On-Policy Self-Distillation boosts Qwen3-ASR dialect accuracy without hurting Mandarin
New OPSD method refines dialect ASR while keeping Mandarin CER flat—no trade-off.
Researchers led by Shuiyuan Wang (with 8 co-authors) tackled the perennial problem of adapting large ASR models to Chinese dialects without sacrificing Mandarin accuracy. Their approach, detailed in arXiv:2608.11898, uses a three-stage pipeline: continual pre-training (CPT), dialect supervised fine-tuning (SFT), and a final refinement stage called On-Policy Self-Distillation (OPSD). OPSD directly addresses the train-test mismatch in autoregressive models by letting the student decode its own prefixes while a frozen teacher—given the reference transcript as privileged context—supplies soft token-level targets. This distillation replaces hard cross-entropy updates on dialect data, preserving existing Mandarin knowledge.
Instantiated with Qwen3-ASR-1.7B, the method was evaluated on both public and internal Mandarin and dialect test sets. Results show that under matched refinement data and schedule, OPSD improves dialect recognition without raising Mandarin CER, whereas continued teacher-forced fine-tuning increases Mandarin CER. The team is open-sourcing the model weights and evaluation scripts, making it possible for developers to apply the same technique to other ASR models and dialect corpora.
- OPSD (On-Policy Self-Distillation) trains ASR on student's own decoded prefixes instead of ground-truth tokens, closing train-test gap.
- Qwen3-ASR-1.7B adapted with CPT + dialect SFT + OPSD improved dialect CER without raising Mandarin CER.
- Authors will release model weights and evaluation scripts on arXiv.
Why It Matters
Enables developers to extend Chinese dialect support for production ASR without regressing Mandarin, a common deployment blocker.