SPEECH SYNTHESIS DEVICE AND PROGRAM

Japanese Patent JP2023045835

A

To cause a model for synthesizing a speech of high quality to learn using smaller-scale training data.SOLUTION: A speech synthesis device is configured to: analyze, using a first model, peripheral context information, symbol by symbol, based upon text data representing a symbol string having a plurality of symbols; generate a noise component having a predetermined probability distribution; determine output information, using a second model, based upon input information associated with information on one of the peripheral context information and frequency characteristics of a speech of a preceding frame; compute, using a third model, a level of the noise component based upon the input information; generate noise-added output information having the noise component added to the output information; estimate, using a fourth model, frequency characteristics of a speech of a current frame based upon the noise-added output information, and information on the other of the peripheral context information and the frequency characteristics of the speech of the preceding frame; and update parameters of the first model, second model, third model, and fourth model so as to decrease a loss function including an estimated error of the frequency characteristics of the speech of the current frame and passage information amount loss of the noise-added output information.SELECTED DRAWING: Figure 1

KUMANO TADASHI
SEIYAMA NOBUMASA
SAITO REIKO

JP2021154428A

April 03, 2023

September 22, 2021

Click for automatic bibliography generation  

JAPAN BROADCASTING CORP

G10L13/06

Shu Oikawa
Naoyuki Takada
Hiroyuki Matsumoto
Ikuichiro Kinoshita

Previous Patent: LAMINATED BODY, ELECTRON SOURCE AND ELECTRONIC DEVICE COMPRISING LAMINATED BODY, AND METHOD FOR MANU...

Next Patent: stirrer