JPH0772897A - Method and device for synthesizing speech - Google Patents

Abstract

PURPOSE:To provide method and device for synthesizing a speech occurring no large waveform discontinuity in a connection point of a sound waveform element piece, performing waveform connection where a spectrum is changed smoothly and occurring less noise. CONSTITUTION:In a voice synthesizer connecting plural sound waveforms and obtaining a synthetic sound, the waveform of the end of the preceding sound waveform and the waveform of the tip of the succeeding sound waveform are cut out by a window function provided with a double length of a pitch length in a waveform cut-out part in sound waveform buffers 4, 5, and these cut out waveforms are added matching with the central position of the window function while weighting in an interpolation waveform generation part 9, and the preceding sound waveform and the succeeding sound waveform are connected in a waveform connection part 10 through an interpolation waveform.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing method and apparatus, and more particularly to a speech synthesizing method and apparatus for connecting speech waveforms to obtain a desired synthesized speech.

[0002]

2. Description of the Related Art Speech synthesis based on a rule capable of synthesizing arbitrary speech is conventionally performed by an LPC (Linear Pr
It was common to use the editive coding technology. The reason is that the LPC analysis / synthesis method assumes independence of vocal tract information and sound source information, so that they can be handled separately, and speech with various intonations can be processed from a relatively small number of speech units. This is because they can be synthesized.

However, with regard to voice, there is essentially an inseparable relationship between vocal tract information and sound source information. Nevertheless, as a result of assuming that they are independent, the spectrum mismatch between the two occurs due to this, resulting in abnormal amplitude, nasal stuffy noise, and other quality deterioration. It was Therefore, a method of using the original waveform as it is for speech synthesis has been researched and developed because the vocal tract information and the sound source information are not separated and the analysis and synthesis method itself does not cause deterioration of synthesized speech. This method cuts out a voice waveform segment from a large amount of original voice data, along with voiced phonological environment information, fundamental frequency information, duration information, amplitude information and other voice information, and stores it as a voice waveform segment file. Then, an arbitrary voice is synthesized by selecting and connecting a voice waveform segment that best matches a desired phoneme condition and prosody condition from the accumulated voice waveform segment files.

[0004]

However, if the speech waveform pieces obtained as described above are simply connected to create a synthesized sound, discontinuity occurs in the speech waveform at the connection point, which results in a click. Cause sound. In order to solve this problem, the cut-out start position of the speech waveform is set near the phonological boundary and immediately after the sound waveform data exceeds zero, and the cut-out end position is set near the phonological boundary and the sound waveform data is A method of just before crossing zero can be considered. In this case, the discontinuity of the voice waveform itself can be avoided, but this causes the discontinuity in the pitch structure, which in turn causes noise. In order to connect while maintaining the pitch structure, it is necessary to preset the cutout position for each pitch in the voice waveform data. However, even in this case, when different voice waveform data are simply connected, the continuity of the spectrum is not guaranteed and noise is often generated.

According to the present invention, interpolation processing is performed between voice waveforms to be connected during voice waveform connection, and noise is generated by preventing large waveform discontinuities and spectral gaps from occurring at the connection points of voice waveform segments. It is intended to provide a method and a device for synthesizing a voice that are prevented in advance.

[0006]

In a voice synthesizing method in which a plurality of voice waveforms are connected to obtain a synthesized voice, a waveform at the end of a preceding voice waveform and a waveform at a leading end of a subsequent voice waveform are set to have twice the pitch length. An interpolating waveform is generated by cutting out with a window function having a length, weighting the cut out waveforms, and adding the center positions of the window functions together, and then interpolating the preceding and following speech waveforms. We constructed a voice synthesis method to connect via.

Then, in the above-mentioned voice synthesizing method, a voice synthesizing method is constructed in which a plurality of interpolated waveforms are generated and the interpolated waveforms are added while being shifted by a pitch interval. Further, in a voice synthesizing apparatus for connecting a plurality of voice waveforms to obtain a synthesized voice, a waveform at the terminal end of the preceding voice waveform and a waveform at the front end of the following voice waveform are converted by a window function having a length twice the pitch length. The speech waveform buffers 4 and 5 are provided with a waveform clipping section for clipping, and an interpolation waveform generating section 8 for adding the central positions of the window functions while adding weights to these clipped waveforms is provided. A voice synthesizer having a waveform data connection unit 9 for connecting a subsequent voice waveform via an interpolating waveform is also configured.

[0008]

Embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a flow chart of a voice synthesizing process according to the present invention, FIG. 2 is a block diagram of a voice synthesizing apparatus according to the present invention, and FIG. 3 is a diagram showing connected voice waveform units and generated interpolation waveforms.

The outline of the present invention will now be described with particular reference to FIG. (A) shows the preceding speech waveform segment,
(E) shows a subsequent speech waveform segment to be connected to the preceding speech waveform segment. (B) shows the pitch waveform at the end of the preceding speech waveform segment cut out by the window function,
(D) shows the leading pitch waveform of the following speech waveform segment cut out by the window function.

(C) shows an interpolated waveform generated by adding the waveforms of (B) and (D) while weighting them. This is a case where the number of interpolation pitch waveforms is m = 2, and two interpolation waveforms are generated. (F) shows a speech waveform that is connected and synthesized. The connected speech waveforms are the preceding speech waveform segment shown in (A), the two interpolated waveforms shown in (C), and the subsequent speech waveform segment shown in (E) in FIG. It is formed by moving in parallel vertically downward and superimposing them.

