JP4630038B2 - Speech waveform database construction method, apparatus and program for implementing this method - Google Patents

Description

  The present invention relates to a speech waveform database construction method, an apparatus and a program for executing the method, and more particularly to construction of a speech waveform database used in a waveform-connected text speech synthesizer. The present invention relates to a speech waveform database construction method for implementing high-quality speech synthesis with a small execution time and calculation amount required for speech synthesis, and an apparatus and program for implementing this method.

In the recent speech synthesis technology field, a large volume storage device stores a large amount of tens of minutes to several tens of hours of real voice data to construct a speech waveform database. A waveform connection type speech synthesis method has been proposed in which speech waveforms having an appropriate length are cut out from a speech waveform database and connected to generate synthesized speech (see Patent Document 1). In this waveform connection type speech synthesis method, when a speech waveform having an appropriate pitch and length does not exist in the speech waveform database, intonation of the synthesized speech to be generated may become unnatural. Therefore, in order to give a desired intonation to the synthesized speech, a method has been proposed in which signal processing is performed on the speech waveform cut out from the speech waveform database and the pitch, length, and other speech parameters are changed (Patent Literature). 1).
Japanese Patent No. 2761552 Specification

However, when the pitch and length of the speech waveform are changed by signal processing, there is a problem that the sound quality deteriorates when the change amount is large. Further, when the signal processing is performed, there is a problem that the amount of calculation is further increased and the time required for the synthetic speech generation processing is increased as compared with the case where the waveform connection is simply performed.
The present invention has been made in view of the above problems, and in waveform-connected text-to-speech synthesis, by adding a modified speech waveform in which a speech parameter is changed to a real voice waveform, a speech waveform database is configured. A speech waveform database construction method capable of generating high-quality synthesized speech with a small execution time and amount of calculation required for speech synthesis, a method for constructing a speech waveform database, an apparatus for implementing this method, and a program are provided. is there.

Claim 1: In a speech waveform database construction method for constructing a speech waveform database used in a waveform-connected text speech synthesizer, designation is made within a range in which sound quality deterioration of a speech waveform does not occur with respect to a prepared real voice waveform Based on the changed amount parameter, the pitch, length, power, or other speech parameter is changed by signal processing to newly generate a prosody changed speech waveform, The amount of change made in the step of generating the prosody modified speech waveform from the given phoneme label information indicating the content and position of the utterance, power information, pitch information, pitch mark information, and other speech information The step of generating new audio information by executing a change corresponding to the In addition to voice information, performing the step of generating a speech waveform database by using the the real voice one or more types of prosody Changed speech waveform and the voice information is derived from the speech waveform, a speech waveform database building method Configured.

Claim 2: In a speech waveform database construction device for constructing a speech waveform database used in a waveform-connected text speech synthesizer, within a range in which sound quality deterioration of the speech waveform does not occur with respect to a prepared real voice speech waveform A voice waveform changing unit 4 that changes a pitch, length, power, or other voice parameters by signal processing based on the change amount parameter specified in step S4 to generate a new prosody changed voice waveform; The phoneme label information indicating the utterance content and its position, or the power information, the pitch information, the pitch mark information, or other voice information added to the real voice waveform is generated in the generation of the prosody modified voice waveform. A voice information changing unit 5 for generating new voice information by changing the amount corresponding to the amount of change; In addition to voice information, voice with a speech waveform database creation unit 8 for generating a speech waveform database by using the the real voice one or more types of prosody Changed speech waveform and the voice information is derived from the speech waveform A waveform database construction device was constructed.

  A third aspect of the present invention is a speech waveform database construction program for instructing a computer to execute the steps described in the first aspect in this order.

In the course of constructing a speech waveform database used in a waveform connected text speech synthesizer,
A stage in which a voice waveform for constructing a voice waveform database, a phonetic label information indicating the utterance content and its position, power information, pitch information, pitch mark information, or other voice information is prepared in the voice waveform. so,
For all real voice waveforms in the voice waveform database, the pitch or length or other voice parameters are changed by signal processing based on the change amount parameter specified by the voice waveform change unit, and the prosody has been changed. Generate a new waveform,
The phonetic label information for indicating the utterance content and its position, the power information, the pitch information or other voice information given to the real voice utterance in the voice waveform database To generate new audio information by changing in accordance with the amount of change made in
A configuration in which a speech waveform database is constructed by a speech waveform database generation unit using the above-described real voice waveform and its speech information, and a speech waveform whose prosody has been changed and its speech information is adopted. As a result, speech waveforms having various prosodic variations by signal processing are included in the speech waveform database in advance, and an appropriate pitch, length, power or other speech corresponding to the text to be synthesized is included in the speech waveform database. There is a high possibility that a parameter speech waveform exists, and high-quality speech synthesis with natural intonation can be performed.

