Speech recognition for Indian courts
Most of my current work, at Adalat AI.
- Training fast in-domain ASR models for low resource languages
- Focus on dictation ready formatting for speech to legal text
- Multilingual and code-switched Indic ASR
- Conformer + Transducer training recipes
Speaker diarization
Who spoke when.... and what?
- Tracking speaker identities across long sessions
- One model for Diarization + ASR with session context
- Focus on Indic Speech
NLProc
Where I started.
- Parameter-efficient cross-lingual transfer for low-resource languages
- Linguistic resources for Lambani, an under-resourced tribal language
- Cross-lingual phrase retrieval for Indian languages
- Interpretability of sentiment analysis models for low-resource languages
- Disentangling speaker and language information from speech
- Precise profanity localization in speech
- Mitigating Whisper hallucinations with adaptive layer attention and knowledge distillation
- Efficient replacements for self-attention when fine-tuning SSL speech models
What I want to do next
My bets for the next 1 year
- Speaker-aware speech recognition and synthesis for long-form speech
- Language coverage and accent robustness for Indic ASR
- Recipes for domain-specific ASR