KR19980079119A - Speech Synthesis Database, How to Create It, and Speech Synthesis Method Using the Same - Google Patents

Abstract

The present invention relates to a speech synthesis method based on a rhyme rule, and more particularly, to a database required for speech synthesis, a method of creating such a database, and a speech synthesis method using the same.

Database for speech synthesis according to the present invention; A voice language data unit for storing voice signals reading a plurality of sentences; And an index table storing positions, lengths, sizes, positions of pitches, and variations of phoneme strings corresponding to the synthesis unit extracted from the speech language data stored in the speech language data unit.

If the database according to the present invention is provided in the form of a CD-ROM, there is an advantage that a DB of various voices can be implemented in a short time than the conventional one.

Description

Speech Synthesis Database, How to Create It, and Speech Synthesis Method Using the Same

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis method based on a rhyme rule, and more particularly, to a database required for speech synthesis, a method of creating such a database, and a speech synthesis method using the same.

In the conventional speech synthesis technology, since a database storing a speech signal is made of a synthesis unit, for example, a phoneme or a syllable, it is difficult to add or change the database. To apply a new synthesis unit, a database is newly created from the beginning. There are drawbacks to this.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a database which is free to change the synthesis unit and can improve the naturalness of the synthesized sound.

Another object of the present invention is to provide a method for creating the database.

It is still another object of the present invention to provide a speech synthesis method capable of freely changing the synthesis unit and improving the naturalness of the synthesized sound.

Figure 1 shows a speech synthesis method according to the prior art.

2 shows the structure of a database according to the present invention.

3 schematically shows a method of creating a database according to the present invention.

4 schematically shows a speech synthesis method according to the present invention.

FIG. 5 shows the speech synthesis part in more detail as shown in FIG. 4.

Database for speech synthesis according to the present invention to achieve the above object; A voice language data unit for storing voice signals reading a plurality of sentences; And an index table storing positions, lengths, sizes, positions of pitches, and variations of phoneme strings corresponding to the synthesis unit extracted from the speech language data stored in the speech language data unit.

According to another aspect of the present invention, there is provided a method of creating a speech synthesis database, the method comprising: collecting speech signals reading a plurality of sentences to create speech language data; Extracting phoneme position, length, size, pitch position, and amount of change in each sentence by analyzing the written speech language data; And extracting and indexing only phoneme strings corresponding to the synthesis unit from the extracted data, and storing the phoneme strings together with phoneme positions, lengths, sizes, pitch positions, and amounts of change.

According to another aspect of the present invention, there is provided a speech synthesis method comprising: a phoneme string corresponding to a speech unit storing voice signals reading a plurality of sentences and a synthesis unit extracted from speech language data stored in the speech language data unit. Creating a database including an index table storing positions, lengths, sizes, positions of pitches, and variations; Extracting a phoneme string corresponding to the synthesis unit and a length, a size, a position and a change amount of the phoneme corresponding to each of the phonemes from the input document to generate the synthesized sound; Searching for indexes having information matching the phonemes included in the phoneme strings, their lengths, sizes, positions of pitches, and amounts of change by referring to an index table stored in the database; Extracting speech signals for synthesis from speech language data stored in the database by referring to phoneme location information of the searched indices; And generating a synthesized sound by connecting the extracted voice signals. Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

1 shows a method of creating a database used in a conventional speech synthesis method. In Fig. 1, reference numeral 10 is a word for recording, 12 is a source unit database, 14 is a rhythm rule processor, and 16 is a voice processor.

The database 12 shown in FIG. 1 is a collection of synthesis units of a previously promised form. The method for creating this database 12 is as follows.

First, the speech synthesis unit is determined. For example, select the appropriate unit from units such as phoneme, diphone, and triphone. Since phoneme has too few units and triphone has too many, you usually choose diphone.

Second, in order to make unlimited speech synthesis with a set unit, all the pronunciation units in Korean are made. For a diphone you need about 4,000 units. To make all of these units, we generate meaningless words that can be pronounced for each diphone, which the announcer reads and records.

Thirdly, the recorded voice is analyzed to extract the parameter for synthesis and data for prosody control, and segmentation is performed by a skilled expert only on the diphone part necessary for synthesis.

Finally, the partitioned units are collected in a certain order to create a database 12 for speech synthesis.

The biggest problem in the prior art as described above is that when changing to a new unit or replacing a bad recording segment, all of the above processes must be repeated from the beginning.

In addition, because the meaningless words are recorded and uttered, very unnatural synthesized sounds may occur in the actual synthesized sounds. Because the pronunciation of words and sentences are pronounced differently, even if they sound the same, the method of pronunciation is significantly different, and because the pronunciation of meaningless words, the synthesized sound is very uniform and monotonous, unlike the sound of a real person.

2 shows the structure of a database according to the present invention. In Fig. 2, reference numeral 20 denotes a speech language data portion, and 22 denotes an index table.

The speech language data unit 20 stores voice signals reading a plurality of sentences.

The index table 22 stores positions, lengths, sizes, pitch positions, and amounts of change of phoneme strings corresponding to the synthesis unit extracted from the speech language data stored in the speech language data unit 20.

