Disentanglement of Speaker Identity for Accented Speech Recognition

Automatic speech recognition (ASR) systems have achieved good performance on standard English speech. However, recognition accuracy significantly degrades for non-native speakers with accented speech. Existing multi-accent ASR models utilize accent embeddings as extra input to handle diverse accents, but these embeddings often intertwine other speech attributes, impeding model performance. In this paper, we propose a speaker identity disentangled accent modeling approach, which extracts the accent embedding without any speaker-specific information, and uses it to improve the performance of the multi-accent speech recognition systems. Experiments conducted on the CommonVoice dataset demonstrate that our proposed method attains a 15% reduction in word error rate (WER) over baselines.