JPH11282484A - Voice synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform natural and smooth voice synthesis without enlarging the storage capacity of a phoneme unit waveform storage part so much. SOLUTION: Text data inputted from a text data input part 1 are sent to a phonetic symbol generation part 2 and phonetic symbols are generated there. A phoneme unit waveform calling part 3 calls data pertinent to first phoneme unit waveform data A0 from the phoneme unit waveform storage part 4 based on the generated phonetic symbol. For the text data inputted from the text data input part 1, a phonetic rule is analyzed by a phonetic rule part 6 and a phonetic parameter generation part 7 generates a phonetic parameter based on the analysis. Then, a pitch pattern generation part 8 generates a pitch pattern based on the phonetic parameter. The phoneme waveform data string generation means 12 of a waveform data connection part 5A receives a parameter numerical value from a voice quality parameter storage part 11 based on information from the pitch pattern generation part 8 and generates the phoneme unit waveform data A1 to follow next based on it and the data A0 from the calling part 3.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for performing speech synthesis based on a rule synthesis method.

[0002]

2. Description of the Related Art FIG. 3 is a block diagram showing a configuration of a conventional speech synthesizer. In the figure, a text data input unit 1 converts input text data into a phoneme symbol generation unit 2.
The phoneme symbol generation unit 2 generates a phoneme symbol corresponding to the text data. The phoneme unit waveform calling unit 3 includes a phoneme symbol generation unit 2
The phoneme unit waveform data corresponding to the generated phoneme symbol is called from the phoneme unit waveform storage unit 4. The waveform data connection unit 5 connects the phoneme unit waveform data sequentially called by the phoneme unit waveform calling unit 3 to generate a phoneme waveform data sequence.

FIG. 4 is a waveform diagram showing an example of the phoneme waveform data sequence generated in this manner. The N phoneme unit waveform data A0, A1,... , An, An + 1,..., AN-1 are connected to form one phoneme waveform data string.

On the other hand, the text data input unit 1 also outputs the input text data to the prosody rule unit 6, and the prosody rule unit 6 analyzes the prosody rules in the text data. . Prosody parameter generation unit 7
Generates prosody parameters based on the prosody rules analyzed by the prosody rule unit 6. The pitch pattern generation unit 8 generates a pitch pattern based on the prosody parameters generated by the prosody parameter generation unit 7.

[0005] The waveform synthesizing section 9 synthesizes a voice waveform based on the input of the phoneme waveform data string generated by the waveform data connecting section 5 and the pitch pattern generated by the pitch pattern generating section 8. The audio output unit 10 outputs audio based on the audio waveform synthesized by the waveform synthesizing unit 9.

[0006]

By the way, in order to synthesize speech as naturally and smoothly as possible, phoneme unit data corresponding to all phoneme data is stored in the phoneme unit waveform storage unit 4, and from among these, It is necessary to call the optimal phoneme unit data. However, in order to store the phoneme unit data corresponding to all phoneme data, the storage capacity of the phoneme unit waveform storage unit 4 becomes very large, and the cost becomes high.

On the other hand, if the storage capacity of the phoneme unit waveform storage unit 4 is limited to a certain value or less in order to avoid such a high cost, a sufficient amount of phoneme unit data cannot be stored and unnaturalness occurs. As a result, an unnatural voice synthesis is performed.

The present invention has been made in view of the above circumstances, and provides a speech synthesis apparatus capable of performing natural and smooth speech synthesis without significantly increasing the storage capacity of a phoneme unit waveform storage unit. It is intended to be.

[0009]

As means for solving the above-mentioned problems, the invention according to claim 1 comprises a phoneme symbol generation section for generating a phoneme symbol from input text data, and a phoneme corresponding to the phoneme symbol. A phoneme unit waveform storage unit that stores unit waveform data, a waveform data connection unit that connects each phoneme unit waveform data sequentially called from the phoneme unit waveform storage unit to generate a phoneme waveform data sequence, and a pitch from the text data. A voice pattern synthesizer for synthesizing a phoneme waveform data string output from the waveform data connection unit based on the pitch pattern generated by the pitch pattern generation unit, comprising: The waveform data connection unit stores a voice parameter value for a parameter set in advance for voice quality. Unit, parameter values selected from among those stored in the voice quality parameter storage unit according to the pitch pattern generated by the pitch pattern generation unit, and the previously called phoneme unit waveform data is input, and then Phoneme waveform data generation means for generating phoneme unit waveform data to be continued, and connection means for sequentially connecting and outputting the phoneme unit waveform data generated by the phoneme waveform data string generation means, It is characterized by the following.

