Synthesis of speech from text - Patent Review 5463713

Description Submit all comments and votes

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to improvements to an apparatus provided for speech synthesis of text by means of a regular synthetic method, and more particularly, to improvements in an apparatus for speech synthesis in which the accent of text data is controlled by an accent control method.

2. Description of the Prior Art

The automatic conversion of text to synthetic speech is commonly known as text to speech conversion or text to speech synthesis. A number of different techniques have been developed to make speech synthesis apparatus practical on a commercial basis. FIG. 5 shows a typical speech synthesis apparatus in which speech is synthesized by a regular synthetic method such as by using a connection rule of mora or rule of phoneme. The speech synthesis apparatus includes an accent control section 13 where a phrase pattern calculating section 13a is arranged to calculate a phrase component (which indicates the height of the voice in the part sandwiched between pauses) according to the number of mora contained in the text, and an accent pattern calculating section 13b is arranged to calculate an accent component (which shows the height of the sound of each word). The phrase component and the accent component are added to each other in a speech synthesizing section 14, and an accent control pattern is calculated as shown in FIG. 6. In general, the phrase component is continuously changed from a high pitch to a low pitch due to the lowering of the pressure under the glottis. The interpolation of the accent component is carried out by putting a pitch target value to each analysis element and linearly interpolating between pitches, or by putting three pitch target values to each analysis element and linearly interpolating among their pitches.

With the above mentioned accent control method in the speech synthesizer, an accent is applied to the synthesized speech by calculating the phrase component and the accent component. The accent component is determined by applying plural target pitches to each mora and linearly interpolating among their pitches.

However, since the pitch of the accent component is simply determined according to the height of the accent, the synthesized speech sounds mechanical due to its uniform change in the pitch. Further, since the interconnections between the syllables, and between the clauses are not taken into consideration, it is apt to cause unsmoothness in the height change of the accent and between moras. Accordingly, the synthesized speech generated by this method sounds unnatural.

In order to solve the above mentioned problem, another accent control method has been proposed, in which the changing coefficient of the pitch in the mora is determined by the linear function calculation according to the accent environment, in detail, according to the height of accent, the position in phrase, cotinuative phoneme or not, the accent height of forward and back mora of the mora, positional relationship with clause, and the target value at forward and back in the mora.

With such an accent control method, improved synthesized speech is provided. However, it is difficult to easily understand the changed accent pattern during the maintenance or when the a variable number is defined since the changing coefficient includes the variable number for controlling. This difficulty is further increased in proportion to the increase in the accent pattern. Furthermore, the calculating operations become more complicated since the function for generating the accent pattern and the defining of the variable number become complex.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide an improved apparatus for speech synthesis which is free of the above mentioned drawbacks.

An apparatus for synthesizing speech from text, in accordance with the present invention, comprises a language processing section which determines an accent environment of each mora for each phrase of the text. In a basic accent pattern table, a basic accent pattern is classified according to the accent environment of the mora. The basic accent pattern includes a pitch data which is edited from real voice data according to the accent environment. A basic accent pattern processing section selects the basic accent pattern of each mora from the basic accent pattern table according to the accent environment and processes the basic accent pattern in the pitch according to the accent environment. A correcting section receives the basic accent pattern in the pitch in the basic accent pattern processing section and corrects the pitch according to the number of moras in each phrase and the position of the mora in the phrase so as to correct the data in the corrected accent component. A phrase pattern processing section determines a phrase component according to the number of moras in each phrase which is of the accent environment. A speech synthesizing section synthesizes speech according to an accent control pattern of the text which is obtained by adding the basic accent pattern and the basic phrase pattern.

With this arrangement, the accent pattern is easily understood by being imaged from the table data, and the maintenance of the speech synthesizer is easily carried out by correcting the data of the basic accent pattern table.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will be described in greater detail by reference to the following description taken in connection with the accompanying drawings, in which:

FIG. 1 is a block diagram of an apparatus of a first embodiment of a speech synthesizer according to the present invention;

FIG. 2 shows tables which disclose a procedure for translating the accent pattern into the form of digitized table by leveling in use for the first embodiment;

FIG. 3 is an accent pattern table which is used in a second embodiment of the speech synthesis apparatus according to the present invention;

FIG. 4 is another accent pattern table which is used in a third embodiment of the speech synthesis apparatus according to the present invention;

FIG. 5 is a block diagram of a conventional speech synthesizing apparatus; and

FIG. 6 shows graphs for explaining the generation of an accent control pattern by the apparatus of FIG. 5.

