MyanSpeech: Joint CTC-Attention and RNN Language Model for End-To-End Read Speech Recognition

This paper proposed MyanSpeech¹, an end-to-end Automatic Speech Recognition (ASR) system designed for the Myanmar language, exploiting a joint CTC-Attention architecture for the acoustic model in combination with a Recurrent Neural Network Language Model (RNN-LM). The integration of CTC-Attention addresses the challenge of aligning audio with text in Myanmar, effectively combining the strengths of Connectionist Temporal Classification (CTC) for stable framewise alignment with Attention mechanisms for handling longer dependencies and variable input lengths. This dual approach enhances recognition accuracy, particularly for low-resource languages. The RNN-LM further supports the model by capturing language-specific grammar and contextual relationships. The dataset includes 36 hours, 33 minutes, and 51 seconds of read speech from 20 males and 18 females speakers, comprising 35,468 utterances of daily conversation from the Linguistic Committee (LC). Among various configurations, the best models achieved a word error rate (WER) of 7.093% and character error rate (CER) of 4.474% on a speaker-dependent (closed test), and a WER of 26.829% and CER of 17.597% on a speaker-independent (open test), respectively. These results highlight the effectiveness of integrating CTC-Attention and RNN-LM for Myanmar ASR, offering a reliable solution for speech recognition in low-resource language.