Open-source work on Malayalam speech AI and edge ML. Each project links to code, a model or a demo.
Malayalam Speech Recognition
Problem: Malayalam has few open speech models, and the best public result I found was 61.84% WER.
Method: I fine-tuned Whisper-small on a corpus of 96.6k clips from six sources. My v1 model reached 37.64% WER on Common Voice.
What I found: When I audited my own data, 41.6% of the test set had leaked from training. I rebuilt a clean test set and retrained.
Result: v2 scores 31.2% WER and 4.0% CER on the leakage-free test set. I also exported it to GGML/GGUF for CPU use.
Model | Dataset | Demo | Benchmark code
Blog series: Part 1: the leakage audit | Part 2: clean-data re-test | Part 3: retraining and an evaluation bug
Malayalam Text-to-Speech
Two voice models trained on my own voice. The VITS model is a fine-tune of MMS-VITS. The F5-TTS voice clone runs about 5x faster on CPU.
VITS model | VITS demo | F5-TTS model
Malayalam Voice Assistant
A voice assistant that runs fully on a CPU. It uses my Whisper v2 model for listening, Gemma 4 E2B for replies and my VITS voice for speaking, with hands-free voice detection in the browser.
Try it
Acoustic Rain Gauge (ICFOSS)
I led the ML for Kerala’s first IoT acoustic rain gauge. It estimates rainfall from sound on a low-cost edge device, with 15% MAPE. Published at IEEE RAICS 2024.
Paper | Code | Demo video
More projects
- ICTSahayi: a fine-tuned GPT-2 FAQ chatbot built for ICT Academy of Kerala. Code
- GPT Story Teller: a small GPT built from scratch in JAX/Flax on 200K stories. Code
- AI Assignment Grader: grades notebooks against a rubric with a local LLM. Code | Demo video
- OmniSense: semantic search inside videos, with timestamp seek. Code