Improving Noisy Student Training on Non-target Domain Data for Automatic Speech Recognition
This work addresses a domain-specific challenge in ASR by incrementally enhancing data filtering for noisy student training.
The paper tackles the problem of improving Noisy Student Training for automatic speech recognition on non-target domain data by proposing a data selection strategy called LM Filter, which uses CER differences between hypotheses with and without a language model as a filter threshold, resulting in a 10.4% improvement over baselines and achieving 3.31% CER on the AISHELL-1 test set.
Noisy Student Training (NST) has recently demonstrated extremely strong performance in Automatic Speech Recognition(ASR). In this paper, we propose a data selection strategy named LM Filter to improve the performance of NST on non-target domain data in ASR tasks. Hypotheses with and without a Language Model are generated and the CER differences between them are utilized as a filter threshold. Results reveal that significant improvements of 10.4% compared with no data filtering baselines. We can achieve 3.31% CER in AISHELL-1 test set, which is best result from our knowledge without any other supervised data. We also perform evaluations on the supervised 1000 hour AISHELL-2 dataset and competitive results of 4.73% CER can be achieved.