Based on Pre-Trained Transfer Learning and Improved Decoder for Myanmar Speech Recognition

Abstract

End-to-end (E2E) speech recognition based on deep neural networks has become the mainstream way to build high-performance speech recognition systems. Training end-to-end (E2E) models relies on large-scale “audio-text” pair datasets, which poses a challenge for studying speech recognition of non-generic languages under low-resource conditions. Unsupervised pre-training methods have achieved good performance in speech recognition for many low- resource languages. We build a Myanmar language speech recognition baseline system based on self-supervised speech pre-trained models under low-resource conditions. To further study the methods to improve the accuracy of Myanmar speech recognition, we propose two optimization methods: (1) Different from the current mainstream method of transfer learning in frontend modules, we use the HuBERT self-supervised speech representation model, which pre-trained on massive unlabeled common language speech data, as the encoder of the E2E Myanmar speech recognition system to extract high performance speech recognition features; (2) To better adapt to long sequence-dependent tasks during the decoding process, we proposed and implemented an improved decoder framework based on the S4 module. The experimental results show that the character error rate of the baseline Myanmar speech recognition system constructed in this paper is 12.6%. After introducing the two improvement methods proposed in this paper, its character error rate decreases to 8.2%.