JPH04277B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech synthesis method for synthesizing arbitrary continuous speech based on speech elements smaller than words.

<Prior art> Methods for synthesizing speech of arbitrary words and sentences have traditionally been based on phonemes (a, i,
p, k, s, etc.)), and a combination of two consecutive phonemes (vowel-consonant,
Dyad, VCV which is vowel-vowel, consonant-consonant)
A method of synthesizing from complex elements of phonemes such as (vowel-consonant-vowel) is known. The method of synthesizing from phonemes requires ``rules'' to realize the movement part from one phoneme to another phoneme, the transformation associated with articulatory combination, etc., but at present, all of these knowledge has not been clarified. However, it does not necessarily provide a good quality synthesized sound. The method using complex elements of phonemes is a relatively realistic method, and there are examples in which it has been put to practical use. In the conventional method using VCV and CV as elements, the combination of synthesis elements is mainly performed at the vowel part. The vowel part is the part that is most audible, and this has the disadvantage that combined distortion of spectral parameters and changes in timbre occur, impairing the quality of the synthesized speech.

<Purpose of the invention> This invention aims to improve the quality of synthesized speech.
This invention provides a speech synthesis method using new elements that eliminates the drawbacks of conventional composite elements.

<Principle of the Invention> As already mentioned, the combination of synthesis elements in the vowel stationary region may have a negative effect on the quality of the synthesized speech. Therefore, a CVC (C: consonant, V: vowel) type synthesis element is set. In this element, the vowel part is influenced by the preceding and following consonants. In other words, it is an element that naturally includes the influence of articulatory combination and can realize smooth mouth movements. In addition, in order to give a sense of rhythm to the synthesized sound, the duration of the vowel must be set appropriately, but in the CVC type element, realistic vowel lengths are achieved due to the influence of the consonants on both sides.

Pronunciation is made as a single syllable (vowel-plosive),
It is convenient to include (a sequence of vowels and vowels i) in one element. This is because these phoneme sequences are not only extremely tight, but also occur frequently in Japanese.

As described above, in the present invention, CVC and CVMC (M is a phonic sound N or a vowel i) are set as the basic formats of speech synthesis elements. However, considering the types of phonemes in Japanese, there are over 6,000 types of speech synthesis elements in this basic format, which exceeds the realistic number of speech synthesis elements. In reality, among these speech synthesis elements, there are probably some phoneme sequences that are rarely used in Japanese, and even if they are used, there are definitely some that appear very rarely. Therefore, in this invention, based on the frequency of appearance of Japanese phonemes, the most frequently used ones are
By using CVC or CVMC and compositing using elements such as CV, VC, VV, VN, and NC for infrequently used items, the storage memory of composite elements can be reduced and the amount of work for creating composite elements can be reduced. Reduce it to a large width. Figure 1 shows, as an example, the cumulative percentages of CVC and CVMC obtained from the analysis of Japanese dictionary headwords.
It is understood that 90% of Japanese words are covered by 1000 CVC and CVMC phonological sequences. Therefore, 1000 long unit elements like this and CV, VC,
Synthesis is performed using short unit elements such as VV, VN, and CN.

<Example> FIG. 2 shows an example of the present invention. When a character sequence of a word or sentence to be synthesized is input from the input terminal 1, the composition element name sequence creation unit 2 immediately divides the input character sequence into a series of long unit elements of CVC or CMC, and calculates the length of each of them. The element name of the unit element is sent to the composite element name collation unit 3. The synthesis element name collation unit 3 reads the name of the speech synthesis element from the long unit synthesis element name storage memory 4, collates this with the long unit element name transferred from the synthesis element name series creation unit 2, and stores the name in the memory. Determine whether or not there is a name of the speech synthesis element in 4.
The result is transferred to the composite element name series creation section 2.
If there is no corresponding long unit element in the long unit composite element name storage memory 4, the composite element name series creation section converts the missing long unit element CVC into CV, VC, and CVMC.
Divide into a series of short unit elements of CV, VM, and MC. For example, sakurawa kireidesu (cherry blossoms are beautiful) is sak+kur+raw+wa kir+re+ei+id+des+
The underlined part is an example of dividing into three short unit elements because there was no speech synthesis element for the long unit element [reid]. In addition, at the end of the word, [wa],
Short unit elements are used, such as [su].

When a synthetic element name sequence is obtained from the synthetic element name sequence creation section 2 as described above, the sequence is sent to the speech synthesis control section 5. For various CVC and CVMC, long unit composite element parameter storage memory 6
For all CVs, VCs, etc., voice synthesis spectrum parameters are stored in the short unit synthesis element parameter storage memory 7. The speech synthesis control section 5 controls the parameter reading section 8 to read the parameters for each speech synthesis element of the input synthesis element name series from either the memory 6 or 7, and sends the information to the parameter smoothing section 9.