DETAILED DESCRIPTION OF THE INVENTION

Referring now to FIGS. 1 to 3, there is shown a first embodiment of an apparatus S for speech synthesis according to the present invention.

The apparatus S for speech synthesis comprises a text input section 1 at which text is inputted for being enunciated, as shown in FIG. 1. The text input section 1 is connected to a language processing section 2 in which the text content is analyzed by means of the morpheme analysis. The data processed in the language processing section 2 is sent to an accent control section 3 in which a phrase pattern calculating section 4 and an accent pattern processing section 5 are parallelly arranged. The phrase pattern calculating section 4 and the accent pattern processing section 5 receive the respective data from the language processing section 2. The accent pattern precessing section 5 includes a basic accent pattern processing section 6 and a correct processing section 8. The basic accent pattern processing section 6 is communicated with a basic accent pattern table 7 through which a proper basic accent pattern is selected. The data from the phrase pattern calculating section 4 and the accent pattern processing section 5 are sent to a speech synthetic section 9.

As a result of the language processing of the inputted text, the number of moras of each clause and the accent data of each mora in the text data are determined and sent to the accent pattern processing section 5. At the accent pattern processing section 5, the basic accent pattern is looked up from the basic accent pattern table 7 in accordance with the input accent environment. The basic accent pattern table 7 previously stores the date which is of a table of the pattern data gained by the pitch analysis of the original (real) speech. At the basic accent pattern table 7, a plurality of accent amount of each mora are classified according to the combination between the accent of the mora and the accent of its forward and back mora of the main mora. At the correcting section 8, the correction value of the basic accent pattern is determined according to the number of moras between the space and the position of the mora of the accent pattern processed in the language processing section 2. In accordance with the correction value, the basic accent pattern is corrected and the corrected accent pattern data is sent to the speech synthesizing section 9 for being combined with the phrase pattern data. The accent component and the phrase component are overlapped and function as an accent control pattern in the speech synthesizing section 9. The pitch of each syllable is controlled at a speech synthesizing section 9 according to the accent control pattern. The synthesized speech is outputted through an articulation filter according to the voice wave pattern and the parameter of the articulation filter which are in cooperated with each syllable.

FIG. 2 shows an original accent pattern table (a) in which the accent pattern is shown in the form of the graph produced, and a digitized accent pattern table (b) which is produced by digitizing the original table (a). The real speech data-base shown by the original table (a) is manually classified by means of the pitch analysis of the vocalization (speaking) according to each accent environment. For example, a plurality of the pitch data of the original speaking, in which the form of the accent combination (the combination of low accent L and high accent H) is LHL, are classified according to every accent environment in which the mora is analyzed with respect to the existence of the contiguous phoneme, the accent height of forward and back mora, the position of punctuation and the like. Accordingly, the proper accent pattern data is selected according to the accent environment even if the height change (LHL) of the accent of the mora is the same as that of the other moras. The basic accent pattern table (b) is determined by leveling and classifying the data of the real speaking data-base shown in the table (a) in accordance with the accent environment, and the table (b) of the basic accent pattern is memorized in the basic accent pattern table 7.

The manner of operation of the thus arranged apparatus S for speech synthesis will be discussed hereinafter with reference to one popular sentence of Japanese language.

For illustration the content of the text is assumed to be "Kyo wa i tenki desu." which means "It is fine today." in English. The sentence is normally described in Japanese (Kanji and Kana), it is herein described by Romaji which is a method for writing Japanese in Roman character in order to facilitate the understanding of the discussion. The text content is inputted from the text input section 1 and sent to the language processing section 2. In the language precessing section 2, the sentence described in Japanese (Kanji and Kana) is translated into Romaji. Since Japanese is a an isosyllabic language, the sentence described in Romaji directly indicates the pronunciation of the sentence. Furthermore, the following table is obtained in the language processing section 2:

TABLE 1 ______________________________________ ANALYSIS ELE. ACCENT CV/C SEGMENT ______________________________________ KYO 2 CV (HIGH) WA 1 CV (CLAUSE B.P) (LOW) I 2 V (CLAUSE B.P.) TE 1 CV N 1 V KI 1 CV DE 1 CV SU 1 CV (PAUSE) ______________________________________ where CV is a consonant + vowel and V vowel.

where CV is consonant +vowel and V vowel.

