Synthesis of everyday conversational speech based on fine-tuning with a corpus for speech synthesis

Abstract

In this letter, we propose a separate modeling of prosodic and segmental features for everyday conversational speech synthesis, addressing challenges posed by low-quality recordings in the Corpus of Everyday Japanese Conversation (CEJC). Initially, the FastSpeech 2 model is trained on the conversation corpus and subsequently fine-tuned on a corpus for speech synthesis. Experimental results show that this fine-tuning approach enhances synthesis quality while preserving the nuances of everyday conversations.