FunASR

Use this root skill to route FunASR requests. Start here when the user asks for speech transcription, subtitles, model choice, service deployment, LLM-ASR routing, punctuation cleanup, manifests, or training/export guidance.

Start here

• Package provenance and staleness check: references/repo-provenance.md
• Router metadata for repo-skills-router: references/repo-routing-metadata.json
• Package commands and entry points: references/cli-reference.md
• Public API shapes and input/output forms: references/api-reference.md
• Model-family choice guide: references/model-overview.md
• Cross-skill workflow map: references/workflows.md
• Shared input/output schemas: references/data-formats.md
• Cross-cutting failure recovery: references/troubleshooting.md
• Safe environment smoke helper: scripts/check_funasr_env.py

Route map

Need	Go to
Plain transcription, batch ASR, subtitles, audio decoding, hotwords, timestamps, or common non-LLM model choice	sub-skills/python-asr-pipelines/SKILL.md
Punctuation cleanup, inverse text normalization, or text normalization	sub-skills/text-normalization/SKILL.md
OpenAI-compatible HTTP API, realtime WebSocket, MCP, browser/client integration, or edge/runtime guidance	sub-skills/serving-and-runtime/SKILL.md
Fun-ASR-Nano, Fun-ASR-MLT-Nano, GLM-ASR-Nano, Qwen3-ASR, or AutoModelVLLM	sub-skills/llm-asr-and-vllm/SKILL.md
Training data prep, manifests, distributed config, export, or local inference after export	sub-skills/training-data-and-export/SKILL.md

What this root skill owns

• The first decision about which FunASR workflow to use.
• Quick package inspection and safe help/version checks.
• The common vocabulary for ASR, VAD, punctuation, speaker labels, hotwords, subtitles, service deployment, and manifests.
• Cross-cutting troubleshooting that applies before a user narrows to a sub-skill.

How to start safely

1. Run python scripts/check_funasr_env.py --check-cli --check-torch to confirm the package and entry points are visible in the current environment.
2. Use import funasr as the minimal package import check.
3. Remember that import funasr may work before PyTorch is installed, but from funasr import AutoModel needs a compatible torch build.
4. If the request names a model family, use references/model-overview.md before choosing a route.

Where to read next

• Transcription and subtitles: python-asr-pipelines
• Optional punctuation cleanup or ITN/TN: text-normalization
• HTTP/WebSocket/MCP/runtime deployment: serving-and-runtime
• Nano/GLM/Qwen3/vLLM backend choice: llm-asr-and-vllm
• Training and export: training-data-and-export

Do not do here

• Do not bury service, model-family, normalization, or training guidance inside this router.
• Do not tell users to open the original source checkout from runtime instructions.
• Do not claim GPU, vLLM, or Pynini support from the root skill alone; route to the owning sub-skill and its references.

Quick model instincts

• CPU-friendly first try: SenseVoiceSmall
• Mandarin production and subtitles: paraformer-zh
• GPU-heavy LLM-ASR / batch throughput: Fun-ASR-Nano or GLM-ASR-Nano through llm-asr-and-vllm
• Server or client integration: serving-and-runtime

Cross-cutting checks

• Use references/troubleshooting.md for install/import, hub, device, and download issues.
• Use the sub-skill troubleshooting pages for workflow-specific failures.
• Keep review-only outputs in skills/tests/funasr/, not inside the runtime skill tree.