JP3532064B2 - Speech synthesis method and speech synthesis device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To synthesize a speech of high quality by accurate driving point detection. SOLUTION: A speech signal is divided into analytic frames (S200) and processed by low-pass filteration (S201), the largest value max nearby the center in an analytic frame is detected, and the time coordinate tm of the largest value is set as a tentative driving point (S204). Further, a maximum value is detected by tracing the frame back by a specific value tp×a from the tentative driving point (S205) and all maximum values detected in the traced-back time coordinate area are compared with a threshold value max×b individually (S208). When a maximum value is larger than the above-mentioned threshold value, the time coordinate of the maximum value is replaced as a driving point (S209), the speech signal is segmented about the driving point (S212) to generate an element piece previously, and the driving point in the element piece is put in the center of superposition to perform window superposition while it is shifted by a pitch cycle.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing method and a voice synthesizing apparatus for synthesizing an arbitrary voice according to a rule, and more particularly to a voice synthesizing method and a voice synthesizing apparatus for connecting voice waveforms to obtain a synthetic voice.

[0002]

2. Description of the Related Art As a conventional text-to-speech conversion apparatus, that is, a text-speech conversion apparatus for converting a text sentence into speech and outputting the speech, an apparatus composed of a text analysis section, a parameter generation section and a speech synthesis section is generally known. Has been. In the text analysis unit, a kanji / kana mixed sentence is input, morphological analysis is performed with reference to a word dictionary, pronunciation, accent, and intonation are determined, and a phonetic symbol with a prosody symbol (intermediate language) is output. In the parameter generator,
The pitch frequency pattern, phoneme duration, etc. are set. The voice synthesis unit performs a voice synthesis process. As a speech synthesis process in this speech synthesis unit, a linear prediction method or the like has been used before, but these methods result in deterioration of information. That is, the deterioration of the sound quality due to the separate handling of the vocal tract information and the sound source information that are originally related to each other and the deterioration of the sound quality due to the restriction due to the modeling of the sound generation process are unavoidable. Therefore, in recent years, vocal tract information and sound source information are not clearly separated, and the original speech waveform is used as it is, that is, the subtle fluctuations contained in the original speech waveform are utilized without artificial modeling. However, a method for obtaining high-quality synthesized speech with little deterioration has been used.

As a method of directly using a voice waveform, there is a conventional method: "FJ CHARPENTIER, MG STELL.
A, DIPHONE SYNTHESIS USING AN OVERLAP-ADD TECHNIQU
E FOR SPEECH WAVEFORMS CONCATENATION ”, Proc.Int.C
onf.ASSP, TOKYO, 1986 PP 2015-2018 ”is known. In this method, a pitch mark (reference point) is attached to the voice waveform in advance, the voice waveform is cut out around the position of the pitch mark, and the pitch mark position is overlapped by shifting the pitch mark position in accordance with the synthetic pitch period during synthesis. PSOLA (Pitch-Synchronous O
verlap Add method "Pitch synchronization waveform superposition method").

FIG. 2 is cited from the above document and is a schematic diagram showing a pitch synchronization waveform superimposing method for superimposing a voice waveform while changing a pitch. This schematic diagram shows an example in which the pitch period during synthesis is increased (pitch is lowered) as compared with the time of analysis (during segment production).

In the pitch synchronization waveform superimposing method, the pitch can be changed as necessary, and thus it has been widely used as a voice synthesizing unit in text-to-speech conversion. In this case, the pitch mark needs to be attached to a specific position for each pitch of the voice waveform, and the following positions have been proposed as the position of the pitch mark.

(1) For setting the peak of the voice waveform as the pitch mark setting position, for example, Japanese Patent Laid-Open No. 4-3729.
There is a "voice pitch conversion method" described in Japanese Patent Publication No. 99.

In this case, since the energy is concentrated at the local peak position of the voice waveform, it is considered to be suitable for storing the spectrum of the cut-out waveform.

(2) For setting the peak of the short time power as the setting position of the pitch mark, for example, "" Examination of speech synthesis system with waveform segment connection "Tsune Kawai, Yoshio Higuchi, Tohru Shimizu, Seiichi Yamamoto Technical report SP93-9 (1993-05)
The Institute of Electronics, Information and Communication Engineers.

Also in this case, as in the case of (1) above,
Since the energy is concentrated at the local peak position of the short-time power of the voice waveform, it is considered to be suitable for storing the spectrum of the cut-out waveform.

(3) A method in which the peak after the pitch filter is set as the pitch mark setting position is disclosed in, for example, Japanese Patent Application Laid-Open No. 7-
"Speech synthesis method and device" described in Japanese Patent No. 72897.
There is.

