Accent conversion using discrete units with parallel data synthesized from controllable accented TTS
Tuan Nam Nguyen, Ngoc Quan Pham, Alexander Waibel
TL;DR
The paper tackles accent conversion without reference utterances, aiming to transform non-native speech into native accents while preserving content and speaker identity. It introduces a discrete-unit framework where self-supervised native speech representations are quantized into units that a pronunciation corrector maps from accented speech, followed by a multi-speaker unit-to-speech model that reconstructs native speech conditioned on speaker identity. A data-augmentation pipeline using controllable multi-accent TTS (SYNTACC) and native TTS (YourTTS) generates large parallel data to train the pronunciation corrector in low-resource settings. Experimental results show improved accentedness and fluency with good speaker similarity, demonstrating the approach's practicality for cross-accent communication and potential benefits for downstream ASR/translation.
Abstract
The goal of accent conversion (AC) is to convert speech accents while preserving content and speaker identity. Previous methods either required reference utterances during inference, did not preserve speaker identity well, or used one-to-one systems that could only be trained for each non-native accent. This paper presents a promising AC model that can convert many accents into native to overcome these issues. Our approach utilizes discrete units, derived from clustering self-supervised representations of native speech, as an intermediary target for accent conversion. Leveraging multi-speaker text-to-speech synthesis, it transforms these discrete representations back into native speech while retaining the speaker identity. Additionally, we develop an efficient data augmentation method to train the system without demanding a lot of non-native resources. Our system is proved to improve non-native speaker fluency, sound like a native accent, and preserve original speaker identity well.