Introduction

---

Here we present synthesis samples generated by our proposed system and those generated by our implementation of the paper "Learning to Speak Fluently in a Foreign Language: Multilingual Speech Synthesis and Cross-Language Voice Cloning". However, we have made several changes to the original TTS model: 1. We have removed the residual encoder. 2. We have replaced the hybrid attention module with the forward attention with transition agent for higher stability. 3. The original is not appropriate for code-switching audio samples from the DB4 dataset. So we have made a simple adaptation that assigns the code-switching audio samples with a new language accent code '2', while '0' is for English and '1' is for Chinese. The model from our implementation of the paper "Learning to Speak Fluently in a Foreign Language: Multilingual Speech Synthesis and Cross-Language Voice Cloning" is notated with 'our implementation of CLMS/BLMS in [1]', and the following samples from that model are all synthesized with respect to the language of the text. In other words, '0' for English text, '1' for Chinese text, and '2' for code-switching text.

Original Voices

English Sentense Synthesis

Mandarin Sentense Synthesis

Code-switching Sentense Synthesis

Reference