Here, the index table can be variably adjusted according to the synthesis unit, and even if only a speech language material is provided, even when implementing a multilingual speech synthesis system, the index table can be implemented.

The database creation method according to the present invention for creating the database shown in FIG. 2 is as follows.

Firstly, voice signals are read from a plurality of sentences to prepare voice language data.

Second, the phonetic language data is analyzed and the phoneme location, length, size, pitch location and variation in each sentence are extracted.

The first and second processes will be described with reference to FIG. 3.

First, a document in which the phoneme frequency is properly harmonized is made, and the announcer records the pronunciation.

Detects voice sections in the recorded voice and detects pitch in each voice section. Firstly, pitch is detected by automatic pitch detection algorithm, and secondly by expert. When pitch marking is completed, a phoneme string to be pronounced is generated using the pitch marking, and the start position of each phoneme is labeled on the pronounced voice signal to match the pitch. Through this process, the position, length, size information, pitch position and variation of each phoneme in the sentence can be accurately known.

Third, only the phoneme strings corresponding to the synthesis unit are extracted and indexed from the extracted data and stored along with the phoneme position, length, size, pitch position and variation.

If you want to use the synthesis unit as a diphone, the index is created and stored by combining only the diphones in the speech language material. If you want to make a triphone, the index is created by combining only the triphones. At this time, the pitch information necessary for the rhyme control, the length and size information of the note are also extracted.

4 schematically shows a speech synthesis method according to the present invention.

In Fig. 4, reference numeral 40 denotes a recording document, 42 denotes a speech language data unit, 44 denotes an index table, 46 denotes a rhythm rule processing unit, and 48 denotes an input document for generating synthesized sounds. Here, the speech language data section 42 and the index table 44 constitute a database according to the present invention shown in FIG.

The speech synthesis method according to the present invention is as follows.

First, a speech language data section storing speech signals reading a plurality of sentences;

A database is created that includes an index table that stores positions, lengths, sizes, positions of pitches, and variations of phoneme strings corresponding to the synthesis unit extracted from the speech language data stored in the speech language data unit. This process is the same as shown in FIG.

Second, a phoneme sequence corresponding to the synthesis unit and a length, a size, a position and a change amount of the phoneme corresponding to the phoneme sequence are extracted from the input document to generate the synthesized sound.

Third, the indexes stored in the database are searched for the phonemes included in the phoneme strings and the indexes having information matching the lengths, sizes, positions of the pitches, and amounts of change.

Fourth, voice signals required for synthesis are extracted from voice language data stored in the database by referring to phoneme location information of the searched indices.

Finally, the extracted speech signals are connected to generate a synthesized sound.

After the second process will be described with reference to FIG. First, the input sequence is analyzed to produce a phoneme sequence to be synthesized. By the phoneme sequence and the rhyme control rules, the synthesis unit, pitch value, voice length, size information, etc. which each phoneme should have are generated. In this case, the segment most similar to the required synthesis unit is obtained from the speech language data by using the index generated by analyzing the speech language data.

Here, the process of finding the most similar segment is expressed by feature weighting. For example, in order to find ahK (ㅏ + 에서 의 in science), it is a process to find a rule for determining a segment having a value most similar to a required value among various kinds of pitch values and length size information with an appropriate weight value. Through this process, when the segments matching the phoneme sequence are reorganized, each segment is connected and synthesized so that no heterogeneous sound is generated.

As described above, the speech synthesis method according to the present invention has the advantage that the synthesis unit can be freely changed and synthesized.

In addition, since the synthesized sound is generated based on the voice segment extracted from the natural sentence, the naturalness of the synthesized sound can be improved.

In addition, multilingual speech synthesizers can be easily implemented as long as the speech language materials are well equipped.

Since most of the speech synthesizer is taken up by the database, if the database according to the present invention is provided in the form of a CD-ROM, there is an advantage that a DB of various voices can be implemented in a short time than the conventional one.

Claims (3)

A voice language data unit for storing voice signals reading a plurality of sentences; And And an index table for storing positions, lengths, sizes, positions of pitches, and variations of phoneme strings corresponding to the synthesis unit extracted from the speech language data stored in the speech language data unit. Preparing voice language data by collecting voice signals reading a plurality of sentences; Extracting phoneme position, length, size, pitch position, and amount of change in each sentence by analyzing the written speech language data; And And extracting and indexing only phoneme strings corresponding to the synthesis unit from the extracted data, and storing the phoneme strings together with phoneme positions, lengths, sizes, pitch positions, and amounts of change. An index storing the position, length, size, pitch position, and amount of change of the phoneme strings corresponding to the synthesis unit extracted from the speech language data section storing the speech signals reading a plurality of sentences and the speech language data stored in the speech language data section. Creating a database comprising a table; Extracting a phoneme string corresponding to the synthesis unit and a length, a size, a position and a change amount of the phoneme corresponding to each of the phonemes from the input document to generate the synthesized sound; Searching for indexes having information matching the phonemes included in the phoneme strings, their lengths, sizes, positions of pitches, and amounts of change by referring to an index table stored in the database; Extracting speech signals for synthesis from speech language data stored in the database by referring to phoneme location information of the searched indices; And And generating a synthesized sound by connecting the extracted speech signals.