System and method for identifying accent of input sound

Abstract

Disclosed are a method and a system of input speech stress recognition, with a training vocabulary, training speech data and training boundary data stored. Hereafter, after the candidates of boundary data are input, the first probabilities are calculated based on the input vocabulary, speech and boundary data. The first probabilities are the corresponding probabilities between the words of input text, each prosodic word boundary and each boundary candidate. In addition, the second probabilities also can be calculated based on the input vocabulary, speech and boundary data when the input speech has boundary of the prosodic word assigned by one of the input boundary data candidates. The second probabilities are the corresponding probabilities between the word speech of input text and input speech data. Thereafter, the optimized the boundary candidate is searched out as the output result by maximizing the products of the first and the second probabilities.

Description

Technical field [0001] The present invention relates to speech recognition technology. Specifically, the present invention relates to a technique for recognizing the accent of an input voice. Background technique [0002] In recent years, attention has been paid to speech synthesis for reading the input text using natural pronunciation without requiring accompanying information such as the reading of the text. In this speech synthesis technology, in order to produce a natural-sounding sound to the listener, it is important not only to accurately reproduce the pronunciation of a word, but also to accurately reproduce its accent. If speech can be synthesized by accurately reproducing the higher H type or lower L type pronunciation of each mora that constitutes a word, it is possible to make the resulting speech sound natural to the listener. [0003] Most of the currently used speech synthesis systems are systems constructed through statistical training of the systems. In order to ...

AI Technical Summary

Problems solved by technology

However, because accents are relative in nature, it is difficult to generate training data from data such as phonetic deviations

In fact, although automatic recognition of accents based on such speech data has been attempted (see Kikuo Emoto, Heiga Zen, Keiichi Tokuda, and Tadashi Kitamura "Accent Type Recognition for Automatic Prosodic Labeling," Proc. of Autumn Meeting of the Acoustical Society of Japan ( September, 2003) (Kikuo Emoto, Heiga Zen, Keiichi Tokuda, and Tadashi Kitamura, "Speech Type Recognition for Automatic Prosodic Labeling", Fall Meeting of the Acoustical Society of Japan (September 2003))), but the accuracy is not satisfactory enough to make The identification is put into practice

Images