JPS5948399B2 - Speech synthesis method - Google Patents

Description

[Detailed Description of the Invention] The present invention relates to a speech synthesis method using a segment editing method, and an object of the present invention is to reduce the number of data required for a segment without impairing the phonological quality of the synthesized sound. It is something to do.

In the conventional segment editing method, a required speech segment is quantized into multiple bits (usually 6 to 12 bits), and then recorded and reproduced according to a certain synthesis method.

However, since fricative consonants in natural speech are generally considered to be noise that has passed through a filter with certain characteristics, their waveforms do not have periodicity. When attempting to synthesize the same waveforms, oscillations were generated due to the repetition of the same waveform, which degraded the sound quality of the synthesized sound. Long enough to suppress the above oscillation noise (
It is conceivable to use an elemental piece (length of about 30 msec or more), but this requires a large amount of digital data, which has the drawback of increasing costs and equipment structure. The present invention eliminates the above problems in fricative consonant synthesis. First, an example of a speech waveform synthesized by the speech synthesis method of the present invention will be shown. Figure 1A is “Hachi (
This is a natural speech waveform of a male voice when pronouncing "8)".

The waveform portions indicated by 1 and 2 in the figure are respectively and [tf
]. Figure B shows a synthesized waveform obtained by the elemental piece synthesis method of this natural waveform. The parts indicated by 3 and 4 in the figure are
Extract a part from the child waveform part shown in 1 and 2 in the above figure,
This process was repeated twice and connected. On the other hand, in Figure C, the child waveforms 3 and 4 in Figure B are converted into zero-crossing waveforms and applied to the corresponding portions as waveforms 5 and 6, respectively. Here, a zero-crossing wave is a waveform in which the absolute value of the amplitude is always constant and vibrates underground with ±o as the boundary, so to speak, it is a waveform in which only the phase information of the waveform remains and the amplitude information is missing. . FIG. 1D is an enlarged view of a part of the natural waveform of the consonant [tf] and its zero-crossing waveform. Next, one embodiment of the present invention will be described with reference to FIG.

In Figure 2, 1 is a conversion table for bit patterns given externally, 2 is a table for reading prime group start addresses for word construction, 3 is a prime group table, and 4 is a 1-bit output for zero-crossing wave or normal 5 is the zero-crossing wave segment file, 6 is the normal multi-bit data segment file, T is D-
A converter, 8 is data for output level control, and 9 is an attenuator for output level control. FIG. 2 is a block diagram for appropriately selecting and outputting a zero-crossing wave and a normal elemental waveform. In Fig. 2, 5 is a memory for the waveform shape of a zero-crossing wave, for example, 1 for a positive signal.
, φ is written for the negative signal. For example, when using a memory with 8 bits per word, the writing order starts from the LSB and MSB for each word.
It should end in B. In the case of fricative consonants, the segment length can generally be selected regardless of the phonetic pitch, so it is convenient when reading out data for each segment, starting from the L-cut of one memory word and ending with the MSB of another memory word. It is. 6 is a memory of elemental shapes other than fricative consonants, 1
The shape of a segment is expressed using all bits of one word per sample point.

Since the above 5 and e are memories regarding the shape of the elemental piece,
Hereafter, it will be called a segment file. 3 is a memory in which, in addition to the start address, stop address, and number of repetitions of the fragments for reading the fragment file mentioned above, the following data 4 and 8 are written, which consists of several repetitions of the same fragment. It defines waveforms (called prime groups) and is called a prime group table.

4 is data in this prime group table, which indicates whether or not the specified elementary piece is a fricative consonant using l bits.

Also, in the figure, 7 is a D-A converter, 9 is an attenuator that can be controlled digitally, and data 8 in the elementary group table indicates the reproduction level of the elementary piece, which controls the attenuator 9. . To synthesize a certain word, it is necessary to connect and output multiple prime groups mentioned above, but the start addresses of various prime group tables are written in succession in memory 2, and this By sequentially reading out the memory 2, the prime groups necessary for synthesizing the target word speech can be output one after another. Memory 1 is a memory for converting an externally applied bit pattern into a start address of memory 2. On the other hand, the program necessary for speech synthesis is shown in FIG.

In step 10, data 4 in the prime group table mentioned above is referenced to specify the branch direction of the program.

When the output segment is a fricative consonant, read one word in the segment file and output it one bit at a time starting from the LSB, and when all bits in the l word have been output, address the next word. It's getting old. On the other hand, if the output segment is not a fricative consonant, it is sufficient to output one entire word. As described above, by operating the function block in Figure 2 with the program in Figure 3, you can output zero-crossing waves for fricative consonants and normal segment data for non-fricative consonants to edit the segments. It is possible to obtain a synthesized speech signal using the method.

The speech synthesis method of the present invention has the above-mentioned configuration, and the zero-crossing wave is applied to the fragments for fricative consonants, which require a longer fragment length compared to the fragments for vowels, etc. It is possible to reduce the data, and on the other hand, since normal multi-bit word data is used for the parts other than the fricative consonant part, it has the advantage that synthesized speech with less phonological defects can be obtained.

[Brief explanation of drawings]

Figure 1A is a natural speech waveform diagram, Figure 1B is a synthetic speech waveform diagram, Figure 1C is a synthetic speech waveform diagram using zero-crossing waves in the fricative consonant part, and Figure 1D is a diagram of the fricative consonant part (Tf ) and its zero-crossing waveform diagram, FIG. 2 is a block diagram of an apparatus implementing the speech synthesis method of the present invention, and FIG. 3 is a program flowchart diagram of the same apparatus. 1... Conversion table, 2... Table for reading prime group start address, 3... Prime group table, 4... Bit data for switching, 5
・・・・・・Element file for zero crossing wave, 6・・・・・・
Multiple bit data segment file, T...D-
A converter, 8...Output level control data, 9...Attenuator.

Claims (1)

[Claims] 1. For fragments of fricative consonants or similar noise waveforms, use the zero-cross or wave of the corresponding fricative consonants or similar noise waveforms, while for fragments other than the above fricative consonants or similar noise waveforms. On the other hand, there is a speech synthesis method characterized by using digital data segments quantized with multiple bits.