Showing 1–1 of 1 results for author: Madotto, A
-
Learning Fast Adaptation on Cross-Accented Speech Recognition
Authors:
Genta Indra Winata,
Samuel Cahyawijaya,
Zihan Liu,
Zhaojiang Lin,
Andrea Madotto,
Peng Xu,
Pascale Fung
Abstract:
Local dialects influence people to pronounce words of the same language differently from each other. The great variability and complex characteristics of accents creates a major challenge for training a robust and accent-agnostic automatic speech recognition (ASR) system. In this paper, we introduce a cross-accented English speech recognition task as a benchmark for measuring the ability of the mo…
▽ More
Local dialects influence people to pronounce words of the same language differently from each other. The great variability and complex characteristics of accents creates a major challenge for training a robust and accent-agnostic automatic speech recognition (ASR) system. In this paper, we introduce a cross-accented English speech recognition task as a benchmark for measuring the ability of the model to adapt to unseen accents using the existing CommonVoice corpus. We also propose an accent-agnostic approach that extends the model-agnostic meta-learning (MAML) algorithm for fast adaptation to unseen accents. Our approach significantly outperforms joint training in both zero-shot, few-shot, and all-shot in the mixed-region and cross-region settings in terms of word error rate.
△ Less
Submitted 4 March, 2020;
originally announced March 2020.
