JP3317458B2 - Voice synthesis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention is applied to, for example, converting an arbitrary text into a voice, selects a voice waveform according to an input phoneme and an input pitch frequency, and converts a pitch frequency of the selected voice waveform into an input pitch frequency. The present invention also relates to a speech synthesis method for controlling a selected speech waveform to have a length according to an input phoneme duration and a power according to an input magnitude to sequentially superimpose a waveform to obtain a synthesized speech.

[0002]

2. Description of the Related Art Conventional speech synthesis techniques include LPC.
A method of synthesizing speech using a spectrum parameter obtained by analysis and a driving sound source signal such as a pulse has been generally used. In this method, a voice is modeled and the voice is represented by a finite number of parameters. The voice can be easily analyzed and synthesized according to a certain procedure. Further, there is a method of selecting an appropriate waveform using a waveform dictionary and performing pitch control using a waveform superposition method to generate a synthesized sound (Japanese Patent Laid-Open No. 01-284898). In this method, the quality of the sound is good because the waveform of the natural sound is used as it is,
It is possible to obtain speech equivalent to natural speech.

[0003]

However, the former method using the spectral parameters and the driving sound source signal has a problem that the sound quality is poor. In the latter editing and synthesis, since the correlation between the pitch frequency and the spectral structure is not taken into consideration, the quality of a part of the voice is deteriorated by the pitch frequency changing process.

An object of the present invention is to provide a voice synthesizing method in which a quality control due to a pitch frequency changing process is reduced in a method of generating a synthesized voice by performing pitch control using a waveform superposition method.

[0005]

According to the present invention, in a method of generating a synthesized sound by performing pitch control using a waveform superposition method, a target formant frequency is determined according to an input pitch frequency and a pitch frequency of a selected waveform. Then, the formant frequency of the pitch-controlled voice waveform is converted according to the target formant frequency.

[0006]

FIG. 1 shows a flow chart of an embodiment of the method of the present invention. In the present invention, a phoneme (input phoneme) and a pitch frequency (input pitch frequency ) are input. For example, when a text is converted to speech, the text is analyzed to be a phoneme sequence, and the pitch for each phoneme is further determined. A pitch pattern that takes into account the continuity of the pitch frequency between the frequency and the phoneme,
A phoneme duration for each phoneme, a power (magnitude) for each phoneme, and a power pattern in consideration of continuity between the phonemes are set.

In the essential part of the present invention, each input phoneme which is a phoneme sequence and an input pitch frequency for each phoneme are mainly related. That is, first, in the waveform selection step S1, a waveform is selected from the waveform dictionary using an appropriate evaluation function based on the input phoneme and the input pitch frequency. Here, for the evaluation function and the waveform selection method, for example, a waveform selection method using an evaluation function (refer to “Waveform Selection Method in Waveform Editing Rule Speech Synthesis Method”, Sound Lecture 1-2-21 (1988)
-10) Hirokawa et al.) Can be used. In this method, an optimal waveform, that is, the smallest W, is selected from the waveform dictionary based on an evaluation function W such as Expression (1), using a target phoneme type, a pitch in a phoneme, a duration time, a phoneme coincidence rate, and the like as parameters. Select the waveform that becomes

[0008] _{W = Σω i | Pi t -Pi} s | (1) Σ weighting coefficients ωi is the i-th parameter i = 1 to n, Pi _t is a target value of the i-th parameter (input value), Pi _s the actual value of the i-th parameter (value of the waveform in the waveform dictionary). Next, in a formant frequency setting step S2, a formant frequency is determined from the input pitch frequency and the pitch frequency of the selected waveform using a table as shown in FIG. For example, if the pitch frequency of the selected waveform is P1 (P _i-1 ≤P1 <P _i ) and the pitch frequency is changed to the input pitch frequency P2 (P _j-1 ≤P2 <P _j ), the _first frequency is changed from FIG. the formant frequencies _{F1 t F1 t = F1 0 +} f1 ij [Hz] (F1 ₀ first formant frequency of the selected waveform) is set as (2). Such a table is prepared, for example, from the first formant to the third formant, which are considered to be greatly affected by the pitch, and referring to the table for each formant based on the pitch frequency and the input pitch frequency of the selected waveform. Determine the desired formant frequency.

In the pitch control step S3, the pitch frequency of the selected waveform is changed according to the input pitch frequency. Here, as a pitch frequency changing method, for example, a waveform superposition method (reference document “Study of pitch control method in waveform editing type rule synthesizing method”, sound lecture 1-4-7 (1990-3)
Hirokawa et al.) Can be used. In this approach, if the input pitch period, as shown in FIG. 3 is T, you gradually decreases front portion and rear portion with respect to the central portion has a length of 2T
Using windows 11, 1 cut waveform 12 selected in pitch unit, i.e. cut out using a window 11 for each pitch cycle T around the large peak waveform 12, overlapping their waveforms at the input pitch period T In addition, a waveform 13 having an input pitch period is obtained.

