Bangla ASR Engine

End-to-end Bangla speech recognition and voice command system for low-resource language processing

Bangla Automated Speech Recognition Engine was a Samsung Research Bangladesh-funded research project on low-resource Bangla speech recognition, voice command understanding, and speech/text resource development.

The project explored end-to-end automatic speech recognition (ASR) for Bangla, including speech corpus preparation, grapheme-to-phoneme modeling, and contextual rescoring for voice command recognition. This work contributed to peer-reviewed publications at ICASSP 2020, LREC 2020, and NAACL 2019.

Highlights

  • Worked on a Samsung Research Bangladesh-funded Bangla ASR research project.
  • Contributed to an end-to-end Bangla voice command recognition system using contextual rescoring.
  • Explored topic-modeling-based contextual information to improve voice command recognition.
  • Helped prepare Bangla speech corpus resources from publicly available audio and text.
  • Worked on grapheme-to-phoneme modeling for non-trivial Bangla transcription problems.
  • Built and evaluated sequence-to-sequence and attention-based models for Bangla ASR and text/speech processing.
  • Contributed to research publications at ICASSP, LREC, and NAACL.
  • Bangla Voice Command Recognition in End-to-End System Using Topic Modeling Based Contextual Rescoring, ICASSP 2020.
  • Preparation of Bangla Speech Corpus from Publicly Available Audio & Text, LREC 2020.
  • Customizing Grapheme-to-Phoneme System for Non-Trivial Transcription Problems in Bangla Language, NAACL 2019.

Keywords

Bangla ASR, speech recognition, voice command recognition, low-resource NLP, contextual rescoring, topic modeling, grapheme-to-phoneme conversion, speech corpus development, Samsung Research Bangladesh.