According to a second aspect of the present invention, in the first aspect of the present invention, the phoneme unit data string generating means is stored in the voice quality parameter storing section in accordance with the pitch pattern generated by the pitch pattern generating section. It is configured by a hierarchical neural network that receives a parameter value selected from among the parameters and the previously called phoneme unit waveform data, and outputs the next phoneme unit waveform data to be continued.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. However, the same components as those shown in FIG. 3 are denoted by the same reference numerals, and redundant description will be omitted. FIG. 1 is a block diagram showing the configuration of the embodiment of the present invention. The main difference between FIG. 1 and FIG.
A is that it has a voice quality parameter storage unit 11, a phoneme waveform data string generation unit 12, and a connection unit 13. However, the phoneme unit waveform storage unit 4 in FIG. 3 has a very large storage capacity for storing phoneme unit data corresponding to all phoneme data, but the phoneme unit waveform storage unit 4 in FIG. Since it is not necessary to store phoneme unit data corresponding to all phoneme data, the storage capacity is small.

The voice quality parameter storage unit 11 stores numerical values preset for predetermined parameters for determining voice quality. The predetermined parameter is, for example, “accent data”, “pitch of voice”, “utterance speed”, “sex (many types of men and women are prepared)” and the like.

The phoneme waveform data string generating means 12 determines the numerical values of the parameters input from the voice quality parameter storage section 11 based on the pitch pattern information generated by the pitch pattern generating section 8, and further generates a phoneme unit waveform calling section 3.
Inputs the value of the previously called phoneme unit waveform data.
Then, the phoneme waveform data sequence generating means 12 is adapted to generate the next successive phoneme unit waveform data based on these inputs. The connection means 13 sequentially connects and outputs the phoneme unit waveform data generated by the phoneme waveform data string generation means 12.

FIG. 2 is a block diagram of the phoneme waveform data string generating means 12. As shown in FIG. 2, the phoneme waveform data string generating means 12 is constituted by a neural network. This neural network has an input layer, an intermediate layer, and an output layer. The input layer is composed of q neurons R1, R2, ..., for inputting phoneme unit waveform data.
Rq and m neurons S1, S2,..., Sm for inputting voice quality parameters. The intermediate layer is composed of q neurons M1, M2, ..., Mq, and the output layer is composed of q neurons V1, V2, ..., Vq for outputting phoneme unit waveform data. The neural network has a learning function (back propagation method) based on the input and output as described above.

The neurons R1, R2,..., Rq in the input layer
The phoneme unit waveform data to be input to, for example, is an amplitude value of the phoneme unit waveform data A0 in FIG. 4 for each predetermined cycle (period obtained by dividing the entire time of A0 by q). For inputting voice quality parameters, one or more neurons are used for one input data. In this case, the required number of neurons may be prepared by converting the input data into a binary number, and “0” or “1” may be input thereto.

Next, the operation of the present embodiment configured as described above will be described. The text data input from the text data input unit 1 is sent to the phoneme symbol generation unit 2,
The phoneme symbol generation unit 2 generates a phoneme symbol corresponding to the text data. The phoneme unit waveform calling unit 3 calls data corresponding to the first phoneme unit waveform data A0 from the phoneme unit waveform storage unit 4 based on the generated phoneme symbols.

On the other hand, the text data input from the text data input unit 1 is also sent to the prosody rule unit 6 to analyze the prosody rules in the text data, and the prosody parameter generation unit 7 generates a prosody parameter based on the analysis. Generate Then, the pitch pattern generation unit 8 generates a pitch pattern based on the generated prosody parameters.
The numerical values of the phoneme unit waveform data A0 called by the phoneme unit waveform calling unit 3 from the phoneme unit waveform storage unit 4 are stored in the neuron R1, R2,.
.., Rq. Also, a pitch pattern generation unit 8
Are also sent to the phoneme waveform data string generating means 12, and the phoneme waveform data string generating means 12 inputs the voice quality parameter values corresponding to the pitch pattern to the neurons S1, S2,..., Sm.

The neuron R1, R2,..., Rq of the phoneme waveform data generation means 12 and the neurons S1, S2,.
.., Vq, a predetermined calculation is performed, and a numerical value for the next phoneme unit waveform data A1 is output from the neurons V1, V2,. Therefore, in FIG. 1, the phoneme waveform data string generating means 12 outputs the phoneme unit waveform data A0 corresponding to this numerical value.
From the phoneme unit waveform storage unit 4 to generate phoneme unit waveform data A1.

