Work

構音障害者のための複数ドメインのデータに基づく機械学習を用いた支援技術

Year: 2020
Type: dissertation
Abstract: The person with an articulation disorder has problems forming speech sounds properly.Although the voice is one of the most natural communication tools for human beings, the voices of persons with articulation disorders are difficult to understand for people around them.To improve the quality of their life, there is a critical need for voice-driven assistive systems.In the field of speech processing, numerous researches have been proposed and remarkable progress has been made in fields with access to a large amount of training data.However, there has been very little benefit for persons with speech disorders because these systems are trained on "typical" speech.To achieve a system that works well for disordered speech, in this paper, we tackle the following two problems: limited data availability and difference in a speech style.First, it is difficult to collect a sufficient amount of speech data to train the model owing to a physical burden.Second, their speech style differs significantly from that of physically-unimpaired persons.Therefore, an approach specifically tailored to overcome these problems of impaired speech data is needed.To solve these problems, this paper proposes the following two approaches: conversion and recognition.Voice conversion (VC) is a technique for converting paralinguistic information in speech, while preserving the linguistic information in the utterance.Although the most popular VC application is speaker conversion, VC also can be applied for emotion conversion and assistive technology.
Institution: Kobe University
Language: English
Open Access status: green
OpenAlex ID: W7208045339
Cites: 122
Cited by:
Related to: 10