Emotional speech from machine

Abstract

Emotional speech is the expressiveness in speech that is transmitted through changes in pitch, loudness, timbre, speech rate and pauses that convey emotion. Although the current TTS technology is capable of converting a given text into speech, they sound monotonous and lack emotion and naturalness. In order to improve artificial voices, application of emotion is highly evaluated. In this thesis, we will be creating a system that makes use of speech mark-up language to produce emotion in speech by analysing the tone of given text. For this purpose, we combine IBM tone analyser with TTS that accepts the speech mark-up language. In this research, we perform empirical study on two experimental implementation using two TTS and two speech mark-up language. The first combination involves IBM TTS and SSML and the second combination includes MARY TTS and EmotionML. The mark-ups are predefined in EmotionML for four major emotions namely anger, fear, joy and sadness and for SSML prosody value from previous study is used. Therefore, this study describes the two implementations and evaluate their output emotional speech synthesis which is then compares with human voice to define its perfection.