Next, the neuron R1, R2,..., Rq of the phoneme waveform data sequence generating means 12 is supplied with a numerical value for A1 instead of the phoneme unit waveform data A0, and the neurons V1, V2,. Output a numerical value for the next phoneme unit waveform data A2. And
The phoneme waveform data sequence generating means 12 similarly generates phoneme waveform data A2. Phoneme waveform data string generation means 12
Is performed in the same manner as described above, and the phoneme waveform data strings A3, A4,.
An, An + 1,..., AN-1 are sequentially generated.

In the connection means 13, the phoneme waveform data string A generated by the phoneme waveform data string
, An, An + 1,..., AN-1 are connected to each other and output to the waveform synthesizer 9. Based on the pitch pattern from the pitch pattern generating unit 8, the waveform synthesizing unit 9 sequentially transmits A0, A1,.
, An, An + 1,..., AN-1, etc., are used to synthesize a speech waveform. Then, the audio output unit 10
Performs audio output based on the audio waveform synthesized by the waveform synthesizer 9.

As described above, in the configuration of FIG. 1, the phoneme waveform data string generation means 12
, And the phoneme unit waveform data to be continued next is generated based on the input of the phoneme unit waveform data called last time. Therefore, it is possible to perform natural and smooth speech synthesis, and it is not necessary to store phoneme unit waveform data corresponding to all phoneme data in the phoneme unit waveform storage unit 4 unlike the related art. The storage capacity of the phoneme unit waveform storage unit 4 can be reduced.

In the above embodiment, the values of the previously called phoneme unit waveform data are input to the neurons R1, R2,..., Rq in the input layer of the phoneme waveform data sequence generating means 12 shown in FIG. Was supposed to be
The number of neurons may be further increased, and the numerical values of all the previously called waveform data may be input to generate the next phoneme unit waveform data to be continued. According to this, it becomes possible to call up more appropriate phoneme unit waveform data.

Further, in the above-described embodiment, an example is shown in which the phoneme waveform data string generating means 12 is constituted by a neural network. However, the present invention is not limited to the neural network only. For example, a circuit having another function such as a fuzzy arithmetic circuit Can also be used.

[0024]

As described above, according to the present invention, based on the voice quality parameter and the previously called phoneme unit waveform data, the phoneme unit waveform data to be succeeded next is generated. Natural and smooth speech synthesis can be performed without increasing the storage capacity of the waveform storage unit so much.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a neural network configuration of a phoneme waveform data string generation unit 12 in FIG.

FIG. 3 is a configuration diagram of a conventional example.

FIG. 4 is a waveform data connection unit 5, 5 in FIG. 1 or FIG.
FIG. 4 is a waveform chart showing an example of a phoneme waveform data string generated by A.

[Explanation of symbols]

 Reference Signs List 1 Text data input unit 2 Phoneme symbol generation unit 3 Phoneme unit waveform calling unit 4 Phoneme unit waveform storage unit 5, 5A waveform data connection unit 6 Prosody rule unit 7 Prosody parameter generation unit 8 Pitch pattern generation unit 9 Waveform synthesis unit 10 Voice output Unit 11 voice quality parameter storage unit 12 phoneme waveform data string generation means 13 connection means

Claims (2)

[Claims] 1. A phoneme symbol generation unit for generating a phoneme symbol from input text data, a phoneme unit waveform storage unit for storing phoneme unit waveform data corresponding to the phoneme symbol, and sequentially called from the phoneme unit waveform storage unit. And a pitch pattern generation unit that generates a phoneme waveform data string by connecting each of the phoneme unit waveform data, and a pitch pattern generation unit that generates a pitch pattern from the text data, and is output from the waveform data connection unit. A speech synthesizer for synthesizing a phoneme waveform data sequence based on a pitch pattern generated by the pitch pattern generation unit, wherein the waveform data connection unit stores a parameter value for a preset item for voice quality. A parameter storage unit; A phoneme waveform data string generating means for receiving a parameter numerical value selected from those stored in the voice quality parameter storage unit and the previously called phoneme unit waveform data and generating phoneme unit waveform data to be continued next; And a connection means for sequentially connecting and outputting phoneme unit waveform data generated by the phoneme waveform data string generation means, and a speech synthesizing apparatus. 2. The phoneme unit data string generation means according to a pitch pattern generated by the pitch pattern generation section, a parameter numerical value selected from the voice quality parameter storage section and a previously called phoneme. The speech synthesis apparatus according to claim 1, wherein the speech synthesis apparatus is configured by a hierarchical neural network that receives unit waveform data and outputs phoneme unit waveform data to be continued next.