Rafat Haameem

At Decodis, I built and shipped production ASR for Swahili, Yoruba, Azerbaijani, and Hausa at roughly 100,000-audio scale. I reduced word error rate on noisy, overlapped field audio from 60% to 28% for Swahili and from 36% to 18% for Azerbaijani.

My work combines end-to-end and streaming ASR, speech-to-speech modelling, speech LLMs, and model efficiency. I focus on evaluation that reveals where models fail, including morphology-aware and code-switching speech evaluation.

At Apurba Technologies and the NSU R&D Lab, I lead Latent-RAG, an end-to-end low-resource speech-to-speech research effort, and serve as Co-Principal Investigator on MenoChat, a voice-first health assistant for low-resource Bangla settings. I also mentor junior researchers from raw data through publication.

I've first-authored work published at INTERSPEECH, IEEE JBHI, and PLOS ONE, including research on streaming ASR, rehabilitation assessment, and lower-carbon knowledge distillation. I enjoy turning research into systems people use while making them more accurate, efficient, and accessible.