Is larger than the search end point (j> search end point), it is determined that the center of the window has come to the position of the search end point, and the process proceeds to step ST504 to end the calculation of the area S (j) of the windowing data.

Finally, in step ST504, the time axis coordinate argmax [S of the center of the window giving the maximum area is set.
(J)], and this is called the center of gravity of the window.
The pitch mark is obtained in accordance with the above equation, and the present flow ends.

Pitch mark (i) = argmax [S (j)] + window length / 2 Expression (4) According to this flow, when the pitch mark of the data of the i-th frame is detected, step ST304 in FIG. Then, a segment creation process is performed by extracting audio data before and after the pitch mark.

As described above, the speech synthesizing method and the speech synthesizing apparatus according to the second embodiment use the accurate pitch period obtained from the cepstrum analysis as the window length. Can be set to the pitch frequency, and it is possible to cope with a change in the pitch cycle in the phoneme, so that a more stable pitch mark can be obtained and the pitch mark error is reduced. Thereby, in the synthesis sound listening, the connectivity of phonemes could be improved.

FIG. 11 is a block diagram showing a configuration of a speech synthesis method and a speech synthesis device according to the third embodiment of the present invention. In the description of the speech synthesis method according to the present embodiment, the same components as those of the speech synthesis method and the speech synthesis apparatus shown in FIG.

In FIG. 11, 101 is a text analysis unit, 102 is a word dictionary, 103 is a parameter generation unit,
04 is an audio signal input unit, 300 is a unit creation unit, 106 is a unit dictionary, and 108 is a synthesized speech unit.
0 comprises a pitch mark calculation unit 310 and a windowing unit 311.

The segment generating section 300 performs a process of setting the center of gravity of the data as a pitch mark, and is a main part of the present voice synthesizing method, and will be described later in detail with reference to FIGS.

After inputting a speech signal, the segment dictionary 106 is created by the segment creating section 300.

The speech synthesizing unit 108
Select the segment in the ΡSOLA (pitch-synchronous
Speech synthesis using Overlap Add method).

Hereinafter, the operation of the speech synthesis method configured as described above will be described.

In the third embodiment, similarly to the second embodiment, when detecting the center of gravity of data, a pitch cycle having high reliability is obtained in advance by cepstrum analysis, and the pitch cycle is determined by the center of gravity. Used for the window function at the time of extraction.

FIG. 12 is a flowchart showing the operation of the segment generating unit 300, and FIG. 13 is a flowchart showing the details of the center-of-gravity point detection processing of FIG. The same processing steps as those of the segment creation processing shown in FIGS. 8 and 10 are denoted by the same step numbers.

It is assumed that the audio signal is input from a disk or the like by the audio signal input unit 104.

First, in step ST101, the audio signal data is divided into sections called analysis frames. In the present embodiment, the length of one frame is 32 ms, and the next frame is shifted by 8 ms. Here, it is assumed that the total number of frames is N.

Next, in step ST102, the frame number i to be processed is initialized.

Next, in step ST301, the pitch period Tp of the sound in the i-th frame is detected. As the pitch period detection method, the cepstrum method is used to calculate the pitch period more precisely, as in the second embodiment.

Next, in step ST302, a window length for detecting the center of gravity is determined based on the pitch period obtained by the cepstrum method, and the center of gravity is detected in step ST601.

Steps ST302 and ST60
Reference numeral 1 denotes a core part of the present speech synthesis method, and in this step, a pitch mark is detected. See Figure 1 for details.
3 will be described later.

In the present embodiment, similarly to the first and second embodiments, first, a range for searching for a pitch mark is given to the data of the i-th frame, and data used in searching for the pitch mark (in the search range and within the search range). The minimum value is detected for the audio signal from the start point / end point of the search range to the time axis coordinate of half the window length), and an offset is added so that the minimum value becomes zero. A window of a fixed length is applied to the data to which the offset is added, and the area of the windowing data at the center position of each window is calculated while shifting the center of the window by one point, and the center of the window giving the maximum of the area is calculated. , And this is referred to as the center of gravity of the data of the i-th frame, and is used as the pitch mark of the data of the i-th frame.

When a pitch mark of the data of the i-th frame is detected, audio data before and after the pitch mark is cut out in step ST304, and the centering is performed so that the pitch mark is located at the center. In the present embodiment, the cut-out length is set to 12 ms, which is a length of the longest pitch period for males with a margin, by a preliminary experiment.