Next, in the formant conversion step S4, a waveform in which the formant frequency is changed to the target formant frequency determined in the formant frequency setting step S2 is generated with respect to the pitch control waveform obtained in step S3. Here, the formant is changed using, for example, the “formant control method” (Japanese Patent Application No. 4-261825). That is, as shown in FIG. 4, the spectral envelope (spectral density function) P (w) of the selected waveform whose pitch has been controlled in the pitch control step S3 is extracted (S ₁ ), and the voice waveform is subjected to fast Fourier transform (FFT). Is converted into a spectrum X (w) (S ₂ ), and a formant is extracted (S ₃ ). Frequency for the purpose of frequency of the formant in which the extracted, i.e. to change the purpose formant frequency determined by the formant frequency setting step S2 (S _4). For the selected voice waveform whose pitch has been controlled, the spectrum envelope corresponding to the voice waveform in which the formant frequency has been changed to the target frequency is determined (S
₅ ). The distortion D between the spectral density P ′ (2πΔT · F ′ _i ) at the converted formant frequency F ′ _i and the desired spectral density Pt (2πΔT · F ′ _i ) is determined (S
₆₎ If the distortion D is not sufficiently small (S _7), the flow returns to step S ₅ by changing the bandwidth to be changed formant (S _8). When distortion D is determined to be sufficiently small in step S _6, 'spectrum X corresponding with (w)' (w P ( w) and X (w) because the spectrum envelope P which finally determined in step S ₅ ) is obtained (S _9), the X 'Request spectral distortion d of the (w) and _{X (w) (S 10)} , the distortion d is sufficiently small that only lever _{(S 11), X' (} w ) Is X (w) and X ′
spectral envelope P of (w) "and (w) P '(w) as the flow returns to step S ₉ (S _12). distortion repeated this d is finally obtained at step S ₉ After sufficiently small inverse FFT of X '(w) which are converted into speech waveform (S _13).

The pitch frequency of the selected waveform is changed to the input pitch frequency, the formant frequency of the waveform whose pitch frequency has been changed is changed to the target formant frequency, and a voice waveform is obtained. Control the length and power of each phoneme according to the phoneme duration set according to the input text and the power pattern for each phoneme and the continuity between the phonemes, and sequentially superimpose and synthesize waveforms Get audio.

The present invention can be applied not only to the case where text is converted to speech but also to the case where a speech waveform is selected according to an input phoneme and an input pitch frequency and speech synthesis is performed.

[0013]

As described above, according to the present invention, the pitch frequency of the speech waveform is set to the target frequency in consideration of the correlation between the pitch frequency of the natural speech and the spectrum structure, and both pitch frequencies are set in accordance with both pitch frequencies. Since the formant frequency is changed, that is, the spectrum is changed, a voice whose pitch frequency is changed can be obtained without deteriorating the quality of the voice, and a synthesized voice having a quality close to natural voice can be obtained.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an embodiment of the present invention.

FIG. 2 is a diagram showing a formant frequency change table.

FIG. 3 is a waveform diagram schematically showing a pitch frequency changing process.

FIG. 4 is a flowchart illustrating an example of a formant conversion process.

Continuation of the front page (56) References JP-A-58-129500 (JP, A) JP-A-2-47700 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 21 / 04 G10L 13/00

Claims (1)

(57) [Claims] An optimum speech waveform is selected using a waveform dictionary according to an input phoneme and an input pitch frequency, and a frequency obtained by using a target formant frequency determination table according to a pitch frequency of the selected speech waveform and the input pitch frequency. If, to determine the desired formant frequency and a formant frequency of the selected speech waveform to obtain a pitch control selection 択波 shaped by superimposing waveforms cut out every period of the input pitch, from the pitch control selection 択波 type Change the extracted formant frequency to the target formant frequency, and pair with the audio waveform changed to the target formant frequency.
A corresponding spectrum envelope P ′ (w) is obtained, and the spectrum envelope P (w) of the selected waveform and the spectrum X
(W), the spectrum envelope P ′ (w)
The spectrum X ′ (w) is obtained, and the spectrum X ′ (w) is defined as a spectrum X (w).
The spectral envelope corresponding to the spectrum X '(w)
P "(w) as spectrum envelope P '(w)
X ′ (w) is repeated, and the spectrum X (w) and the spectrum X ′ (w) are
Spectrum X ′ (w) in which the distortion of
It is Once, inverse FFT and the resulting spectrum X '(w) and speech synthesis method according to feature to be converted speech waveform.