The embodiments of the present invention will be described in more detail below. In FIG. 2, reference numeral 1 is an input terminal to which a voice waveform segment to be connected is input in a number format. Reference numeral 2 is a control unit, and 3 is a voice waveform segment file. In the voice waveform segment file 3, voice waveform segments and numbers are associated with each other. Reference numerals 4 and 5 are voice waveform buffers, both of which temporarily store the voice waveform elements read from the voice waveform element file and cut out the voice waveform elements. 6 and 7 are pitch waveform buffers. Reference numeral 8 is an interpolation waveform generation unit, 9 is a waveform data connection unit, and 10 is an output terminal. The operation of the speech synthesizer of the present invention will be described below.

First, the number sequence of speech waveform segments to be connected is input from the input terminal 1. Then, the control unit 2 analyzes these input numbers, sequentially reads the corresponding speech waveform segment from the speech waveform segment file 3, and stores them in the speech waveform buffer 4 and the speech waveform buffer 5. . The preceding speech waveform segment among the speech waveform segments connected to each other is accumulated in the speech waveform buffer 4, and the succeeding speech waveform segment is accumulated in the speech waveform buffer 5. Then, pitch marks are provided in advance to each of the voice waveform segments. The pitch mark is passed through the low-pass filter having a cut-off frequency of not less than the fundamental frequency, that is, about 256 Hz, and the original voice waveform is passed, and the pitch mark is given to each local peak position of the waveform after passing through the low-pass filter of the original voice waveform. To do. The pitch mark pitches thus imparted satisfactorily represent the pitch pitches of the speech waveform segments. Further, the cut-out start position and end position of the voice waveform segment are determined in synchronization with the pitch mark.

Next, in the voice waveform buffer 4, the waveform data is cut out by a window function having a window length twice the pitch length, centering on the last pitch mark of the preceding voice waveform segment accumulated here. , And sends it to the pitch waveform buffer 6. Similarly, in the voice waveform buffer 5, a pitch waveform is cut out around the head pitch mark of the subsequent voice waveform segment accumulated here, and this is sent to the pitch waveform buffer 7. The window function may have various shapes, but a cosin wave shape, for example, a Hanning window is considered effective in maintaining the original waveform. 3 (B) and 3 (D)
Shows a window function having a window length twice the pitch length and waveform data cut out by the window function.

In order to generate the interpolated waveform, the interpolated waveform generator 8 inputs the pitch waveform data from the pitch waveform buffer 6 and the pitch waveform buffer 7, and performs addition processing while weighting these data to form the interpolated waveform. To generate. In this interpolation waveform generation process, the last pitch waveform of the preceding speech waveform segment cut out by the window function is x (i), the beginning pitch waveform of the succeeding speech waveform segment is y (j), and the generated pitch is When the waveform is z (k), it is represented by z (k) = w _m × _x (i) + (1−w _m ) × y (j) (1). Here, w _m is a weighting function, and if the number of interpolation pitch waveforms to be generated is m, the n-th weighting function is represented by w _m = n / (m + 1) (2). In the equation (1), the centers of the respective pitch waveforms are made to coincide with each other and added. The number m of interpolation pitch waveforms to be generated can be set to 1.

The waveform data connection unit 9 is supplied with the interpolated waveform generated as described above by the interpolated waveform generation unit 8 and the preceding speech waveform segment data which is the original waveform for generating the interpolated waveform. Is input from the voice waveform buffer 4, and the following voice waveform segment data is input from the voice waveform buffer 5. Then, the preceding speech waveform and the succeeding speech waveform are connected via an interpolating waveform to generate synthetic speech.

[0016]

As described above, according to the present invention, when two speech waveform segments having different characteristics are connected to each other, the waveforms at the end of the preceding speech waveform segment and An interpolating waveform is generated by cutting out the waveform at the tip of the following speech waveform segment and adding the both waveforms that have been weighted while weighting them, and connecting the preceding speech waveform and the succeeding speech waveform via the interpolation waveform. To do. Therefore, a large waveform discontinuity does not occur at the connection point of the speech waveform segment, and since the waveform is connected so that the spectrum changes smoothly, the generation of noise is extremely small.

[Brief description of drawings]

FIG. 1 is a diagram showing a processing flow of the present invention.

FIG. 2 is a block diagram of the present invention.

FIG. 3 is a diagram showing connected voice waveforms and generated interpolated pitch waveforms.

[Explanation of symbols]

 1 Input Terminal 2 Control Section 3 Voice Waveform File 4 Voice Waveform Buffer 5 Voice Waveform Buffer 6 Pitch Waveform Buffer 7 Pitch Waveform Buffer 8 Interpolation Waveform Generation Section 9 Waveform Data Connection Section 10 Output Terminal

Claims (3)

[Claims] 1. A voice synthesizing method for connecting a plurality of voice waveforms to obtain a synthesized voice, wherein a waveform at the end of a preceding voice waveform and a waveform at the front end of a following voice waveform have a pitch length of 2
It is cut out by a window function having a doubled length, and the extracted waveforms are weighted while the center positions of the window functions are aligned and added to generate an interpolated waveform, and a preceding speech waveform and a succeeding speech waveform are generated. A voice synthesizing method characterized by connecting through an interpolated waveform. 2. The speech synthesis method according to claim 1, wherein a plurality of interpolated waveforms are generated and the interpolated waveforms are added while being shifted by a pitch interval. 3. A voice synthesizer for connecting a plurality of voice waveforms to obtain a synthesized voice, wherein the end waveform of the preceding voice waveform and the leading waveform of the following voice waveform have a pitch length of 2
A waveform cutout unit that cuts out by a window function having a double length is provided, and an interpolation waveform generation unit that adds the center positions of the window functions while weighting the cutout waveforms and adds the preceding waveform A voice synthesizing apparatus comprising: a waveform connecting section for connecting a subsequent voice waveform via an interpolating waveform.