Hishab’s NLP team has researched on the development of domain-agnostic ASR (Artificial Speech Recognition) dataset and the research paper is accepted by EMNLP workshop.
Rabindranath Nandi (Senior NLP Engineer, Hishab), Mehadi Hasan (Senior NLP Engineer, Hishab), Tareq-Al-Muntasir (Head of NLP, Hishab), Sagor Sarker (NLP Engineer, Hishab), Quazi Sarwar Muhtaseem (Junior NLP Engineer, Hishab) & Md. Tariqul Islam (Machine Learning Engineer, Hishab) was working with Shammur Absar Chowdhury & Firoj Alam from Qatar Computing Research Institute (QCRI), Doha, Qatar to complete this research paper.
Addressing a critical bottleneck in the development of Automatic Speech Recognition (ASR) systems for low-resource languages, a team of engineers at Hishab and QCRI has introduced a groundbreaking pseudo-labeling approach. This pioneering methodology has resulted in the creation of a large-scale, domain-agnostic ASR dataset, boasting over 20,000 hours of labeled Bangla speech.
Utilizing the extensive corpus, the team designed a conformer-based ASR system, leveraging cutting-edge technology to enhance accuracy and efficiency. The trained ASR system underwent rigorous benchmarking against publicly available datasets and human-annotated domain-agnostic test sets including telephony data, showcasing its prowess when compared to existing models in the field.
The research paper is publicly available at https://arxiv.org/pdf/2311.03196.pdf
What is EMNLP?
Empirical Methods in Natural Language Process (EMNLP) is a leading conference in the area of natural language process and artificial intelligence.
Hishab is a world leader in telephony based Conversational AI technology. Hishab holds over 20 patents in over 27 countries across the world. Hishab’s team is represented by people from 10 different countries across Europe and Asia. Bangladeshi Engineering meets with Japanese and Indian leadership to bring the world’s MOST innovative solution for digitalization of the mass across the world.