Next, in step ST306, step S306 is executed.
Windowing is performed based on the window data C0 from T305, and the windowed data is written in a segment dictionary in step ST307, and the audio data cut out in step ST107 is sequentially written as a unit dictionary on a storage medium such as a disk. The windowing in step ST306 corresponds to the operation of the windowing unit 107 (FIG. 1) of the first embodiment.

As described above, in the present embodiment, as in the second embodiment, since the exact pitch period of the i-th frame has already been obtained in step ST301, step ST3 is performed.
In step 06, the windowed data can be written in the segment dictionary, and the speech synthesis does not require the multiplication of the windowing for each pitch required in the first embodiment. And the processing amount of the speech synthesis processing amount can be greatly reduced. Therefore, an apparatus using the present speech synthesis method can be realized by a normal CPU without using an arithmetic processor such as DSΡ.

In step ST108, the total number of frames N
Is greater than the frame number i (i <N), it is determined whether all frames have been completed. If i <N, it is determined that the processing has not been completed, and the frame number to be processed in step 109 is determined. Update (i = i + 1) step ST30
Proceed to 1 to continue the subsequent processing. Step ST
If it is determined in step 108 that the processing of all the frames has been completed, a disc closing process or the like (not shown) is performed, and the operation of the segment generating unit 200 ends.

FIG. 13 is a flowchart showing the detailed operation of the center-of-gravity point detection process (step ST601) of FIG.

First, in step ST201, the audio signal is divided into sections called analysis frames, and data data [k], (k = 1, i) of the i-th frame of the divided audio signal.
2, ...). In the present embodiment, the length of one frame is 32 ms, and the next frame is shifted by 8 ms.

Next, in step ST501, the pitch period is detected, and the detected pitch period is set as the window length, and the window data window [k], (k = −window length / 2−2)
(Window length / 2) is created.

Next, the center-of-gravity point detection processing (step S
As T601), a pitch mark search range (start point / end point) is set in step ST203.

In step ST701, step ST2
The minimum value of the audio signal, minimum [i], from the point time half the window length before the start point of the search range set in 03 to the time point half the window length after the end point of the search range is detected.

In Step ST205, Step ST7 is executed.
An offset is added to the data in the frame so that the minimum value detected in 01 becomes 0. This offset addition is represented by the above equation (2).

Next, j is the time axis coordinate which is the center of the pitch mark search window, and in step ST206, the time axis coordinate of the pitch mark search start point set in step ST203 is substituted as the initial value of j.
From T207 to step ST210, the fixed length window is applied to the data of the window length from the search start point, and the area S (j) of the windowed data is calculated.

That is, in step ST207, k is set to −window length / 2, area S (j) is set to 0, and in step ST702, the pitch length window shown in the second embodiment is applied to a predetermined power of the data to which the offset is added. , And the area S (j) of the windowing offset data is calculated by the equation (3). The area S (j) of the windowing offset data is calculated by equation (5).

S (j) = S (j) + [(Off_Add_Data (k)) ⁿ × window [k]] Expression (5) In step ST209, k is larger than window length / 2 (k
> Window length / 2), and k is equal to or less than window length / 2 (k ≦
In the case of (window length / 2), it is determined that the calculation of the window length has not been completed, and k is incremented by 1 (k = k) in step ST210.
+1), and returns to step ST702.

K is larger than window length / 2 (k> window length / 2)
In the case of, it is determined that the calculation of the window length has been completed, and step ST2 is performed.
At 11, it is determined whether or not j is larger than the search end point (j> search end point). If j is equal to or smaller than the search end point (j ≦ search end point), the area S (j) of the windowing data has not been calculated. In step ST212, j is incremented by 1 (j
= J + 1), and returns to step ST207.

As described above, the center of the window is set at step ST21.
The area S (j) of the windowing data is calculated while shifting one point at a time by 2 in step ST211.
Is larger than the search end point (j> search end point), it is determined that the center of the window has come to the position of the search end point, and the process proceeds to step ST703 to terminate the calculation of the area S (j) of the windowing data.

Finally, in step ST703, the time axis coordinate argmax [S of the center of the window giving the maximum area is set.
(J)], and this is called the center of gravity of the window, and the center of gravity point detection processing is terminated as a pitch mark.

According to this flow, when the pitch mark of the data of the i-th frame is detected, step S in FIG.
The process shifts to T304 to perform a segment creation process by cutting out audio data before and after the pitch mark.