The peak after the pitch filter is the peak of the drive waveform of a 1-pitch vocal cord, and according to the above document, it is reported that it is a good representative of the pitch interval.

(4) As an example in which the 15% delay point of the impulse driving point is set as the pitch mark setting position, for example, "A study of pitch waveform extraction position" Yasuhiko Arai, Nishimura
Hiroshi Yoshida, Toshimitsu Minowa IEICE Technical Report SP95-8 (1995-0
5) The Institute of Electronics, Information and Communication Engineers.

According to this document, it is reported that the spectral distortion is minimized.

(5) As an example of setting the glottal closing point as the pitch mark setting position, for example, "About Japanese text-to-speech synthesis system using waveform superposition method" S. Sakamoto, T. Saito, K. Suzuki, Y. Hidehashi, Kobayashi There is Meishin Giho SP95-6 (1995-05) The Institute of Electronics, Information and Communication Engineers.

The glottal closing point in this document is considered to be the same as the impulse driving point (excitation point of one pitch waveform). In order to stably extract this glottal closing point,
Dynamic Wavelet transform is used.

[0016]

However, the conventional pitch mark position as described above has the following problems.

As shown in FIGS. 3 and 4, the peak of the voice waveform of (1) may be difficult to recognize depending on the phoneme (particularly / a / or / e /). That is, when the pitch mark is automatically added, it is easy to make a mistake in the position,
When it is given manually, it is difficult to judge due to subtle differences.

Here, in FIG. 3, / e / pronounced as / he /
In the voice waveform of the part, the pronunciation of / me / in Figure 4 is / e /
Is the voice waveform of the part. In these figures, the pitch mark should be originally given to all points a. However, when it is attempted to automatically extract the peaks of the voice waveform and add pitch marks, b, a, a, a, a, b in FIG.
In FIG. 4, pitch marks are added to the points b, b, b, a, a, and a. If you do voice synthesis based on this,
Since the pitch interval fluctuates, the synthesized sound has distorted sound quality.

In addition, when a pitch mark is manually added to the peak of a voice waveform, there is no clear reference as to whether it is the point a or the point b. I am lost and not efficient.

On the other hand, in the above-mentioned (3) in which the peak after the pitch filter is used as the set position of the pitch mark, there are the following problems. According to an experiment by the applicant, the cutoff 25 described in JP-A-7-72897 is disclosed.
With the 6 Hz LPF, the waveform is considerably rounded. For this reason, there is a deviation from the peak position of the waveform, the pitch fluctuates due to this deviation, and the sound becomes jumbled and distorted. Rather, the result of (1) is relatively good when the peak of the voice waveform is set as the pitch mark setting position.

In the case of (2) in which the short-time power peak is set as the set position of the pitch mark, the maximum value and the minimum value are evaluated equally, so that some utterers cause pitch fluctuations, resulting in a synthesized sound. Distortion may occur. In the case where the pitch mark setting position is the 15% delay point of the impulse driving point in (4), the processing amount for specifying the setting position is large, resulting in a delay in processing, and depending on the individual and the type of phoneme. , The delay point of 15% is not always the best. In the case where the glottal closing point in (5) is set as the pitch mark setting position,
The Dynamic Wavelet transform performed for extracting the glottal closing point has a large amount of processing, and causes a delay in processing as in (4) above.

In view of the above problems, the present invention has been made in consideration of the fact that the influence of the glottal drive impulse theoretically appears at the first peak of the voice waveform in the voice generation process, and there is little pitch fluctuation. An object of the present invention is to realize a high-quality voice synthesizing method and voice synthesizing device by enabling pitch mark setting.

[0023]

In order to solve the above-mentioned problems, a voice synthesizing method according to a first aspect of the present invention includes a step of detecting a first maximum point of a one-pitch waveform of a voice signal, A voice synthesis unit is created in advance by a step of centering the first maximum point and cutting out the voice signal, and the maximum point in the voice synthesis unit is set as the center of superimposition, and a pitch period The feature is that the windows are overlapped while being shifted.

As described above, since the first maximum point of the one-pitch waveform of the audio signal is the center point of the audio signal to be cut out, the center point for superimposing can be easily set by a simple process. . This makes it possible to reduce pitch fluctuation and reduce spectral distortion. As a result, the rumbling sound was reduced.

The voice synthesizing method according to the second invention is characterized in that the low-pass filtered waveform of the voice signal is used when the first maximum point is detected.

As described above, when the low-pass filtered waveform of the audio signal is used, fine fluctuations of high frequency are removed to form a smooth waveform, which facilitates detection of the maximum point. This allows
The center point for superimposing can be easily set.

