TTS-Guided Training for Accent Conversion Without Parallel Data

Accent Conversion (AC) seeks to change the accent of speech from one (source) another (target) while preserving content and speaker identity. However, many existing AC approaches rely on source-target parallel data during training or reference at run-time. We propose a novel conversion framework without need for either speech. Specifically, text-to-speech (TTS) system is first pretrained with target-accented data. This TTS model its hidden representations are expected be associated only target accent. Then, encoder trained convert under supervision model. In doing so, source-accented corresponding transcription forwarded TTS, respectively. The output optimized same as text embedding in system. At run-time, combined decoder toward target. experiments, we converted English two source accents (Chinese/Indian) (American/British/Canadian). Both objective metrics subjective listening tests successfully validate that proposed approach generates samples close high quality.