Table 1 shows the accent environment of the inputted text, in detail, the height of accent, the accent position in phrase, the kind of mora and the like. Language processing section 2, analyzes whether each mora is cotinuative phoneme or not, now high is the accent height of forward and back mora of each mora, what positional relationship does each mora have in the clause, what is the target value at forward and back in the mora and the like. According to the data obtained in the language processing section 2, a phrase component is calculated. Further, in the basic accent pattern processing section 6, an accent component of each mora of the text is selected from the basic accent pattern table 7 according to the accent environment. In the correcting section 8, the correction value of the basic accent pattern is determined according to the number of moras between the space and the position of the mora of the accent pattern processed in the language processing section 2. In accordance with the correction value, the basic accent pattern is corrected and the corrected accent pattern data is sent to the speech synthesizing section 9 for being combined with the phrase pattern data.

With the thus arranged apparatus for speech synthesis, the accent pattern data corresponding to all of the accent environment is stored in the accent pattern table 7. Accordingly, the accent pattern data is easily looked up by using the accent environment as an index when the maintenance of the data is carried out, or the correction value is determined. Furthermore, since the accent pattern is stored in the accent pattern table 7 in the form of a plural accent amount (pitch) for every mora, the accent pattern for every mora is imaginably impressed. Therefore, the accent pattern is easily understood and amended as compared with a conventional method in which the accent component is calculated from functions and coefficient data. Additionally, since the accent pattern data has been generated from a real voice, the clear and realistic voice is easily obtained.

Referring to FIG. 3, there is shown another accent pattern table 7a of a second embodiment of an apparatus S for speech synthesis according to the present invention. The second embodiment is similar to the first embodiment except for the basic accent pattern table 7.

The basic accent pattern table 7a of the second embodiment is classified on the basis of the real voice data-base so that the accent amount (pitch) of each mora is determined according to the accent environment, more particularly, according to the boundary of accent phrase such as whether the accent is positioned at a forward or back position, or whether the accent does not exist.

With the thus arranged apparatus for speech synthesis, the basic accent pattern is classified according to the boundary position of the accent in the basic accent pattern table 7a. Accordingly, the boundary position of the accent becomes clear and therefore the synthetic voice has a clear boundary position of the accent, that is, the synthetic speech sounds with modulation.

Referring to FIG. 4, there is shown another basic accent pattern table 7b of a third embodiment of the apparatus S for speech synthesis according to the present invention. The third embodiment of the apparatus S for speech synthesis is similar to the first embodiment except for the basic accent pattern 7b.

The basic accent pattern table 7b is classified so that the accent amount in each mora is determined according to the property of mora which indicates the difference of the structure of mora. The property of mora is supplied from the language processing section 2 to the basic accent pattern processing section 6 with the number of moras in each boundary and the accent pattern. Accordingly, the proper accent pattern is looked up from the basic accent pattern table 7b according to the property of the mora and the accent environment.

With the thus arranged apparatus S for speech synthesis, the basic accent pattern is prepared according to the property of mora. Accordingly, even if the mora has the same accent environment as the other, the proper accent pattern for the mora is selected since the basic accent pattern table 7b is classified in accordance with whether the pattern of the mora is vowel mora, vocal consonant+vowel (mora), or voiceless consonant+vowel (mora). Furthermore, the accent pattern is classified according to whether the vowel part has a long sound or not. Accordingly, the synthesized sound further approaches the human voice which has a different accent according to the difference of the mora property.

While the embodiments of the present invention have been shown and described so that the apparatus processes the text written in the Japanese language, it will be appreciated that the principle of the present invention may be applied to other languages.

* * * * *