In the speech synthesis method according to the third aspect of the invention, the step of dividing the speech signal into analysis frames, the step of low-pass filtering the speech signal, and detecting the maximum value near the center of the frame within the analysis frame. Then, the process of setting the time coordinate of the maximum value as the temporary driving point, and the process of detecting the maximum value in the time coordinate range traced back from the time coordinate of the temporary driving point,
The process of individually comparing all the maximum values detected in the traced time coordinate area and the threshold value, and as a result of the comparison with the threshold value, when the maximum value is larger, the time coordinate of the maximum value is driven. A voice synthesis unit is created in advance by the process of cutting out the voice signal by centering the drive point as a center, and the drive point in the voice synthesis unit is set as the center of superposition, and the pitch cycle is set. It is characterized in that the windows are overlapped while being shifted.

As described above, the maximum value in the vicinity of the center of the frame is set as the temporary driving point, the maximum value is detected in the time coordinate area traced back from that point, and the final value is compared in comparison with the temporary driving point. Since the driving point that is the center of cutting is set in,
It is possible to detect the first peak as an influence of the glottal drive impulse in the voice generation process, and it is possible to accurately set the pitch mark.

Further, the center point at the time of superimposing can be easily set by a simple process, and the fluctuation of the pitch can be reduced and the spectral distortion can be reduced. As a result, the rumbling sound was reduced.

Further, by treating the speech units in units of a fixed length and performing frame processing, it becomes easier to control the speech waveform data at the time of speech synthesis.

In the voice synthesizing method according to the fourth aspect of the present invention, the pitch period in the analysis frame is detected, and the maximum value is detected by tracing back from the temporary driving point in a time coordinate range of a predetermined number of times the pitch period. It is characterized by doing.

With the above configuration, the area for detecting the maximum value is limited to a certain range, so that the pitch mark can be set accurately.

A speech synthesis method according to a fifth aspect of the invention is characterized in that the threshold value is set to a predetermined multiple of a maximum value detected as a temporary driving point near the center of the frame in the analysis frame. .

With the above configuration, the threshold value is set to a predetermined number of times the maximum value detected as the temporary driving point, so that a small extra maximum value will not be erroneously detected. As a result, even if a noise component is superposed, stable operation is achieved without being affected by an extra maximum value.

In the voice synthesizing method according to the sixth aspect of the present invention, a predetermined value is set according to the type of phoneme of the voice signal, and a maximum value is set in a time coordinate range that is traced back from the temporary driving point by the predetermined value. It is characterized by detecting.

With the above configuration, the region for detecting the maximum value is limited to a certain range, so that the pitch mark can be set accurately.

A speech synthesizer according to a seventh aspect of the present invention is a maximum point detecting means for detecting a first maximum point of a one-pitch waveform of a voice signal, and centering the voice signal around the first maximum point. Voice signal cutting-out means for cutting out the voice signal, a voice synthesis unit storage means for storing the voice synthesis unit cut out by the voice signal cutting means, and a voice stored in the voice synthesis unit storage means. It is characterized by further comprising a speech synthesis unit for windowing and superimposing while shifting the local maximum point in the synthesis segment as a center of superimposition and shifting it by a pitch period.

As described above, since the first maximum point of the one-pitch waveform of the audio signal is set as the center point of the audio signal to be cut out, the center point for superimposing can be simplified as in the first invention method. It can be easily set by various processes.
This makes it possible to reduce pitch fluctuation and reduce spectral distortion. As a result, the rumbling sound was reduced.

The speech synthesizer according to the eighth aspect of the invention is characterized by using the low-pass filtered waveform of the speech signal when detecting the first local maximum.

With the above configuration, since the low-pass filtered waveform of the audio signal is used, the maximum point can be easily detected and the center point for superimposing can be easily set, as in the second invention method. You can

A speech synthesizer according to a ninth aspect of the invention is a dividing means for dividing an audio signal into analysis frames, a low-pass filtering means for low-pass filtering the audio signal, and a portion near the center of the frame within the analysis frame. Maximum value detecting means for detecting the maximum value and setting the time coordinate of the maximum value as a temporary driving point, and maximum value detecting means for detecting the maximum value in the time coordinate range traced back from the time coordinate of the temporary driving point. And, comparing means for individually comparing all the maximum values detected by the maximum value detecting means with the threshold value, and the result of the comparison by the comparing means, if the maximum value is larger, the time coordinate of the maximum value is calculated. A voice signal segmentation unit that replaces as a drive point and centers the drive point to extract the voice signal, and a voice synthesis unit segment that stores the voice synthesis unit segment cut out by the voice signal segmentation unit. Means and a voice synthesizer for windowing and superimposing while shifting the driving point in the voice synthesizer stored in the voice synthesizer storage means as a center of superposition while shifting by a pitch period. .

