Cutting word error on low-resource Indic ASR
Error analysis, data curation and multi-stage fine-tuning of Whisper, Wav2Vec2 and comparable models, moving word error rate on multilingual holdout sets rather than on a leaderboard split.
- Relative WER reduction
- 25—50%
- Audio curated
- 100k+ h
- Training
- multi-GPU DDP
- Inference
- vLLM
समस्या
State-of-the-art ASR checkpoints are competitive in English and weak in low-resource Indic languages, where the failures are systematic rather than random: specific phonetic contrasts, specific code-mixing patterns, specific channel conditions.
Scaling data indiscriminately does not fix a systematic error. It buys a smaller version of the same error at higher cost.
निर्णय
Each round started with error analysis on holdout audio, and the curation followed from it — cleaning, quality filtering, metadata generation, and assembly of task-specific training sets drawn from a corpus of over 100,000 hours, rather than a single undifferentiated pool.
Fine-tuning then ran in stages against those targeted sets, with distributed PyTorch multi-GPU training optimised for throughput and high-throughput vLLM inference pipelines for the evaluation and pseudo-labelling passes that fed the next round.
Baselines were established across low-resource Indic languages first, so each subsequent gain was measured against a known number rather than an impression.
परिणाम
25 to 50 percent relative WER reduction across multilingual holdout evaluation sets, with competitive baselines published internally for the low-resource languages that had none.