|
Bengali.AI
I contribute to Bengali.AI, an open research community focused on building datasets, benchmarks, and models for Bengali language technologies—spanning ASR, NLP, and computer vision—and releasing them publicly (often via Kaggle) to accelerate research.
What I did
- Built and released open-source benchmarks/datasets (with clear train/eval protocols) and helped drive community adoption via competitions and public baselines.
- Worked end-to-end: data collection/QA → dataset/benchmark design → baseline training + evaluation → documentation + release.
- Collaborated with a distributed team of researchers/engineers to ship artifacts that are usable by the research community (not just a paper).
Highlights (selected)
OOD-Speech (ASR benchmark)
- Large Bengali out-of-distribution speech recognition benchmark.
- 1100+ hours, 22,000+ contributors, 17 domains.
- Whisper fine-tuning + evaluation for regional Bengali ASR.
- Links: paper, Kaggle, demo
BaDLAD (Document AI dataset)
- 33,695 annotated Bengali document samples across multiple domains.
- Trained Mask R-CNN / YOLO-based layout detectors for document layout analysis.
- Links: paper
Where to find the work
News
|