Self-supervised ASR-TTS models suffer in out-of-domain data conditions. Here we propose an enhanced ASR-TTS (EAT) model that incorporates two main features: 1) The ASR→TTS direction is equipped with a language model reward to penalize the ASR hypotheses before forwarding it to TTS. 2) In the TTS→ASR direction, a hyper-parameter is introduced to scale the attention context from synthesized speech before sending it to ASR to handle out-of-domain data. Training strategies and the effectiveness of the EAT model are explored under out-of-domain data conditions. The results show that EAT reduces the performance gap between supervised and self-supervised training significantly by absolute 2.6\% and 2.7\% on Librispeech and BABEL respectively.
@article{arxiv.2104.07474,
title = {EAT: Enhanced ASR-TTS for Self-supervised Speech Recognition},
author = {Murali Karthick Baskar and Lukáš Burget and Shinji Watanabe and Ramon Fernandez Astudillo and Jan "Honza'' Černocký},
journal= {arXiv preprint arXiv:2104.07474},
year = {2021}
}