Here, in the third embodiment, in the above-mentioned steps ST207 to ST210, the area of the window offset addition data at the center position of each window is obtained, but the processing of step ST702 is a processing newly proposed. It is.

That is, in step ST702,
The data to which the offset has been added is raised to a predetermined power, the window set in step ST501 is applied to the data raised in the predetermined power to determine its area, and the center of gravity is detected in step ST703.

Using this center of gravity as a pitch mark, re-cutting and centering of the audio signal are performed, and furthermore, step ST5
Using the pitch period obtained in step 01 to write data into the segment dictionary over a window, this process may be the same as in the second embodiment.

The third embodiment is characterized in that the data to which the offset has been added is raised to a predetermined power, a window having a pitch period length is applied, and the center of gravity is detected.

FIG. 14 is a diagram showing the transition of the center of gravity of predetermined power data of a waveform to which an offset has been added. The waveform shown in this figure is a waveform of a female voice uttered as a (/ a /),
On both the vertical and horizontal axes, as in FIG. 5, the vertical axis represents the size of the data to which the offset was added, and the horizontal axis represents time. The center of gravity (pitch mark).

As described above, the voice synthesizing method and the voice synthesizing apparatus according to the third embodiment multiply the data to which the offset has been added by a predetermined power to open a window of a pitch period length,
Since the center of gravity is detected, more appropriate pitch marks can be set by simple processing, and high-quality speech synthesis can be performed.

Hereinafter, the effects of the present embodiment will be specifically described.

FIG. 14 is a diagram showing the behavior of the position of the center of gravity based on a predetermined power of the data to which the offset has been added. When the pitch added window is directly applied to the data to which the offset has been added (the first power), the same result as in the second embodiment can be obtained. Next, the data with the offset
When a pitch length window is applied to the squared, fourth, and eighth powers and the center of the window that gives the maximum area is detected, it can be seen that the center of gravity approaches the waveform peak as the multiplier increases. Here, the reason why the peak of the waveform is emphasized in the pitch mark assignment is to obtain a more natural speech synthesized sound with less deformation into a formant.

In the simple waveform peak detection, since the pitch mark position has a slight fluctuation for each individual waveform, the synthesized sound becomes a rough sound with poor connection. Since the power of the multiplied waveform appears as a wave close to a gentle sine wave for each pitch period, the vicinity of the peak of the waveform can be detected stably, and a smooth synthesized sound can be obtained.

In preliminary experiments by the present inventors, it has been found that when a multiplier of 8 and a Hanning window having a pitch length as a window function are used, the center of gravity substantially coincides with the peak of the waveform.

In each of the above embodiments, the detection of the center of gravity is performed only on the voiced portion of the audio signal. For the unvoiced sound portion, audio data is used as it is.

In the second and third embodiments, the cepstral method is used as the pitch period detecting method. However, other methods such as the autocorrelation method and the modified autonomous correlation which is the autocorrelation of the linear prediction residual are used. A method such as a correlation method can also be used.

[0170] The segment creation unit in the speech synthesis method and apparatus in each of the above embodiments changes the pitch of the original speech to change the pitch of the voice, that is, the pitch mark setting in a so-called speech pitch conversion device. It is also possible to adapt to the processing of various audio output devices.

The speech synthesizing method and apparatus according to each of the embodiments described above can be applied to any speech synthesizing method using text data as input. Digital speech waveform data is converted into an analog signal to synthesize speech. Any of the voice synthesis methods may be used, and may be a part of a circuit incorporated in various terminals.

Furthermore, it goes without saying that the number of dictionaries and various circuit parts and the type of connection, which constitute the speech synthesis method and apparatus according to each of the above embodiments, are not limited to the above embodiments. It goes without saying that various processing means, for example, a pitch mark search range, a method of calculating the area of the windowing data, and the like in the segment creation unit can be appropriately changed within the scope of the technical idea of the present invention.

[0173]

According to the speech synthesizing method and the speech synthesizing apparatus according to the present invention, means for detecting a minimum value of a voice signal, adding an offset to the voice signal so that the minimum value is non-negative, and calculating addition data; Prepare a window of a predetermined length, apply a window to the addition data while moving the center of the predetermined length window, calculate the area of the windowing data at each center position of the window, and calculate the center of the window that gives the maximum area. Means for detecting time axis coordinates, and means for centering the time axis coordinates to cut out an audio signal, and for windowing and superimposing while shifting the pitch axis by the time axis coordinate of the center of the extracted audio signal as the center of superimposition. As a result, a pitch mark with less fluctuation can be set by simple processing, and high-quality speech synthesis can be realized.

