Real world Automatic Speech Recognition (ASR) system development requires rigorous performance review under varying real world conditions. This paper reports our effort on ASR resource creation, transcription, system building and performance assessment for connected and continuous word applications in Bengali language (ranked seventh worldwide) using GMM-HMM and DNN framework on available open source toolkits. Baseline models are built from merging Bengali dataset hosted on government sites with application specific 100 hours indigenous audio collected under target deployment scenario. After feedback analysis of live systems by real users, novel Error Handling Techniques like Signal Analysis and Decision, Confidence based ASR output Polling and Runtime LM are implemented which results around 2%- 11% overall gain in WER with encouraging task success rates in final field trials. Results suggest that recent approaches along with application, environment, target user and runtime resource specific appropriate strategies will yield better acceptability of live ASR systems in India.
Cite as: Khan, S., Pal, M., Basu, J., Samirakshma Bepari, M., Roy, R. (2018) Assessing Performance of Bengali Speech Recognizers Under Real World Conditions using GMM-HMM and DNN based Methods. Proc. 6th Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU 2018), 192-196, doi: 10.21437/SLTU.2018-40
@inproceedings{khan18_sltu,
author={Soma Khan and Madhab Pal and Joyanta Basu and Milton {Samirakshma Bepari} and Rajib Roy},
title={{Assessing Performance of Bengali Speech Recognizers Under Real World Conditions using GMM-HMM and DNN based Methods}},
year=2018,
booktitle={Proc. 6th Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU 2018)},
pages={192--196},
doi={10.21437/SLTU.2018-40}
}