Kazi Rafat Haa Meem
Kazi Rafat Haa Meem
Speech & Language AI Researcher
Apurba–NSU R&D Lab · Dhaka, Bangladesh
Publishes as Kazi Rafat · also cited as K. Rafat
Summa Cum Laude · North South University Merit-Based Scholarship · nationwide competitive admission $34.5K+ in research grants awarded Adaption Labs inaugural cohort

I’m a speech researcher at the Apurba–NSU R&D Lab in Dhaka. I study the representations inside speech models, what they capture, how that information is organised, and how it can be drawn out and improved. What drives me is putting that understanding to work, building speech systems that make better use of what models already know, and designing models whose training is shaped by what their representations reveal. Before this, I spent two and a half years at Decodis shipping production speech systems, which taught me to care about the data behind a model and about evaluation that shows where it breaks. The next billion people will talk to machines rather than type, and today’s models barely hear them.

The next billion people will talk to machines, not type — and today's models barely hear them.

The problem I'm becoming an expert in

Speech LLMs are converging on a fluent, low-latency, conversational interface — yet we still understand little about what these models actually represent about the speech they hear. They encode how something was said and then lose it before they reply. Aggregate error rates hide exactly the code-switching and suffix-level failures that matter. And systems decode transcripts to do things their own representations could already support. I want to understand where meaning and delivery live inside speech models, and use that understanding to build speech-to-speech systems that stay grounded in what is true, respond to how people actually speak, and are evaluated on metrics that expose where they break instead of averaging over it.

Evidence I've already made progress on it

Speech LLMsSpeech RecognitionSpeech-to-SpeechText-to-SpeechAudio Signal ProcessingAgentic AIKnowledge DistillationModel CompressionHealth AICode-Switching
In progress Understanding the representations inside speech models and using them to build better speech systems. Applying to PhD programs.

Highlights

scroll for more ↓

Systems in Production

At Decodis I shipped ASR for Swahili, Yoruba, Azerbaijani, and Hausa into workflows running at roughly 100k audios. On real noisy field audio with heavy speaker overlap, Swahili word error rate went from 60% to 28% and Azerbaijani from 36% to 18% — achieved by attacking the high-error tail rather than the average, with SNR-based filtering, controlled-ratio overlap augmentation, and pitch-aligned speaker concatenation feeding a recursive training loop.

▶ Hear it — Swahili field audio
ReferenceWhatsApp yangu mimi hutumia kuongea na watu walio nje ya nchi, na kazi zangu pia nyingi ni ya WhatsApp na marafiki pia huwa tunaongea kupitia WhatsApp
Model outputwhatsapp yangu mimi hutumia kuongea na watu walio nje yaaa nchi, na kazi zangu pia nyingi ni ya whatsapp na marafika pia huwa tunaongea kupitia whatsapp
Conversational Swahili from the deployed pipeline. Remaining errors are underlined.

Publications

Dynamic Block-Online Streaming ASR for Low-Resource Agglutinative Code-Switching Speech with Morphology-Aware Evaluation
Interspeech 2026OralFirst author
K. Rafat, A. Imran, M. I. Hossain, M. R. Ali, F. Rahman, S. Rahman, N. Mohammed
A dynamic block-online model trained with global attention, achieving a 17.5% relative error reduction over conventional streaming approaches on code-switched speech. Introduces CS-WER, an evaluation framework measuring suffix-, switch-, and root-level performance separately.
▶ See it — Bangla–English code-switching
Referencecollegeটি ঢাকা শিক্ষা boardের আওতাধীন এবং জাতীয় বিশ্ববিদ্যালয়ের অধিভক্ত
Ours — dynamic block-onlinecollegeটি ঢাকা শিক্ষা bordের আওতাধীন এবং জাতীয় বিশ্ববিদ্যালয়ের অধিভক্ত
Conventional streamingcollegeটি ঢা শিক্ষা bordর আওতাধীন এবং জাতীয় বিশ্ববিদ্যালয়ের অধিভক্ত
Conventional streaming truncates at block boundaries (marked): ঢাকা clipped to ঢা, and the Bangla suffix ের on the English loanword board reduced to র. Exactly the failure aggregate WER hides — and what CS-WER was built to measure.
A Point Cloud Transformer for Remote Monitoring and Automated Assessment of Physical Rehabilitation Exercises
IEEE JBHI 2026IF 7.7 · Q1First author
K. Rafat, Md. I. Hossain, M. M. L. Elahi, S. Momen, F. Rahman, N. Mohammed, S. Rahman
Geometric transformer with low-cost attention that classifies and scores rehabilitation exercise quality from 3D pose and motion data. Supported by the CTRG 2023 national grant.
Mitigating Carbon Footprint of Hyper-parameter Selection During Knowledge Distillation
PLOS ONE 2023Q1First author
K. Rafat, A. A. Mahfug, M. I. Hossain, S. Momen, F. Rahman, S. Rahman, N. Mohammed
A distillation framework that removes computationally expensive hyper-parameter search, generalising across image classification, data-free distillation, and object detection. Began as an undergraduate course project.
DeepMarkerNet: Supervised Duchenne Marker Modeling for Spontaneous Smile Recognition
Pattern Recognition Letters 2024IF 3.9Co-author
M. J. Hasan, K. Rafat, F. Rahman, N. Mohammed, S. Rahman · Elsevier
Precision-Driven Low-Resource Speech Synthesis for Bangla Text-to-Speech System
PML4LRS @ ICLR 2024Co-author
T. S. Shahjahan, M. I. Hossain, K. Rafat, M. R. Amin, F. Rahman, N. Mohammed