As described above, the maximum value in the vicinity of the center of the frame is set as the temporary driving point, the maximum value is detected in the time coordinate area traced back from that point, and the final value is compared in comparison with the temporary driving point. Since the driving point that is the center of cutting is set in,
Similar to the method of the third aspect, the first peak can be detected as the influence of the glottal drive impulse in the voice generation process, and the pitch mark can be set accurately.

Further, the spectrum distortion is also reduced, and the rumbling sound is reduced. Furthermore, it becomes easy to control the voice waveform data during voice synthesis.

A speech synthesizer according to a tenth aspect of the invention detects a pitch period in the analysis frame and detects a maximum value by tracing back from the temporary driving point in a time coordinate range of a predetermined multiple of the pitch period. It is characterized by doing.

With the above construction, the region for detecting the maximum value is limited to a certain range, so that the pitch mark can be set accurately as in the case of the method of the fourth aspect.

The speech synthesizer according to the eleventh aspect of the invention is characterized in that the threshold value is set to a predetermined multiple of the maximum value detected as a temporary driving point near the center of the frame in the analysis frame. .

With the above configuration, the threshold value is set to a predetermined number of times the maximum value detected as the tentative driving point, so that a small extra maximum value is erroneously detected as in the fifth invention method. There is nothing to do. As a result, even if a noise component is superposed, stable operation is achieved without being affected by an extra maximum value.

The voice synthesizer according to the twelfth aspect of the invention sets a predetermined value in accordance with the type of phoneme of the voice signal, and sets a maximum value in the time coordinate area that is traced back by the predetermined value from the temporary driving point. It is characterized by detecting.

In this case also, as in the eleventh invention,
Accurate pitch mark setting is possible and stable operation is achieved.

[0050]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the accompanying drawings.

[First Embodiment] A voice synthesizing method and a voice synthesizing apparatus according to the first embodiment will be described below. FIG. 5 is a block diagram showing the configuration of the speech synthesizer according to the first embodiment.

In the text analysis unit 101 in the figure, when a kanji / kana mixed sentence is input, morphological analysis is performed with reference to the word dictionary 102 to determine readings, accents and intonations of the kanji / kana mixed sentence, and prosodic symbols. Output phonetic symbols (intermediate language). The parameter generator 103 sets a pitch frequency pattern, phoneme duration, and the like. The text analysis unit 101, the word dictionary 102, and the parameter generation unit 103 are the same as the conventional ones.

The voice synthesis unit 104 performs a voice synthesis process. That is, a segment in the segment dictionary 105 is selected, and a windowing unit 106 applies a time window of a later-described time window length Tpl to the segment so that the driving point described later becomes the center of superimposition. Then, the voice synthesis unit 104 performs voice synthesis by the pitch synchronization waveform superposition method.

Here, the time window length Tpl is set as Tpl = C0 × min (Tpa, Tps) where Tpa is the pitch period during analysis and Tps is the pitch period during synthesis. Note that C0 has a value of about 2.0.

The segment dictionary 105 is a dictionary in which segments are written. The segment is created by the segment creating unit 107. The segment creation unit 107 is the main part of the present invention and has the processing function shown in the flowchart of FIG.

The processing in the segment creating unit 107 will be described with reference to FIG. When an audio signal is input to the segment creating unit 107 by the audio signal input unit 108 including a data disc, first, in step S200, the input audio signal data is divided into analysis frames. This analysis frame is audio signal data divided into sections of a fixed length. In this embodiment, one frame length is 32 msec and 8 m
It is delimited so that it shifts to the next frame and moves to the next frame.
Here, the total number of frames is N. In addition, the waveform data is Xn (l), [n = 1, 2, ..., N (frame), l = 1, 2,
..., Fr (points)]. Fr is the frame length (3
2 ms) and the number of samples in one frame determined by the sampling frequency FS. In this embodiment, Fr = 32
× FS / 1000.

In step S201, the waveform data Xn (l)
Is applied to LPF (low pass filter). This is to remove fine fluctuations in the peak fusion processing described below. Further, the waveform data after the LPF processing is subjected to the processing for correcting the delay due to the LPF, and the corrected waveform data is set to yn (l). Note that n = 1, 2,
…, N (frame), l = 1, 2,…, Fr (point)
Is. The LPF used here is preferably a well-known FIR type digital filter having a linear phase for processing based on a waveform to be performed later. Further, this LPF processing is collectively performed in advance.

