Nemotron 0.6B Indic-English Name Recognition Model
This is a fine-tuned version of nvidia/nemotron-speech-streaming-en-0.6b optimized for rare Indian name recognition.
Success Metrics
- Indian Names WERR: 78.8% (Relative reduction in errors)
- LibriSpeech WER: 2.32% (Baseline was 2.79%)
- Custom Dataset: Optimized for a private 37k Indian Name list.
- Artifacts: 0% '??' duration artifacts (Fixed via Phase 2 Acoustic Hardening).
Training Procedure
- Phase 1: MWER discriminative fine-tuning (Frozen Encoder).
- Phase 2: Unfrozen Encoder "Acoustic Polishing" at 1e-5 Learning Rate.
- Data Strategy: 3:1 Mixed Rehearsal (Names:LibriSpeech).
Inference Parameters
To achieve the benchmarked accuracy, use these settings in your NeMo inference pipeline:
- Decoding Strategy:
malsd_batch
- Beam Size: 4
- Depth Scaling: 2.0 (Mandatory for long Indian names)
- Boosting Alpha: 1.2
