JP2000214876A - Japanese speech synthesis method - Google Patents

Abstract

(57) [Problem] To provide a Japanese language that can determine the phoneme duration of each phoneme in a shorter time than in the past, and can prevent a local statistical error from increasing. An object of the present invention is to provide a speech synthesis method. SOLUTION: The Japanese speech synthesis method according to the present invention comprises:
Speech waveform segments corresponding to Japanese phoneme symbols are prepared in advance, and speech waveform segments corresponding to the input phoneme symbol sequence are connected in the order of the input phoneme symbol sequence, and waveform shaping is performed according to the prosody regulation control, In a Japanese speech synthesis method for synthesizing a speech corresponding to an input phoneme symbol string, category classification is performed on all phoneme symbols in advance, and the phoneme duration for each input phoneme in the input phoneme symbol string is determined by the input phoneme. The decision is made based on the category of the input phoneme included in the symbol string.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Japanese speech synthesis method for converting input text information into speech information having a natural prosody.

[0002]

2. Description of the Related Art As a method of controlling a phoneme duration used in a Japanese speech synthesis method, a prediction model is analyzed by performing a quantification class I analysis using parameters considered to affect the phoneme duration. (See Patent Publication No. 2749803 and Patent Publication No. 2749804).

[0003]

However, the above conventional method has the following disadvantages.

(1) Since the phoneme duration of each phoneme is determined in consideration of all parameters that are considered to affect the phoneme duration, it is a long analysis for statistical analysis such as quantification type I analysis. Takes time.

(2) Since the phoneme duration of each phoneme is determined in consideration of all parameters that are considered to affect the phoneme duration, local statistical errors may occur due to inconsistent parameters. May be larger,
Obviously, there may be abnormalities in phonological duration control. If the statistical error is large, deleting the data and re-performing the statistical processing is known as a general method.However, if the effects of the parameters are inconsistent under specific conditions, accurate statistical results can be obtained. I can't.

The present invention provides a Japanese speech synthesis method capable of determining the phoneme duration of each phoneme in a shorter time than in the past and preventing a local statistical error from increasing. The purpose is to:

[0007]

In the method of synthesizing Japanese speech according to the present invention, a speech waveform segment corresponding to a Japanese phoneme symbol is prepared in advance, and the input phoneme symbol sequence is added to the input phoneme symbol sequence in the order of the input phoneme symbol sequence. In the Japanese speech synthesis method of connecting the corresponding speech waveform segments, shaping the waveform in accordance with the prosody regulation control, and synthesizing the speech corresponding to the input phoneme symbol string, category classification is performed in advance for all phoneme symbols. In addition, a phoneme duration for each input phoneme in the input phoneme symbol string is determined based on the category of the input phoneme included in the input phoneme symbol string.

The categories of phoneme symbols are, for example,
(1) vowels and vowels, (2) long vowels, (3) unvoiced plosives, (4) affricates, (5) unvoiced fricatives, (6) semi-vowels and flowing sounds, (7) nasal consonants, (8) Voiced plosives are classified into (9) other independent phonemes.

The phoneme duration for each input phoneme is determined by the category of the phoneme of interest, the category of the phoneme two before the phoneme of interest, the category of the phoneme one before the phoneme of interest, and the category of the phoneme one after the phoneme of interest. It is preferable to make the determination based on the category of the phoneme two steps after the phoneme of interest.

In addition, it is preferable that the phoneme duration for the target phoneme is determined in consideration of the number of moras in the exhalation paragraph including the target phoneme and the mora position of the target phoneme.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention relates to a Japanese speech synthesis method for converting input text information into speech information having a natural prosody, and is characterized by a method for determining a phoneme duration. It is implemented in the form of application software of a personal computer or middleware in an embedded device.

FIG. 1 shows the processing procedure of the speech synthesis method.

The input Japanese kana / kanji mixed text information is converted into a phoneme symbol string by language processing (step 1). Next, a speech waveform segment corresponding to the phoneme symbol string is extracted from the speech database (step 2).

The extracted speech waveform segments are subjected to waveform shaping by prosody rule control (step 3) and output as speech information (step 4). The phoneme duration control is executed in step 3.

FIG. 2 shows a processing procedure of the phoneme duration control executed in step 3.

Here, the input text is "Taro-san, good morning", and the corresponding phoneme symbol string is "tar".
The following describes an example in which o ＾ saN% ohayo ”. In the phoneme symbol string, the pronunciation is represented by Roman alphabet reading. However, the sound-repelling“ n ”is represented by“ N ”. Vowels are denoted by ＾ immediately after aiueo.
The pose is represented by “%”.

Table 1 shows the notation and category classification of each phoneme. For simplicity of explanation, control symbols for accents are omitted.

[0018]

[Table 1]

In the phoneme duration control, the phoneme duration of each phoneme included in the input phoneme symbol string is determined.

First, the process branches based on the target phoneme and the phonemes before and after the target phoneme (step 31).

The following seven branches are provided. (1) When the target phoneme is a prompt sound: proceed to step 311 (2) When the target phoneme is a pause (silent section): proceed to step 312 (3) When the phoneme after the target phoneme is a prompt sound: step 313 (4) When the phoneme after the target phoneme is a pause: go to step 314 (5) When the phoneme immediately before the target phoneme is a prompt sound: go to step 315 (6) One of the target phonemes If the previous phoneme is a pause: go to step 316 (7) If none of the above (1) to (6): go to step 317

Primary search tables are created in advance for each of the above seven types. Table 2 shows part of the contents of the primary search table used when the phoneme after the target phoneme is a pause. The average time (= average phoneme duration time) average time for the pixel of interest in each table is obtained mainly by averaging a plurality of actual utterance data (same speaker) by performing statistical processing.