In step S202, the frame number n of the analysis frame to be processed is initialized. Step S20
3, the pitch period t of the audio signal in the nth frame is
Detect p. As a method of detecting the pitch period tp, a method of detecting the peak interval of the waveform can be considered as a simple method, but in the present embodiment, the cepstrum method is used. This is for more accurately calculating the pitch period. In this cepstrum method, the pitch period tp is detected in the processing steps shown in FIG. First, step S40
The time waveform is input in 1 and windowing is performed in step S402. Next, the windowed time waveform is subjected to discrete Fourier transform (DFT) in step S403, and the square root of the sum of squares of its real part and imaginary part is logarithmically converted in step S404. Then, inverse Fourier transform (IDFT) is performed in step S405, and the cepstrum component is obtained and output in step S406. Thus, the cepstrum method is
The convolution operation is converted into an additive operation. Since the voiced sound signal of speech is a convolution of the sound source component with vocal tract information, the cepstrum method is suitable for separating the two. If the input signal is a voiced voice signal, change the pitch period
If T0, the sound source component has high kefrenshii (long time region)
Appears in the vicinity of T0, and the vocal tract component appears as a low-Keffrenchy (short-time region) component. In order to obtain the pitch period from the cepstrum, the peak of the high kefrenshi part can be obtained and the time from the time origin to this peak can be measured.

Then, in step S204 in FIG. 1, the maximum value (max) near the frame center of each analysis frame.
And the time coordinate tm is detected. The pitch period tp obtained in step S203 is used to determine the vicinity of the center of the frame. A predetermined number of times the pitch period tp is set in the vicinity of the frame center before and after the time coordinate of the frame center. In this embodiment, 0.6 tp before and after the time coordinate at the center of the frame is near the center. As a result, the maximum value max near the center of the frame is: max = maxmum {yn (l); l = (Fr / 2) −
0.6tp, ..., (Fr / 2) + 0.6tp} = yn (tm). The time coordinate tm of the maximum value max is set as a temporary driving point.

In step S205, all the maximum values are detected in the time coordinate range that is a predetermined value back from the time coordinate tm set as the temporary drive point. The section for a predetermined value traced back is [tm-tp × a, tm). Here, the constant a is 0
It is a value between 1 and 1. This constant a is set by statistically calculating the best numerical value on the basis of experimental data. In this embodiment, it is set to about 0.25 to 0.40.

Here, the maximum value is Pk, the total number of detected maximum values Pk is M, and the time coordinate of the maximum value Pk is tpk (k
= 1, 2, ..., M), and Pk = yn (tpk).

Next, in steps S206 to 211, peak fusion processing is performed.

First, in step S206, the time coordinate tm of the maximum value which is a temporary driving point is set as the initial value of the pitch mark K. In step S207, the maximum value counter j is set to M.

In step S208, the magnitude of the threshold value and the maximum value Pj is determined. The threshold value is set to a predetermined multiple of the maximum value max near the center of the frame. That is, max
Set to xb. Here, the constant b is a value between 0 and 1. As with the constant a, the constant b is set by statistically calculating the best numerical value based on experimental data. In this embodiment, it is set to about 0.6 to 0.8.

As a result of the determination in step S208, when the maximum value Pj is larger than the threshold value max × b, the process proceeds to step S209, and the time coordinate tm of the pitch mark K is set to the time coordinate of the maximum value Pj. Replace with tpk. In the determination of step S208, the maximum value Pj is ma
If it is smaller than xxb, the process jumps to step S210 without performing the process of step S209. Then, in step S210, the maximum value counter j is decremented by 1, and in step S211, it is determined whether or not the value of the maximum value counter j after the subtraction is 0. If it is not 0, the process returns to step S208, and the processes of steps S208 to S211 are repeated. As a result, the pitch mark K as the true driving point is finally obtained from the temporarily set driving point.

FIG. 7 shows a schematic diagram of the peak fusion process in steps S206 to 211.

FIG. 7 shows (/ a) of the pronunciations of / ra /.
7A shows the speech waveform before the LPF processing, and FIG. 7B shows the speech waveform after the LPF processing which is the low-pass filtered waveform with the delay corrected. . Figure 7
In both the waveforms (A) and (B), there are four main maximum values in one pitch section. The low-pass filtered waveform of FIG. 7 (B) is shown in FIG. 7 (A).
Compared with the voice waveform of, the high frequency fine fluctuations are removed, and thus it is suitable for pitch mark addition. However, if the maximum value is simply extracted from the low-pass filtered waveform after the LPF processing, all the points are b points. It is considered that the influence of the glottal drive impulse logically appears at the first peak in the voice generation process. Therefore, it is inaccurate to simply give the pitch mark to the low-pass filtered waveform after the LPF processing (point b is used as the pitch mark).

On the other hand, when the peak fusion processing of the present embodiment is performed, the maximum value b point in FIG. 7 (B) obtained in step S204 is moved back to the maximum value b point in step S205. There is a sufficiently large maximum point a, and the pitch mark K is replaced with the point a in steps S208 and S209.