In the signal processing for the real voice waveform performed by the voice waveform changing unit, the prosody change is performed in a range in which the sound quality does not deteriorate, so that high quality speech synthesis without sound quality deterioration can be performed.
In addition, since the signal processing for speech waveforms is performed at the stage of database construction, a signal processing technique that requires little computational complexity but with little deterioration in sound quality can be used, so the sound quality can be increased without increasing the time required for the synthesized speech generation processing. High quality speech synthesis with natural intonation without degradation can be achieved.

The best mode for carrying out the invention will be described with reference to the embodiment of FIG.
Reference numeral 1 denotes an audio waveform database construction device, which is composed of an audio waveform change unit 4, an audio information change unit 5, changed audio waveform data 6, changed audio information 7, and an audio waveform database generation unit 8. Yes.
The speech waveform database construction apparatus 1 constructs a speech waveform database 9 by inputting the real voice waveform data 2 and its phoneme label information, power information, pitch information, pitch mark information or other real voice information 3. This is a speech waveform database construction device. The speech waveform database construction device 1 is also composed of a computer device composed of well-known hardware and a computer program that is software for driving the CPU. The speech waveform change unit 4, the speech information change unit 5, Each of the speech waveform database generation units 8 is configured by one or both of hardware and software.

The real voice waveform data 2 is, for example, voice data obtained by reading a word or a sentence by the voice of a narrator, AD-converted, and stored in a storage area of a storage medium such as a hard disk.
The real voice information 3 corresponds to the real voice waveform data 2 and is data to which phonetic label information, power information, pitch information, pitch mark information, spectrum information and other voice information indicating utterance contents and time information are added. These are stored in a storage area of a storage medium such as a hard disk.

The speech waveform changing unit 4 performs signal processing on all the real voice waveform data stored in the real voice waveform data 2 on the basis of the change amount parameter given to the voice waveform database construction device 1 and corresponding sound. Changed to high and length, and the changed speech waveform data 6 in which the specified change amount is added to the pitch and length is generated. As a signal processing method used here, for example, there is a signal processing method represented by PSOLA (Pitch Sincronous Overlap Add). The changed speech waveform data 6 is stored in a storage area of a storage medium such as a hard disk.
The voice information changing unit 5 changes the information of the real voice information 3 according to the pitch and length of the voice waveform data changed by the voice waveform changing unit 4 and generates changed voice information 7. For example, when the pitch of the voice waveform data changes, the pitch information, pitch mark information, and spectrum information are changed accordingly. Further, when the length of the speech waveform data changes, the time information and the pitch mark information of the phoneme label are changed correspondingly. The changed audio information 7 is stored in a storage area of a storage medium such as a hard disk.

Based on the data stored in the voice waveform data 2, the real voice information 3, the changed voice waveform data 6, and the changed voice information 7, the voice waveform database generation unit 8 generates the voice waveform data 9 and the voice information index 10. A speech waveform database 11 is generated.
The voice waveform data 9 stores the voice waveform which is the content of the real voice waveform data 2 or the changed voice waveform data 6 and is stored in a storage area of a storage medium such as a hard disk.
The voice information index 10 holds the contents of the real voice information 3 or the changed voice information 7 as an index, and uses the desired phoneme information and pitch information as search keys, and the voice waveform data necessary for voice synthesis from the voice waveform data 9. Can be read.

Here, the processing operation will be described with reference to FIG. FIG. 2 is a flowchart showing the processing operation of voice waveform database construction.
First, (S1) is a signal for all the real voice waveform data stored in the real voice waveform data 2 in the voice waveform change unit 4 based on the change amount parameter given to the voice waveform database construction device 1. Apply the processing to change the pitch and length. This amount of change needs to be within a range where sound quality degradation does not occur. Since this allowable amount also depends on the signal processing method, it cannot be defined as one type, and should be determined from experimental values according to the signal processing method. Suppose that the pitch is given by 10%. In this case, for all the real voice waveform data stored in the real voice waveform data 2, a voice waveform whose pitch is uniformly increased by 10% is generated by signal processing. As a signal processing method used here, for example, there is a method represented by PSOLA. In the case of a system that requires pitch mark information during signal processing such as PSOLA, the pitch mark information stored in the real voice information 3 is read and used. The changed voice waveform data is temporarily stored as changed voice waveform data 6.

  Subsequently, in (S2), in the voice information changing unit 5, the information of the real voice information 3 is changed in accordance with the pitch and length of the voice waveform data changed by the voice waveform changing unit 4. In the case of this embodiment, in order to increase the pitch by 10%, the pitch information is changed to a value increased by 10%. Further, since the pitch mark position changes when the pitch changes, the pitch mark position is changed according to the changed pitch mark position. For the pitch mark position after the change, the pitch mark position information used at the time of signal processing of the voice waveform changing unit 4 may be used. The changed audio information data is temporarily stored as changed audio information 7.

