Speech AI · Research / 2026
Listening beyond the baseline
Speech recognition for voices often overlooked.
The problem
General-purpose speech recognition can struggle with dysarthric speech. This project explores whether adapting Whisper-small to Hindi dysarthric speech can improve transcription on unseen speakers and sentences.
How I approached it
I developed the preprocessing and training pipeline: audio preparation, Romanized Hindi normalization, Devanagari conversion, and mixed-precision fine-tuning. The documented split separates both speakers and sentence templates to reduce leakage. I also coordinated a five-person team across research, implementation, and presentation.
What came out of it
The repository reports WER decreasing from 136.03% to about 84.36%, and CER from 110.68% to 61.98%, on a 45-sample test set. WER can exceed 100% when insertions are numerous. The improvement is promising, but the remaining error rate and small test set leave substantial work before practical use.
What I took away
Reliable evaluation starts with the split and the transcripts, well before the training loop.
Project details: repository and documentation . Personal contributions are based on my résumé and project notes.