JPH037995A - Generating device for singing voice synthetic data - Google Patents

Abstract

PURPOSE:To improve the quality of data by segmenting a singing voice by using character string, pitch, strength, and length data generated from a score as reference data, and extracting pitch, strength, and length data. CONSTITUTION:The reference data generating means consisting of a score input part 1 and a phoneme information generation part 2 generates phoneme information and rhythm information as voice data from the score. Then the segmentation processing part 5 of a rhythm information extracting means segments the input singing voice by referring to the reference voice data. Data on the length in the voice data corresponding to the song voice inputted by the segmentation and data on the strength and pitch are found. The found rhythm information from the input song voice and the phoneme information found from the score are stored as voice data on a storage medium or outputted to a transmission line, so that the data is used as the voice data of a voice synthesizing circuit X.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention is directed to the creation of song speech synthesis data that is input to a speech synthesis circuit that performs rule synthesis of song speech based on speech data consisting of phonological information and prosody information. It is related to the device.

[Prior Art] In general, a speech rule synthesis method is a method that generates speech based on phonetic and linguistic rules by inputting phonological parameters that control phonological information and prosodic parameters that control prosodic information. Although the amount of information to be processed (external input information, internally stored information) is very small, it is a speech synthesis method that can output synthesized speech of a large number of words and sentences.

On the other hand, the text synthesis method is the ultimate speech synthesis method that performs speech synthesis only from character strings. It automatically creates phonological information and prosody information from character strings and performs rule synthesis based on that information. It will be realized. However, although it is currently being actively researched, the quality of synthesized speech is still low.

Here, text synthesis consists of two parts: automatically creating phonological information and prosody information from character strings, storing them as audio data on a storage medium, or outputting them to a transmission line, and the other part: It is conceivable to extract voice data from the system and divide it into two parts, a part that outputs the sound by rule synthesis.In this way, the part that outputs the synthesized speech can be made compact, that is, it can be made smaller, lighter, and lower in cost. (See Japanese Patent Application No. 62-263298).

By the way, if we think about the case of synthesizing singing voices, the ultimate form in this case would be to input musical scores and output singing voices.Comparing this to the text synthesis method, in the case of musical scores, text is generated. Since it includes not only strings (lyrics) but also other information, it is easy to generate prosodic information, and it is possible to obtain higher quality with a simpler structure than the text synthesis method. -177314
(see issue).

In this case as well, in the same way as we considered text synthesis, we can automatically create phonological information and prosody information from the musical score, and store it as audio data on a storage medium or output it to a transmission path. It is conceivable to divide it into two parts: a part that extracts audio data from a storage medium or a transmission path, and outputs audio through rule synthesis.

Now, in the ultimate form, when a musical score is input, a singing voice is output, but this is not always the case.

In other words, professional singers may be able to create a song with a more human touch and higher artistic quality by singing it slightly off the sheet music. Therefore, it is conceivable to create audio data by inputting actual singing voices instead of musical scores, and applying the ``audio code creation method'' of Japanese Patent Application No. 138517-1980.

[Problems to be Solved by the Invention] In other words, in order to create audio data by inputting audio, height data, intensity data, and length data for each sound (syllable) are required (see Figure 3). Therefore, we first need to segment the input audio. Considering this segmentation, since a character string (phoneme string) is also input, it is a segmentation with known phoneme, which is easier than segmentation with unknown phoneme performed in speech recognition or the like. However, with the current level of technology, complete segmentation is still not possible, and as a result, there is a problem in that voice data of low quality is created. In addition, the pitch extraction algorithm for obtaining height data is also 1.
There is no such thing as 00% perfect, but 1/2 (half pitch) of the true value.
The value may be doubled (double pitch), and this also causes the problem of creating voice data of low quality.

The present invention has been made in view of the above-mentioned points, and its purpose is to perform highly accurate segmentation of actual singing voices, and as a result, to provide a rule-based synthesizer for singing voices. An object of the present invention is to provide a song voice synthesis data creation device that creates voice data for outputting higher quality song voices.

[Means for Solving the Problems] The song speech synthesis data creation device of the present invention provides song speech synthesis data that is input to a speech synthesis circuit that performs rule synthesis of song speech based on speech data consisting of phonological information and prosody information. The data creation device includes a reference data creation means that creates a character string from the input musical score, height, intensity, and length data according to rules, and segmentation of the input singing voice using each of the above data as reference data. and a prosodic information extracting means for extracting height, intensity, and length data, and the data extracted by the prosodic information extracting means is output as audio data.

[Function] The present invention is configured as described above, and is provided in an apparatus for creating song speech synthesis data that is input to a speech synthesis circuit that performs rule synthesis of song speech based on speech data consisting of phonological information and prosody information. , creates character strings and height, strength, and length data based on rules from the input musical score, and uses the above data as reference data to segment the input singing voice and calculate the height, strength, and length. This is a system that extracts data and outputs the extracted data as audio data.Since the intensity data and length data are extracted based on the actual singing voice, it is a speech synthesis circuit using a rule synthesis method. As the singing voice synthesis data input into the software, voice data for natural singing voices is obtained, and the voice data is extracted by segmenting the singing voice by referring to the voice data created from the musical score. Audio data that can be used to synthesize high-quality singing voices can be obtained at low cost with a simple circuit configuration.