In the voice synthesizing method and the voice synthesizing apparatus according to the present invention, means for detecting a pitch period from a voice signal, detecting a minimum value of the voice signal, and adding an offset to the voice signal so that the minimum value is non-negative. A window having a pitch period length is prepared, a window is added to the added data while moving the center of the window, and the area of the windowing data at each center position of the window is calculated. Means for detecting the time axis coordinate of the center of the window that gives the audio signal, and extracting the audio signal by centering around the time axis coordinate, and shifting the window by shifting the pitch axis by the time axis coordinate of the center of the extracted audio signal as the center of superposition. Since there is a means for multiplying and superimposing, it is possible to set a pitch mark with less fluctuation by simple processing.

Further, since an accurate pitch period is used as the window length, the cutoff frequency of the window can be set to the pitch frequency, and the change in the pitch period during the phoneme can be handled, so that the window is more stable. Pitch marks can be obtained.

In the voice synthesizing method and the voice synthesizing apparatus according to the present invention, means for detecting a pitch period from a voice signal, detecting a minimum value of the voice signal, and adding an offset to the voice signal so that the minimum value is non-negative. A window having a pitch period length is prepared, and while moving the center of the window, a window is formed on the data obtained by raising the added data to a predetermined power, and the area of the windowed data at each center position of the window is calculated. Means for calculating and detecting the time axis coordinate of the center of the window giving the maximum of the area, and cutting out the audio signal by centering around the time axis coordinate, and taking the time axis coordinate of the center of the extracted audio signal as the center of superposition, Since it has a means for windowing and superimposing while shifting by the pitch period, it is possible to set a pitch mark with less fluctuation by simple processing, and furthermore, a gentle sine wave for each pitch period To appear as close waves can be detected near the peak of the stable waveform can be obtained a smooth synthetic sounds.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech synthesis method and a speech synthesis device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the speech synthesis method and the speech synthesis unit of the speech synthesis device.

FIG. 3 is a flowchart illustrating details of a center-of-gravity point detection process of the speech synthesis method and the speech synthesis device.

FIG. 4 is a diagram showing a minimum value detection range and a window function of the speech synthesis method and the speech synthesis device.

FIG. 5 shows the speech synthesis method and the speech synthesis device 64, 9;
It is a figure which shows transition of the area of the windowing data in each window length of 6,128 points.

FIG. 6 is a diagram showing a transition of an area of windowing data by detecting a center of gravity in a waveform having a large high frequency component of the speech synthesis method and the speech synthesis device.

FIG. 7 is a block diagram illustrating a configuration of a speech synthesis method and a speech synthesis device according to a second embodiment of the present invention.

FIG. 8 is a flowchart showing the operation of the speech synthesis method and the segment creation unit of the speech synthesis device.

FIG. 9 is a flowchart showing details of a pitch cycle detection process of the speech synthesis method and the speech synthesis device.

FIG. 10 is a flowchart showing details of a center-of-gravity point detection process of the speech synthesis method and the speech synthesis device.

FIG. 11 is a block diagram illustrating a configuration of a speech synthesis method and a speech synthesis device according to a third embodiment of the present invention.

FIG. 12 is a flowchart showing the operation of the speech synthesis method and the segment creation unit of the speech synthesis device.

FIG. 13 is a flowchart illustrating details of a center-of-gravity point detection process of the speech synthesis method and the speech synthesis device.

FIG. 14 is a diagram showing a transition of a center of gravity of predetermined power data of a waveform to which an offset of the speech synthesis method and the speech synthesis device is added.

FIG. 15 is a schematic diagram showing a conventional speech synthesis method of superimposing while changing the pitch of PSOLA.