[0023]

[Table 2]

The target phoneme is “taro @ saN% oha”
In the case of "N" in "yo", the process proceeds to step 314 because the phoneme after the phoneme of interest is a pause.
Since the processing from step 311 to step 317 is the same as the processing from step 314, only the processing from step 314 will be described.

First, based on the phoneme of interest, the phoneme two phonemes before the phoneme of interest, the phoneme one phoneme before the phoneme of interest, the phoneme one phoneme after the phoneme of interest, and the phoneme two phonemes after the phoneme of interest,
Search the primary search table of Table 2 (step 314)
1).

The phoneme of interest is "taro @ saN% ohay
In the case of "N" in "o", the target phoneme N belongs to category 1, the phoneme s two before belongs to category 5, the phoneme a before belongs to category 1, and the phoneme o two afterwards.
Belongs to category 1 and the other is% (pause). Therefore, from the primary search table in Table 2, the average time (= average phoneme duration) of the target phoneme N and the secondary search table No.
Is required.

In this example, the average time (= average phoneme duration) of the target phoneme N is determined to be 90 msec, and the secondary search table No is determined to be 997. Table 3 shows No. 99
7 shows the contents of the secondary search table.

[0028]

[Table 3]

Next, the average phoneme duration of the target phoneme N is adjusted based on the number of moras in the exhalation paragraph including the target phoneme and the secondary search table specified by the primary search table ( Step 3142).

When a person reads out a character, the character reads out the character while taking a breath. This period of breathing is called an exhalation paragraph. When the number of characters between the punctuation marks is large, the number of mora between the exhalation paragraphs usually increases, and it becomes quick. Therefore, when the number of mora between exhalation paragraphs (the number of characters between punctuation marks) increases, the average phoneme duration of phonemes during that period decreases. In this example, the exhalation paragraph including the target phoneme is “taro @ saN” and has 5 mora. Based on the secondary search table shown in Table 3, 5
The correction value corresponding to the mora is read. In this example,
Since the correction value corresponding to 5 moras is 0 ms, the average phoneme duration of the phoneme N of interest remains unchanged.

Next, it is determined whether or not the phoneme of interest is within 3 moras of the sentence end (step 314).
3). As a trend of modern languages, there is a tendency that the language, that is, the end of the sentence, is not clearly stated. For this reason, the phoneme duration time becomes short at the end of the sentence. Therefore, the phoneme of interest is 3 at the end of the sentence.
If it is within the mora, the average time phoneme duration of the target phoneme is adjusted based on the secondary search table obtained from the primary search table (step 314).
4). Each correction value in the secondary search table is also obtained by performing statistical processing from a plurality of actual utterance data.

In this example, since the target phoneme N is not within 3 moras at the end of the sentence, the average phoneme duration is not adjusted based on the end-of-sentence determination. Therefore, the average phoneme duration of the target phoneme N is finally determined to be 90 msec.

In Table 2, the secondary search table N
o. 997-No. When the contents of 999 are the same, it is preferable to combine these three secondary search tables into one in order to reduce the memory capacity for storing the secondary search table.

[0034]

According to the present invention, the phoneme duration of each phoneme can be reduced in a shorter time than a statistical analysis such as quantification type I analysis, which considers all parameters considered to affect the phoneme duration. Analysis can be performed, and there is no danger that a statistical error locally increases due to inconsistent parameters or the like.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a processing procedure of a speech synthesis method.

FIG. 2 is a flowchart illustrating a processing procedure of a phoneme duration control method.

 ──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Makoto Hashimoto 2-5-5 Keihanhondori, Moriguchi-shi, Osaka F-term in SANYO Electric Co., Ltd. (reference) 5D045 AA07

Claims (4)

[Claims] 1. A speech waveform segment corresponding to a Japanese phoneme symbol is prepared in advance, and a speech waveform segment corresponding to the input phoneme symbol sequence is connected in the order of the input phoneme symbol sequence, and the prosody regulation control is performed. In a Japanese speech synthesis method for performing waveform shaping and synthesizing speech corresponding to an input phoneme symbol string, category classification is performed in advance for all phoneme symbols, and a phoneme duration for each input phoneme in the input phoneme symbol string. A method for synthesizing Japanese speech, comprising determining a length based on a category of an input phoneme included in an input phoneme symbol string. 2. The categories of phoneme symbols are: (1) vowels and vowels, (2) long vowels, (3) unvoiced plosives, (4) affricates, (5) unvoiced fricatives, (6) semi-vowels and streams. sound,
2. The method according to claim 1, wherein the method is classified into (7) nasal consonant, (8) voiced plosive, and (9) other independent phonemes. 3. The phoneme duration for each input phoneme is
Based on the category of the phoneme of interest, the category of the phoneme before the phoneme of interest, the category of the phoneme before the phoneme of interest, the category of the phoneme after the phoneme of interest, and the category of the phoneme after the phoneme of interest. 3. The method for synthesizing Japanese speech according to claim 1, wherein 4. The phoneme duration for a target phoneme is determined in consideration of the number of mora of a breath paragraph containing the target phoneme and the mora position of the target phoneme. Japanese speech synthesis method described in any of them.