[Embodiment] Figure 1 shows an embodiment of the present invention, in which song speech synthesis data is created to be input to a speech synthesis circuit that performs rule synthesis of song speech based on speech data consisting of phonological information and prosody information. The device includes a reference data creation means that creates a character string from the input musical score, height, intensity, and length data according to rules, and performs segmentation of the input singing voice using each of the above data as reference data. and prosodic information extraction means for extracting length, strength, and length data,
The data extracted by the prosody information extraction means is output as audio data. Here, the reference data creation means includes a musical score input unit 1 consisting of a manual key device or an optical musical score reader that inputs character string data and height data of the musical score, and a musical score input unit 1 that extracts prosody information from the input musical score data according to rules. It is composed of a prosodic information creation section 2 that creates prosody information.

In addition, a singing voice input section 3 such as a microphone, a feature extraction section 4 that extracts the features of the input singing voice, and a singing voice from which the features have been extracted are referred to each data created by a reference data creation means. It is comprised of a segmentation processing section 5 that divides each phoneme into segments, and a prosodic information extraction section 6 that extracts height, strength data, and length data of each segmented phoneme.

In addition, string data and height, strength, and length data are
The data is output as audio data via an encoding unit 7 that encodes the data.

Now, the reference data creation means creates phonological information and prosody information as audio data from the musical score. In this embodiment, regarding prosody information, strength data is not necessary and only height and length data are created as reference speech data.Next, the segmentation processing unit 5 of the prosody information extraction means The input song voice is segmented by referring to the reference voice data. This process is performed, for example, as follows.

First, the boundary points of each phoneme are calculated based on the length data in the created reference audio data, and the length of the resulting audio is expanded or contracted so that it matches the length of the input song audio. , and move the boundary points of each phoneme accordingly. The vicinity of the boundary point of each phoneme found in this way is defined as the range where the corresponding boundary point of the input song voice exists (see Figure 2), and the boundary point is determined by considering the properties of the phonemes before and after it. . Furthermore, Figure 2 (IL) shows the boundary points of each phoneme based on the audio data created from the musical score.
Figure (b) shows the case where the boundary points are linearly extended to match the length of the input singing voice, and Figure 2 (
c) shows the energy, and FIG. 2(d) shows the formant frequency.

Next, from the segmentation results obtained in this way, length data (for example, the length of each vowel and consonant) and strength data (for example, the length of each mora) of the voice data corresponding to the input singing voice are calculated. Find the power of the representative point).

Next, obtain height data. For example, if the data is the pitch of the representative point of each mora, when finding the pitch for the desired point, the pitch search range is set near the value of the audio data found from the musical score, and the pitch is found within that range. When obtained in this way, the search range is narrow, so accuracy is high (half-pitch or double-pitch errors are eliminated), and calculations are faster.

The prosody information from the input singing voice obtained in this way and the phonological information obtained from the musical score are converted into audio data, which can be stored in a storage medium or output to a transmission path to form the speech synthesis circuit X.
used as audio data.

As described above, by inputting the voice data created based on the actual singing voice to the voice synthesis circuit X, a natural singing voice can be obtained.

In addition, since the audio data created from the musical score is extracted by performing segmentation of the singing audio, it is possible to obtain audio data that can be synthesized into high-quality singing audio with a simple circuit configuration and at low cost. It has become.

[Effects of the Invention] The present invention is configured as described above, and provides an apparatus for creating song speech synthesis data that is input to a speech synthesis circuit that performs rule synthesis of song speech based on speech data consisting of phonological information and prosody information. In this step, a character string and height, strength, and length data are created according to rules from the input musical score, and the above data is used as reference data to segment the input singing voice and calculate the height, strength, and length. This method extracts the sound data and outputs the extracted data as audio data.Since the intensity data and length data are extracted based on the actual singing voice, it is not possible to synthesize speech using the regular synthesis method. As the singing voice synthesis data input to the circuit, voice data for natural singing voices is obtained, and the voice data is extracted by segmenting the singing voice by referring to the voice data created from the musical score. The effect is that voice data capable of synthesizing high-quality singing voices can be obtained at low cost with a simple circuit configuration.

[Brief explanation of the drawing]

FIG. 1 is an example of a song voice audio data creation device according to the present invention, and FIGS. 2 and 3 are explanatory diagrams of the same operation. X is a speech synthesis circuit, Y is a song speech synthesis data creation device, 1
2 is a musical score input section, 2 is a prosodic information creation section, 3 is a singing voice input section, 4 is a feature extraction section, 5 is a segmentation processing section, 6
7 is a prosodic information extraction section, and 7 is an encoding section.

Claims (1)

[Claims] (1) In a device for creating singing speech synthesis data that is input to a speech synthesis circuit that synthesizes singing speech according to rules based on speech data consisting of phonological information and prosody information, character strings and high reference data creation means for creating pitch, strength and length data; and prosodic information extraction means for extracting pitch, strength and length data by segmenting the input singing voice using each of the above data as reference data. A device for creating song speech synthesis data, characterized in that the data extracted by the prosodic information extraction means is output as speech data.