End-to-End Myanmar Speech Recognition with Human-Machine Cooperation

End-to-end automatic speech recognition based on deep neural networks has achieved satisfactory results in some languages with large-scale trainable data. Due to the small scale of available training data, there is still a significant gap between Myanmar speech recognition and application needs. Based on the end-to-end model of Transformer, this paper studies the influence of character, byte pair encoding (BPE), and syllable label encoding on recognition rate. For the Myanmar speech recognition model in a low-resource environment, we propose and implement two training methods: (1) The self-supervised pre-training speech representations model based on massive English data is introduced into the Myanmar speech recognition model as an acoustic feature extractor. The experimental results show that the pre-training speech representations method has an obvious effect on end-to-end Myanmar speech recognition. (2) Expand the scale of the database using human-machine cooperation, to improve the speech recognition rate. Based on the method (1), through a human-machine cooperation experiment, 2898 new corpora were obtained after 6 iterations. The character error rate of the Myanmar speech recognition model on the development set decreased from 4.8% to 4.1%, and the character error rate of the test set decreased from 5.5% to 4.2%. The time required for manual proofreading of each batch of the speech corpus has decreased from 19.1 hours to 16.3 hours. Experiments have proved the effectiveness of the human-machine cooperation method.