R2T2 critique: unexplained cross-lingual transfer and fixed chunk sizes cap streaming ASR
dr_cintas · x · 2026-11-06
drcintas offers a technical critique of R2T2, the streaming ASR work built on Qwen3-ASR:
- Cross-lingual transfer is impressive but unexplained: trained only on EN+ZH, streaming gains extend to 18 test sets across 10 more languages (average WER/CER at 160ms drops from 30.75 to 8.4–9.6). He attributes it to the frozen encoder's language-agnostic acoustic representations and calls for an ablation on the encoder's multilingual capacity.
- Chunk size is fixed at inference: an adaptive chunking strategy keyed to audio complexity (e.g., larger chunks in noise) could cut latency further without hurting accuracy.
In a prior reply he also notes the LSP target is computed from the base model's full-context output, so R2T2's stability ceiling is bounded by Qwen3-ASR's full-context quality; a self-consistent LSP could push further.
More from Research
- Tsinghua's Coca-Cola compresses agent tool outputs, cutting tokens up to 89% while boosting accuracy — Tsinghua · 2026-11-07
- baselabs unpacks OPSD post-training research: hinted-teacher bias hurts long-horizon learning — JoshPurtell · 2026-11-07
- s³ paper: cooperative multi-agent RL post-training beats best-of-N by up to 166% — StellaLisy · 2026-11-07
- Rough VIX Model Cuts Option Pricing MSE by 13% in Volatility Research Paper — chaumian · 2026-11-07
- Stanford EduNLP lab hiring Applied Research Lead for AI teaching tools — sebgehr · 2026-11-07
- NeurIPS Quest: a pixel-art map to navigate 672 posters across sessions and topics — shangbinfeng · 2026-11-07