JPS59148094A - Voice synthesizer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to speech synthesis, and is particularly applicable to a rule synthesis device for the purpose of synthesizing arbitrary words and sentences.

Conventional configurations and their problems Research on speech synthesis has been actively conducted for several years. Analysis-synthesis is often used to analyze speech, extract feature parameters, and synthesize speech from those parameters.However, analysis-synthesis reduces the amount of speech that can be synthesized, and compresses the amount of information below a certain amount. However, there was a problem in that the sound quality deteriorated significantly and there was a limit to the amount of information compression.

Therefore, research has begun on rule synthesis, which can synthesize arbitrary speech with a small amount of information. Table 1 summarizes a comparison of three methods of rule synthesis based on the unit of storing audio data.

Method (1) is a method of storing voice data in cv units and combining them at the vowel part. For example, to synthesize KOTOBA (word), Noni "KO", 'OTO',
'OBA' No. 8, Table 1 The VCV fragments (hereinafter, basic units such as ■Cv are referred to as fragments) are combined and synthesized.

In this example, the initial CV (KO) + VCV (OTO
) + VCV (OBA). In this method, all the connections between elementary fragments are done vowel-vowel, so the combination rules are relatively simple, but there is a problem that a large number of elementary fragments are required, and the number of such fragments is about 900. .

Method (2) uses cv (one syllable) as the memorization level; for example, KOTOBA (word) is KO'.

It is synthesized by combining 8 CV fragments of 'To' and 'l(A'. In this method, the number of required fragments is approximately 100.
On the other hand, it requires a method (less than the old one, but vowel to consonant <V-C), making the combination rules a little more complicated.

Method (3) uses C and V (phonemes) as memory units, and KOTOBA < N leaves) is written in %/, 0/
, Tt.

The six phonemes %Q/, %T3/, and %A/ are combined and synthesized. The number of elements required for this method is about 20, which is much smaller than methods (1) and (2). Since the combination (C-V) also requires one key, the combination rules become very complicated.

OBJECTS OF THE INVENTION It is an object of the present invention to solve the above-mentioned problems and to provide a speech synthesis device that enables rule synthesis using a small number of fragments and simple combination rules.

In consideration of the fact that the connection rules become more complex as the units of the constituent elements of the invention become smaller, this is because the connection rules differ depending on the type and property of the phoneme itself, and the types and properties of the adjacent phonemes. The result.

The present invention has been made based on the discovery that, depending on the phoneme, even if the combination rules are limited, the intelligibility does not deteriorate significantly.
a means for inputting a character string, a character string analysis means for selecting a segment necessary for synthesis from the input character string, a storage means for storing the segment, and a storage unit for storing the selected segment. Simple reading from the beginning.

an editing buffer for temporarily storing the read segment; a means for estimating utterance duration, prosody, and amplitude for manually inputting the editing buffer signal and storing the result in the editing buffer; Means for combining segment sequences according to a combination rule and means for synthesizing speech waveforms are provided, and the units stored in the segment storage unit include the type and property of the phoneme itself, and the type of adjacent phonemes.

Depending on the nature of the phoneme, the structure is configured to select and decide whether to include only a single phoneme or to include some or all of the adjacent phonemes, and for phonemes that can have limited combination rules, a single phoneme is used. By using a phoneme as a unit, and for other phonemes, including some or all of the adjacent phonemes, it is possible to synthesize rules using simple combination rules and a small number of units.

DESCRIPTION OF THE EMBODIMENTS An embodiment of the regular speech synthesis apparatus according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of this embodiment. The character string to be synthesized is input in Roman letters from the input section (]), and the segment necessary for synthesis is selected in the string analysis section (2). Next, a necessary segment is read out from the segment storage section (3) by the segment reading section (4) and temporarily stored in the editing buffer (5). Using the signal from the editing buffer (5) as input, the time length determination unit (6) determines the time length for each segment by taking into account the influence of adjacent phonemes, and the result is sent to the editing buffer (5).
Store in. Similarly, as a signal or input to the editing buffer (5), the prosody adding section (7) determines the fundamental frequency, and the amplitude adding section (8) takes into account the position and type of each element and indicates the intensity of the synthesized sound. Determine the amplitude and store the results in hli collection buffer 1 (5). Finally, in a combining section (9), the segments are combined according to the connection rules for each segment, and in a synthesizing section θQ, synthesized speech is obtained at an output terminal 01).

The synthesis method used in the synthesis section 01 is a formant vocoder method, and the parameters required for synthesis are the first formant,
Second formant, third formant (hereinafter Fl, F2.F
8 + (z), bandwidth of each formant (hereinafter referred to as B1, B2, B3, unit Hz), amplitude of the famous formant (hereinafter referred to as A1, A2, A3, unit DB), amplitude indicating intensity, fundamental frequency There are 11 parameters, and the parameters are updated every 5 msec. Among these, the amplitude and fundamental frequency indicating the strength are determined by the prosody imparting section (7) and the amplitude imparting section (8).
), the parameters that the segment stored in the segment storage unit (3) should have are F1, F2, F3, A1
, A2, A3, B1, B2, and B3, nine in total.

