Using Fun-ASR-Nano with Transformers 5.17: examples and migration notes
Native Fun-ASR-Nano support is available in the released Transformers 5.17.0 package. This -hf repository is the checkpoint for that route; a Transformers source checkout or trust_remote_code=True is not required.
Run Your Own Audio
The documented reference environment is Linux x86-64, Python 3.12, CPU float32, with matching PyTorch/torchaudio versions and a pinned model revision. English, Chinese, contextual-keyword and padded-batch checks used two short public audio samples. They are smoke tests, not an accuracy or GPU benchmark. Hosted Colab execution is not part of that validation.
Choose the Right Route
The original toolkit checkpoint, native -hf checkpoint, vLLM checkpoint and GGUF files are different artifacts. This native generation export does not provide the CTC-dependent timestamps or speaker diarization, and it is not a realtime server. The model card and guides explain the alternatives.
中文：现在可直接使用正式版 Transformers 加载本仓库的原生权重。上面的中文指南包含安装、本地音频和批量转写入口；请勿将原版工具包、vLLM 或 GGUF 权重直接替换进原生示例，也不要把文本转写能力理解为已经支持时间戳或说话人分离。
For a reproducible problem, please include the package versions, model revision, device/dtype, minimal command and a shareable audio sample or exact error. Avoid including access tokens or private recordings.