Work in Progress

Speech-to-Speech RepresentationsManuscript in preparation
Studying what an end-to-end speech-to-speech model’s internal representations capture about the speech it hears, and using that understanding to make spoken assistants better at grounding their answers in knowledge and at adapting how they speak to the user. Details after review.

Patents

Theme-Based Voice Emotion Analytics MethodPatent Pending
U.S. Patent Application #19/763,175 · Decodis · 2026
An analytics method augmenting speech-based prosodic information, including a timestamping approach for accurate grapheme and prosodic alignment in real-time multilingual transcription.

Selected Projects

MenoChatMenoChat — Voice-First Women's Health Assistant
$30K GrantBest Idea & Best PosterCo-Principal Investigator
An agentic health assistant for women with limited literacy in Bangladesh, combining Bangla ASR and TTS with a multi-stage, comorbidity-aware LLM pipeline that grounds answers in verified medical information and escalates high-risk cases. Refined with public-health experts at icddr,b icddr,b. Integrating Bangla ASR and TTS had stalled others; the end-to-end voice pipeline was made to work here. Prototype scored 75 on an accessibility-adapted SUS.
AdaptBench — Benchmark for LLM Evaluation
Adaption Labs 2026
Selected for the inaugural Adaption Labs grant cohort to build a benchmark for evaluating large language models.
Intelligent Virtual Assistant for Low-Resource Devices
NAACL Mentorship Program
Multilingual, multimodal voice assistant on constrained hardware using distilled language models, with mood recognition and parental controls — reasoning about model compression, multilingual coverage, and on-device deployment simultaneously.

Experience

DecodisJul 2023 – Jan 2026
Machine Learning Engineer · Boston, MA (Remote)
  • Built production ASR for Swahili, Yoruba, Azerbaijani, and Hausa, integrated into end-to-end workflows operating at roughly 100k audios at scale.
  • Reduced Swahili WER from 60% to 28% on a highly noisy, overlapped test set via a voting-based SNR filter, controlled-ratio overlap augmentation, pitch-aligned speaker concatenation, and a recursive training loop.
  • Cut Azerbaijani WER on noisy field audio from 36% to 18% through percentile-based error analysis targeting the high-error tail rather than the average.
  • Designed the cross-language timestamping approach that became the basis of a pending patent; added BERTScore evaluation that surfaced mislabeled training data.
  • Designed and led the company's data-annotation process end to end; delivered $25,000+ in cost savings per project across three projects.
Apurba Technologies · NSU R&D Lab2022 – Present
Research Assistant · Dhaka, Bangladesh · concurrent with Decodis, 2023–2026
  • Lead researcher on speech LLMs, ASR, TTS, knowledge distillation, and agentic AI; published at Interspeech, ICLR, PLOS ONE, PRL, and IEEE JBHI.
  • Appointed Co-Principal Investigator on the funded MenoChat project.
  • Mentored junior researchers, carrying projects from raw data to publication; that mentorship contributed to work at Pattern Recognition Letters and the ICLR PML4LRS workshop.

Collaborators

Teaching

Education

North South UniversityJan 2019 – May 2022
B.Sc. Computer Science & Engineering · Summa Cum Laude
Merit-Based Scholarship throughout (nationwide competitive admission). Core coursework in Pattern Recognition & Neural Networks, Machine Learning, and Probability & Statistics.

Honors & Grants

Contact

I like meeting people who care about speech, multilingual AI, or health. Reach me at kingrafat82@gmail.com.

Download CV