In step S212 in FIG. 1, L audio data Xn are cut out before and after the pitch mark K obtained by the peak fusion process and centered so that K is located at the center. In addition, L minute was set to 12 msec here. This is a value obtained by a preliminary experiment by the inventor of the present invention in which a man having a larger pitch period than a woman has a margin in the longest pitch period.

In step S213, the speech data extracted in step S212 is stored in a storage medium such as a data disk as a segment in the nth frame in the segment dictionary 105.
Are sequentially written. In step S214, it is determined whether writing of the segment has been completed for all analysis frames. If this writing has not been completed, the frame number is updated in step S215 and then step S2 is performed.
Returning to step 03, the processing from step S203 to step S213 is continued. If it is determined in step S214 that all analysis frames have been processed, the segment dictionary 105
The data disc closing process (not shown) and the like are performed, and the operation of the segment creating unit 107 ends.

The target segment is appropriately selected from the segment dictionary 105 in which the segment created by the above processing is written, and the windowing unit 106 performs windowing, and the speech synthesis unit 104. The voice synthesis process is performed.

[Effects] (1) As described above, since the first maximum point of the 1-pitch waveform of the audio signal is set as the center point of the audio signal to be cut out, the center point at the time of superimposing can be easily performed by simple processing. Can be set to. This makes it possible to reduce pitch fluctuation and reduce spectral distortion. As a result, it is possible to reduce audible rumbling sounds.

(2) Since a tentative driving point is defined and the driving point on which actual processing is performed is specified in comparison with this tentative driving point in the time domain traced back from that point, the first driving impulse as the influence of the glottal driving impulse is determined. It is possible to detect the maximum point of, and it is possible to set the pitch mark accurately. As a result, the spectrum distortion is small, the operation is stable, and the processing amount can be reduced.

(3) Since the low-pass filtered waveform of the voice signal is used in the production of the segment, the fine fluctuation of the high frequency is removed to form a smooth waveform, and the maximum point can be easily detected.

(4) Since the time for tracing back from the temporary driving point tm to detect the maximum point is limited to tp × a, it is possible to prevent the tracing back too much and accurately capture the first peak. .

(5) Since the threshold value is set to a predetermined multiple of the maximum value max, that is, max × b, it is possible to prevent erroneous fusion of small extra maximum values. Therefore, even if a noise component is superimposed, it is possible to operate stably without being affected by an extra maximum value.

(6) By handling a voice segment with a fixed length as a unit and performing frame processing, voice waveform data can be easily controlled during voice synthesis.

[Second Embodiment] Next, a second embodiment of the present invention will be described.

FIG. 8 is a block diagram showing the arrangement of a speech synthesizer according to the second embodiment.

The overall structure of the speech synthesizer according to this embodiment is the same as that of the speech synthesizer according to the first embodiment. The feature of the speech synthesis method of the present embodiment is that the segment creating unit is different from the segment creating unit 107 according to the first embodiment in that
The point is that the processing functions shown in the flowchart of FIG.

The processing in this segment creating unit will be described with reference to FIG.

In the speech synthesizer of this embodiment, the processing from step S200 to step S203 is the same as the function of the segment creating unit 107 according to the first embodiment. That is, the audio signal data is divided into analysis frames in step S200, and the LP is divided in step S201.
F is multiplied, the frame number n is initialized in step S202, and the pitch period tp is detected in step S203. The details of these processes are the same as each process in the first embodiment.

Then, in step S303, a predetermined value (c) is set according to the type of phoneme of the analysis frame.
The predetermined value (c) may be generated by a rule or may be configured to draw a table based on the phoneme of the analysis frame.

In step S304, the maximum value (max) near the frame center of each analysis frame and its time coordinate tm are detected. In addition, to determine the vicinity of the center of the frame,
Using the pitch period tp obtained in step S203,
The same process as step S204 of the first embodiment is performed. As a result, the maximum value max near the center of the frame is: max = maxmum {yn (l); l = (Fr / 2) −
0.6tp, ..., (Fr / 2) + 0.6tp} = yn (tm). The time coordinate tm of the maximum value max is set as a temporary driving point.

In step S305, all the maximum values are detected in the time coordinate area which is traced back by the predetermined value c from the time coordinate tm set as the temporary driving point. The section going back is [tm
-C, tm).

The maximum value is Pk, the total number of detected maximum values Pk is M, and the time coordinate of the maximum value Pk is tpk (k = 1, 2,
, M), and Pk = yn (tpk).