  Next, in (S3), the voice waveform database generation unit 8 constructs the voice waveform database 11 using the real voice waveform data 2, the changed voice waveform data 6, the real voice information 3, and the changed voice information 7. . The contents of the real voice information 3 or the changed voice information 7 are held as the voice information index 10 and the pointer information to the corresponding voice waveform data is registered, so that the voice synthesis engine uses the voice waveform database 11 to synthesize the voice. Is performed, the speech waveform data 9 is searched for desired phoneme information, pitch information, etc., and the speech waveform data necessary for synthesis can be read out. According to this embodiment, the voice waveform database 11 to be finally constructed includes, for the voice waveform data, the real voice waveform data 2 and the changed voice data 6 to which the pitch is increased by 10%. As a result, the size of the speech waveform database is doubled.

As described above, according to the present invention, since signal processing based on the change amount parameter is performed in advance, the variation of the speech waveform held in the speech waveform database increases. Therefore, it is desired to synthesize the speech waveform database. There is a high possibility that a speech waveform having an appropriate pitch or length corresponding to the text exists, and high-quality speech synthesis with natural intonation can be performed.
In the signal processing for the real voice waveform performed by the voice waveform changing unit, since the prosody change is performed within a range in which the sound quality does not deteriorate, the voice waveform obtained by performing signal processing on the real voice waveform is used. In spite of this, it is possible to synthesize high-quality speech without deterioration of sound quality.

In addition, since the signal processing for the speech waveform is performed at the stage of database construction, it is possible to use a signal processing method that requires a small amount of computation with little deterioration in sound quality, and without increasing the time required for the synthesized speech generation process, This has the effect of enabling high-quality speech synthesis with natural intonation without sound quality degradation.
In addition, the Example mentioned above can be implemented with the following aspects.
For example, the change amount specified in the change amount parameter does not need to be one type. For example, the pitch change amount is specified by changing four types of 5% up, 10% up, 5% down, 10% down. Four types of the completed speech waveform data 6 and the modified speech information 7 may be generated according to the change amount parameter, and the speech waveform database 11 based on these may be constructed. Moreover, you may specify combining the change of a pitch, and the change of length. In these cases, since there are more pitch variations and pitch variations for a certain phoneme string, the pitch connection between speech segments becomes smoother, and the naturalness of the duration of phonemes. Therefore, the quality of the synthesized speech is further improved. The present invention can also be implemented using a signal processing method other than the previous PSOLA.

The figure explaining an Example. The figure which shows the operation | movement flowchart of an Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Voice waveform database construction apparatus 2 Real voice voice waveform data 3 Real voice voice information 4 Voice waveform change part 5 Voice information change part 6 Changed voice waveform data 7 Changed voice information 8 Voice waveform database generation part
9 Voice waveform data 10 Voice information index 11 Voice waveform database

Claims (3)

In a speech waveform database construction method for constructing a speech waveform database used in a waveform connected text speech synthesizer,
Based on the amount of change parameter specified within the range where sound quality deterioration of the voice waveform does not occur with respect to the prepared real voice waveform, the pitch, length, power, and other voice parameters are processed by signal processing. Change and execute the step of generating a new prosody modified speech waveform,
In the step of generating the prosody modified speech waveform, the phoneme label information indicating the utterance content and its position, the power information, the pitch information, the pitch mark information, or other speech information given to the real voice waveform Performing a step of generating new audio information by changing in accordance with the amount of change made,
In addition to the real voice speech waveform and the voice information, performing the step of generating a speech waveform database by using the the real voice one or more types of prosody Changed speech waveform and the voice information is derived from the speech waveform,
A method of constructing a speech waveform database. In a speech waveform database construction device for constructing a speech waveform database used in a waveform connected text speech synthesizer,
Based on the amount of change parameter specified within the range where sound quality deterioration of the voice waveform does not occur with respect to the prepared real voice waveform, the pitch, length, power, and other voice parameters are processed by signal processing. A voice waveform changing unit that changes and generates a new prosody changed voice waveform;
Corresponding to the amount of change made in the generation of the prosody modified speech waveform, the phoneme label information indicating the utterance content and its position, or the power information, pitch information, or pitch mark information given to the real voice waveform A voice information changing unit that generates new voice information by changing
In addition to the real voice speech waveform and its audio information, and the audio waveform database generation unit for generating a speech waveform database by using the the real voice speech waveform and the voice information of one or more types of prosody Modified which was derived from the speech waveform ,
A speech waveform database construction device characterized by comprising:   A speech waveform database construction program characterized in that a command for executing the steps described in claim 1 in this order is given to a computer.

Images