[Explanation of symbols]

101 text analysis unit, 102 word dictionary, 103
Parameter generation unit, 104 audio signal input unit, 105,
200, 300 unit creation unit, 106 unit dictionary, 10
7, 211, 311 windowing part, 108 synthetic voice part,
210, 310 pitch mark calculation unit

Claims (9)

[Claims] 1. A voice synthesis method for obtaining a synthesized voice by connecting or overlapping voice waveforms, wherein a minimum value of a voice signal is detected, and an offset is added to the voice signal so that the minimum value is non-negative. Calculating the addition data, preparing a window of a predetermined length, moving the center of the predetermined length window, applying the window to the addition data, and calculating the area of the windowing data at each center position of the window. Is calculated, the time axis coordinate of the center of the window that gives the maximum of the area is detected, the audio signal is cut out by centering around the time axis coordinate, and the cut out audio signal is superimposed on the time axis coordinate. A voice synthesizing method, characterized in that windowing and superimposing are performed while being shifted by a pitch period as a center of the speech. 2. A voice synthesizing method for obtaining a synthesized voice by connecting or overlapping voice waveforms, wherein a pitch period is detected from a voice signal, a minimum value of the voice signal is detected, and Calculating addition data by adding an offset so that the minimum value is non-negative, preparing a window of the pitch period length, applying the window to the addition data while moving the center of the window, and applying each window center position. Calculates the area of the windowing data in, detects the time axis coordinate of the center of the window that gives the maximum of the area, cuts out the audio signal by centering around the time axis coordinate, and cuts out the audio signal, Using the time-axis coordinates as the center of superimposition and window-superimposing them while shifting them by a pitch cycle. 3. A voice synthesizing method for obtaining a synthesized voice by connecting or overlapping voice waveforms, wherein a pitch period is detected from a voice signal, a minimum value of the voice signal is detected, and the voice signal is added to the voice signal. Calculating the addition data by adding an offset so that the minimum value is non-negative, preparing a window of the pitch period length, moving the center of the window, and applying the window to data obtained by raising the addition data to a predetermined power. Calculating the area of the windowing data at each center position of the window, detecting the time axis coordinate of the center of the window that gives the maximum of the area, and centering around the time axis coordinate to obtain the audio signal. A speech synthesizing method, comprising: clipping and superimposing the clipped audio signal on a window while shifting the pitch axis by using the time axis coordinate as a center of superposition. 4. The speech synthesis method according to claim 2, wherein the detection of the pitch period is a detection of an accurate pitch period obtained by a cepstrum analysis method. Method. 5. The speech synthesis method according to claim 3, wherein the predetermined value of the power is any one of a square, a fourth and an eighth power. 6. A voice synthesizing apparatus for obtaining a synthesized voice by connecting and overlapping voice waveforms, detecting a minimum value of a voice signal, and adding an offset to the voice signal so that the minimum value is non-negative. Means for calculating addition data, and preparing a window of a predetermined length, applying the window to the addition data while moving the center of the predetermined length window, and applying the windowing data at each center position of the window. Means for calculating the area of the window, and detecting the time axis coordinate of the center of the window that gives the maximum of the area, and cutting out the audio signal by centering around the time axis coordinate, A voice synthesizing device comprising: means for windowing and superimposing while shifting the pitch coordinates by a pitch cycle with the axis coordinates being the center of superimposition. 7. A voice synthesizer for obtaining a synthesized voice by connecting or overlapping voice waveforms, comprising: means for detecting a pitch period from a voice signal; detecting a minimum value of the voice signal; Means for calculating an addition data by adding an offset so that the minimum value is non-negative; preparing a window having the pitch period length; applying the window to the addition data while moving the center of the window; Means for calculating the area of the windowing data at each center position, detecting time axis coordinates of the center of the window that gives the maximum of the area, and centering the time axis coordinates to obtain the audio signal. A speech synthesizing apparatus, comprising: means for clipping and windowing and superimposing the clipped speech signal while shifting the clipped speech signal by a pitch period with the time axis coordinates being the center of superposition. 8. A voice synthesizing apparatus for obtaining a synthesized voice by connecting or overlapping voice waveforms, means for detecting a pitch period from a voice signal, detecting a minimum value of the voice signal, and Means for calculating an addition data by adding an offset so that the minimum value is non-negative; preparing a window having the pitch period length, moving the center of the window, and moving the addition data to a predetermined power. Means for covering the window, calculating an area of the windowing data at each center position of the window, and detecting a time axis coordinate of a center of the window that gives a maximum of the area; and centering around the time axis coordinate. Means for extracting the audio signal, and windowing and superimposing the extracted audio signal while shifting the pitch by a pitch period with the time axis coordinates being the center of superimposition. Speech synthesis apparatus according to. 9. The speech synthesizer according to claim 6, wherein a window length of the window is determined based on a pitch period obtained by a cepstrum analysis method.