Next, in steps S306-311, peak fusion processing is performed. This peak fusion processing is also similar to that of the first embodiment. That is, the time coordinate tm is set as the initial value of the pitch mark K in step S306, and the maximum value counter j is set to M in step S307.
To the threshold value max in step S308.
The magnitude between xb and the maximum value Pj is determined, and step S309
At the time coordinate tpk of the maximum value Pj
The maximum value counter j is decremented by 1 in step S310, and it is determined in step S311 whether the value of the maximum value counter j is 0. Then, the processes of steps S308 to S311 are repeated until the value of the maximum value counter j becomes zero.

Next, in step S312, the voice data Xn is cut out and the pitch mark K is centered in the center, and the cut-out voice data is sequentially written as the segment dictionary 105 in step S313, and in step S31.
In 4, it is determined whether or not the writing of the segment has been completed for all the analysis frames. If this writing is not completed, the frame number is updated in step S315, the process returns to step S203, and the processes from step S203 to step S313 are continued. If it is determined in step S314 that the processing of all analysis frames has ended, the data disc closing processing of the element dictionary 105 (not shown) is performed, and the operation of the element generation unit 107 ends.

The target segment is appropriately selected from the segment dictionary 105 in which the segment created by the above processing is written, and windowing is performed by the windowing section 106, and the voice synthesis section 104 is selected. The voice synthesis process is performed.

[Effect] With the above configuration, the same operation and effect as the voice synthesizing method and the voice synthesizing apparatus of the first embodiment can be obtained.

Further, in the present embodiment, the predetermined value (c) is set according to the type of phoneme of the analysis frame, and the range in which the maximum value is detected retroactively from the temporary driving point is the range of the predetermined value (c). Since it is limited to, the following effects can be obtained.

(1) The predetermined value (c) can be set a priori according to the type of phoneme, and it becomes possible to control within a proper range according to the data to be processed.

(2) Since the range to be traced back is set to the predetermined value (c), the “tp × a” processing which is performed for each analysis frame in the first embodiment is not necessary. As a result, the processing amount can be reduced, and the processing device can be simplified or the processing speed can be increased.

[Modification] In the first and second embodiments, the LPF processing is performed collectively, but it may be performed for each frame.

Although the cepstrum method is used as the pitch period detection method, another method such as an autocorrelation method or a modified autocorrelation method which is an autocorrelation of linear prediction residuals may be used.

Further, the voice synthesis method and voice synthesis apparatus of each of the above embodiments uses the voice synthesis unit to change the pitch of the original voice and change the pitch of the voice, that is, pitch mark setting in a so-called voice pitch conversion apparatus. It is possible to adapt to processing in various audio output devices such as.

[0097]

As described in detail above, according to the present invention, the following effects can be obtained.

(1) Since the first maximum point of the 1-pitch waveform of the audio signal is the center point of the audio signal to be cut out,
It is possible to easily set the center point when the waveforms are superimposed. This makes it possible to reduce pitch fluctuation and reduce spectral distortion. As a result, it is possible to reduce audible rumbling sounds.

(2) Since a tentative driving point is defined and the driving point on which actual processing is performed is specified in comparison with this tentative driving point in the time domain traced back from that point, the first driving pulse as the influence of the glottal driving impulse is determined. It is possible to detect the maximum point of, and it is possible to set the pitch mark accurately. As a result, the spectrum distortion is small, the operation is stable, and the processing amount can be reduced.

(3) Since the low-pass filtered waveform of the audio signal is used in the production of the segment, the fine fluctuation of the high frequency is removed to form a smooth waveform, and the maximum point can be easily detected. This makes it possible to easily set the center point when superimposing.

(4) Since the time for tracing back from the tentative driving point for detecting the maximum point is limited, it is possible to prevent the tracing back too much and accurately capture the first peak.

(5) Since the threshold value is set to a predetermined multiple of the maximum value, a small extra maximum value is erroneously fused,
It will not be affected by noise and will be able to operate stably.

(6) By handling a voice segment with a fixed length as a unit and performing frame processing, it becomes easy to control voice waveform data during voice synthesis.

[Brief description of drawings]

FIG. 1 is a flowchart showing a processing function in a segment creating unit of a speech synthesizer according to a first embodiment of the present invention.

FIG. 2 is a schematic diagram showing a pitch synchronization waveform superimposing method for superimposing a voice waveform while changing a pitch.

FIG. 3 is a schematic diagram showing an example of pitch mark setting for a voice waveform of a portion of / e / pronounced / he /.

FIG. 4 is a schematic diagram showing a pitch mark setting example for a voice waveform of a portion of / e / pronounced / me /.

FIG. 5 is a block diagram showing a configuration of a speech synthesizer according to the first embodiment of the present invention.

FIG. 6 is a flowchart illustrating a cepstrum method.

FIG. 7 shows a schematic diagram of peak fusion processing in a segment creating unit.