Next, the segment storage unit (3) will be explained in detail.

The segment units stored in the segment storage unit (3) are complex because the C-V combination rules are different for each subsequent vowel (since the synthesized speech is Japanese, the subsequent phonemes are limited to vowels). 'g'
, 'd', 't゜, and 'k', the CV units 'ga' and 'do', which include a part of the subsequent phoneme, are used as an elemental piece. For phonemes other than 'g', 'd', 't', and 'k', the C-V combination rules can be limited to several types, so only a single phoneme is used as a segment. That is, the fragments stored in the fragment storage unit (3) are 'ga', 'gi',
'gu', 'ge', 'go', 'da', 'de',
'do', 'ta', 'ti', 'tu', 'te',
'to', 'ka', 'ki', 'ku', 'ke',
18CV fragments of 'ko', 'a', 'i', 'u'
,'e','o','s7,'z','ch7,'n'
, 'm', 'p', 'b', 'r', 'y', 14 phonemes (C, V) in total, 32 phonemes. The consonant + 'y' +
For vowels, or 'g', 'd', 't', and 'k', for example, it is composed of 'gi' + vowel and time length adjustment,
'h' is composed of amplitude information.

The ``n'' is used for the ``n'' sound, and the consonant is constructed by adjusting the duration.

Figure 2 shows an example of an elemental piece used in this device.
' and 'KO' parameters are shown. F1, F2, and F3 are formant frequencies, B1, B2, and B3 are bandwidths in Hz, and A1, A2, and A3 are amplitudes in DB.

The synthesis unit (10) updates the input parameters every 5 msec, so when a CV is used as a segment unit, such as 'KO', a parameter sequence every 5 msec is provided as a segment. 12 in KO12 indicates the length of the parameter string that the elemental piece 'KO' has.

For example, the character string 'KOTO' input from input section (1)
BA' is analyzed by the character string analysis unit (2), and 'KO', 'T', 'B', and 'A' are selected as the fragments necessary for synthesis. Then, the segment reading unit (4) reads the parameters of the segment 'KO', 'TO', 'B', and 'A' from the segment storage unit (3) and stores them in the editing buffer (5).

Next, the coupling rules used in the coupling part (9) will be explained.

The following two combination rules are used here.

1. If the consonants are 'p' and 'b', no parameters are interpolated.

2. Otherwise, perform linear interpolation.

Figures 3(a) and 3(b) show combination rule 1 when the consonants are 'p' and 'b'. and the combination rules for other cases. (1
) indicates the parameter value of the preceding phoneme, (2) indicates the parameter value of the subsequent phoneme, and t1 and t2 indicate the duration length of each phoneme.

For formant vocoder, 'd', 'c', 't'
, 'k', and it has been experimentally confirmed that the clarity does not deteriorate even if the combination rules of C-C and V-C other than 'k' are limited to the above two. In this way, by using CV, C, and V units of elemental fragments as appropriate depending on the nature of the consonant, it is possible to achieve highly clear layer and rule synthesis with only 32 elemental fragments and two simple rules. Can be done.

Effects of the Invention According to the present invention, by appropriately selecting the storage unit of the phoneme according to the characteristics of the phoneme, it is possible to perform rule synthesis using a small number of phonemes and a simple combination rule.

In this example, the synthesis method is a formant vocoder,
Although CV, C, and V are used as storage units, the present invention is also applicable to other composition methods and storage units.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a speech synthesis device showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of a segment used in the present invention, and FIG. 3 is a diagram showing segment combination rules. . (1)...Input section, (2)...Character string analysis section, (3)
...Segment storage section, (4)...Segment reading section, (5)
... Edit buffer, (6) ... Time length determination section, (7
)... Prosody imparting section, (8)... Amplitude imparting section, (9)
...Joining part, (10)...Composition part, (11)...
Output agent Yoshihiro Moriki Figure 1 Figure 2 A63310532491 / 60 60 5Q /
50 70 110PI F2 F3
AI A2 A3 BI B2 83Figure 3 (&N U shell “If (P, j)) 558-

Claims (1)

[Claims] 1. A means for inputting a character string, a character string analysis means for selecting a segment necessary for synthesis from the input character string, a storage means for storing the segment, and a method for storing the selected segment. an editing buffer that temporarily stores the read out segment; an editing buffer that temporarily stores the read segment; and an editing buffer that receives the edited bassoma signal and stores the result in the editing buffer. The units to be stored in the segment storage unit are based on the type and property of the phoneme itself. . A speech synthesis device that selects and determines whether to include only a single phoneme or to include some or all of the adjacent phonemes, depending on the type and property of the adjacent phonemes.