The parameter smoothing section 9 smoothes the parameters by performing the following weighted averaging process on spectral parameters (PARCOR coefficients, LSP parameters, formants, etc.) and amplitudes at the speech synthesis element connection portion and its vicinity.

f= _f+l 〓 ^i+fl W _if Pf However, _l 〓 ^j=-l W _j = 1 Pf: Parameter Wi: Weighting coefficient f: Frame number 2l+1: Frame length of the window to perform weighted average processing This processing is performed as an audio This is done over several frames before and after the joint of composite elements.

FIG. 3 shows an embodiment of the parameter smoothing unit 9, in which the parameter storage memory 6,
The parameters of the speech synthesis element read from 7 are stored in the input buffer memory 21.
The input buffer memory 21 is a double buffer, and the parameters of the next speech synthesis element are also stored at the same time. The arithmetic control section 22 is a section that controls the smoothing operation of parameters, and transfers the parameters to the output buffer register 24 for each frame via the readout control section 23 in the composite element section where smoothing processing is unnecessary. In the parameter smoothing section near the joint of the speech synthesis elements, the parameters of the window section in which the weighted average processing is performed are sequentially read out from the input buffer memory 21, and the multiplier 25 multiplies them with the weighting coefficients stored in the memory 26. At the same time,
A product-sum operation is performed by the adder 27 and the accumulation register 28, and the smoothed parameter values of the window-centered frame are sent to the output buffer register 24. When the transfer of the parameters of all frames of one synthesis element to the output buffer register 24 is completed, the parameters of the next speech synthesis element are written to the input buffer memory 21, and the same operation as described above is repeated. Note that the envelope of the weighting coefficient within the window section is maximum at the center, such as a triangular waveform or a Gaussian waveform.

FIG. 4A shows an example of a time series of LSP parameters before such processing, and FIG. 4B shows an example of the time series of LSP parameters subjected to the averaging process. For example, by averaging the parameters,
The joining part 31 between and "gu" and the joining part 32 between "ru" and "ma" become joining parts 33 and 34, respectively, which give smoothness to the synthesized sound and eliminate differences caused by sudden changes in parameter values at the joining parts. Mixing of sound can be avoided.

Returning to the explanation of FIG. 2, when the parameter smoothing section 9 generates the parameter time series of the speech synthesis elements, the speech synthesis control section 5 supplies the drive sound source signal from the sound source generation section 10 to the speech synthesis digital filter 11, The filter characteristics of the digital filter 11 are controlled by the parameter time series from the parameter smoothing section 9, and a final synthesized speech signal is output to the synthesized speech output terminal 12. The sound source generation section 10 and the speech synthesis digital filter 11 constitute a speech synthesizer.

<Effects> As explained above, this invention has CVC and
In order to use CVMC as a basic element of speech synthesis,
Linear combinations of parameters are avoided in the vowel part, where the audibility is louder, the effects of articulatory combinations are naturally included, smoother mouth movements are achieved, and not only a synthesized sound of better quality than before can be obtained, but also CV ,VC,
The combined use of VV and VN type synthesis elements has the advantage of reducing the total number of speech synthesis elements and making it possible to realize the device economically. In addition, when smoothing the parameters when combining the speech synthesis elements, a smooth synthesized sound can be obtained, and it is also possible to prevent abnormal sounds from being mixed in due to "jumps" in the parameters at the synthesis element combination section.

[Brief explanation of drawings]

Fig. 1 is a diagram showing the cumulative percentage of appearance of CVC and CVMC in Japanese words, Fig. 2 is a block diagram showing an embodiment of the present invention, and Fig. 3 is a specific example of the parameter smoothing unit 9 in Fig. 2. FIG. 4 is a diagram showing the difference between the case where parameters are averaged and the case where they are not averaged. 1: Character sequence input terminal, 2: Synthesis element name series creation section, 3: Synthesis element name collation section, 4: Long unit synthesis element name storage memory, 5: Speech synthesis control section, 6: Long unit synthesis element parameter storage memory , 7: Short unit synthesis element parameter storage memory, 8: Parameter reading section, 9: Parameter smoothing section, 10: Sound source generation section, 11: Speech synthesis digital filter, 12:
Synthetic sound output terminal, 21: Input buffer memory, 2
2: Arithmetic control section, 23: Read control section, 24: Output buffer register, 25: Multiplier, 26: Memory, 27: Adder, 28: Accumulation register.

Claims (1)

[Claims] 1. A first memory that stores names of speech synthesis elements in the form of vowels, vowel sequences, or vowel-phrasal sequences sandwiched between consonants on both sides; and a second memory that stores spectral parameters of each of the speech synthesis elements.
a third memory for storing spectral parameters of a speech synthesis element in a two-phoneme continuous format; and a speech synthesizer; If the spectral parameters are stored, the spectral parameters of the speech synthesis element are read from the second memory, and if the spectral parameters are not stored in the first memory, the corresponding spectral parameters are read from the third memory; 1. A speech synthesis method, comprising: means for mutually combining the obtained spectral parameters to create a continuous spectral parameter time series, and supplying this to the speech synthesizer to synthesize continuous speech.