FIG. 8 is a flowchart showing a processing function in a segment creating unit of the speech synthesis device according to the second embodiment of the present invention.

[Explanation of symbols]

101: text analysis unit, 102: word dictionary, 103:
Parameter generation unit, 14: voice synthesis unit, 105: segment dictionary, 106: windowing unit, 107: segment creation unit, 108:
Audio signal input section.

Claims (12)

(57) [Claims] 1. A speech synthesizer comprising: a step of detecting a first maximum point of a one-pitch waveform of a voice signal; and a step of centering the first maximum point with respect to the voice signal to cut out the voice signal. Create a piece in advance, with the maximum point in the speech synthesis element as the center of superimposition,
A voice synthesis method characterized by performing windowing and superimposing while shifting by a pitch period. 2. The voice synthesizing method according to claim 1, wherein a low-pass filtered waveform of the voice signal is used when the first maximum point is detected. 3. A step of dividing an audio signal into analysis frames, a step of low-pass filtering the audio signal, a maximum value in the vicinity of a frame center in the analysis frame is detected, and a time coordinate of the maximum value is calculated. The process of setting as a temporary driving point, the process of detecting the maximum value in the time coordinate range traced back from the time coordinate of the temporary drive point, and the total maximum value detected in the traced time coordinate range and the threshold value are set. As a result of comparison between the process of individually comparing and the threshold value, when the maximum value is larger, the time coordinate of the maximum value is replaced as a driving point, and the audio signal is centered around the driving point to output the audio signal. Create a voice synthesis unit in advance by the process of cutting out, with the driving point in the voice synthesis unit as the center of superimposition,
A voice synthesis method characterized by performing windowing and superimposing while shifting by a pitch period. 4. The voice synthesis method according to claim 3, wherein the pitch period in the analysis frame is detected, and the maximum value is traced back from the temporary driving point in a time coordinate range that is a predetermined multiple of the pitch period. A method for synthesizing speech, which is characterized by detecting. 5. The speech synthesis method according to claim 3, wherein the threshold value is set to a predetermined multiple of a maximum value detected as a tentative driving point near the center of the frame in the analysis frame. A method for synthesizing speech. 6. The voice synthesizing method according to claim 3, wherein a predetermined value is set according to a phoneme type of the voice signal, and the maximum value is obtained in a time coordinate range that is traced back from the temporary driving point by the predetermined value. A voice synthesis method characterized by detecting a value. 7. A local maximum point detecting means for detecting a first local maximum point of a one-pitch waveform of an audio signal, and an audio signal clipping for centering the first local maximum point with respect to the audio signal to cut out the audio signal. Means, a voice synthesis unit storage means for storing the voice synthesis unit cut out by the voice signal cutout unit, and the maximum point in the voice synthesis unit stored in the voice synthesis unit storage unit. A voice synthesizing apparatus comprising: a voice synthesizing unit that performs windowing and superimposing while shifting by a pitch period as a center of superimposition. 8. The speech synthesizer according to claim 7, wherein a low-pass filtered waveform of the speech signal is used when the first local maximum point is detected. 9. A dividing means for dividing an audio signal into analysis frames, a low-pass filtering means for low-pass filtering the audio signal, a maximum value near a frame center in the analysis frame, and a maximum value thereof. Maximum value detecting means for setting the time coordinate of as the temporary driving point, maximum value detecting means for detecting the maximum value in the time coordinate range traced back from the time coordinate of the temporary driving point, and the maximum value detecting means. The comparison means for individually comparing all the maximum values with the threshold value and the comparison result by the comparison means, when the maximum value is larger, the time coordinate of the maximum value is replaced as the driving point, and the driving point is A voice signal clipping means for centering the voice signal by centering it, a voice synthesis element storage means for storing the voice synthesis element clipped by the voice signal clipping means, and the voice synthesis element. As the center of superimposing the driving point in the speech synthesis fragments stored in 憶 means, speech synthesis apparatus characterized by comprising a speech synthesizer for windowing superimposed while shifting pitch cycle. 10. The speech synthesizer according to claim 9, wherein the pitch period in the analysis frame is detected, and the maximum value is traced back from the temporary driving point within a time coordinate range of a predetermined multiple of the pitch period. A speech synthesizer characterized by detecting the following. 11. The speech synthesis apparatus according to claim 9, wherein the threshold value is set to a predetermined multiple of a maximum value detected as a tentative driving point near the center of the frame in the analysis frame. A speech synthesizer characterized by. 12. The speech synthesizer according to claim 9, wherein a predetermined value is set according to a phoneme type of the voice signal, and the maximum value is obtained in a time coordinate range that is traced back from the temporary driving point by the predetermined value. A voice synthesizer characterized by detecting a value.