FunASR v1.3.14 fixes two inference failures and makes the first-run model choice and benchmark evidence easier to understand.
Runtime fixes
- SenseVoice Triton path imports correctly again. Misplaced text between Triton decorators and kernel definitions caused a Python SyntaxError during auto-registration and prevented the optimized DTW module from loading. The module now compiles normally. (#3027)
- Paraformer no longer crashes when a short or near-silent segment produces no CIF boundary. The no-fire path now returns an empty frame result safely. On the reported 184-file Chinese benchmark, decoding improved from 179/184 to 184/184 while micro-average CER remained 10.17%. (#3028)
Onboarding and benchmark clarity
- Quick Start now leads with the flagship Fun-ASR-Nano path while retaining SenseVoice as the CPU-friendly multilingual option.
- The benchmark table reports Fun-ASR-Nano's vLLM throughput (340x realtime) and adds Chinese CER, with English and Chinese READMEs kept in sync.
- The model-selection guide and duplicated Quick Start sections were cleaned up so installation, model choice, and first inference are easier to scan.
Upgrade
pip install --upgrade funasr==1.3.14
Included changes
- #3019 — lead Quick Start with Fun-ASR-Nano
- #3020 — update model-selection ordering
- #3021 — remove duplicated Quick Start content
- #3023 — report Fun-ASR-Nano vLLM throughput
- #3024 — add Chinese CER to the benchmark
- #3025 — synchronize translated README benchmarks
- #3027 — fix SenseVoice Triton SyntaxError
- #3028 — guard Paraformer CIF no-fire segments
Full changelog: v1.3.13...v1.3.14