A comparison of multi-style DNN-based TTS approaches using small datasets

摘要

Studies have shown that people already perceive the interaction with computers, robots and media in the same way as they perceive social communication with other people. For that reason it is critical for a high-quality text-to-speech system (TTS) to sound as human-like as possible. However, a major obstacle in creating expressive TTS voices is that the amount of style-specific speech needed for training such a system is often not sufficient. This paper presents a comparison between different approaches to multi-style TTS, with focus on cases when only a small dataset per style is available. The described approaches have been originally proposed for efficient modelling of multiple speakers with a limited amount of data per speaker. Among the suggested approaches the approach based on style codes has emerged as the best, regardless of the target speech style.