Practical overview
When it fits
Use Fun Asr when you need model-based transcription and can select a checkpoint that matches the audio languages. The Nano checkpoint targets Chinese, English and Japanese, while the MLT checkpoint covers 31 languages.
Who it is for
- Developers building audio transcription features for Chinese, English or Japanese.
- Teams needing transcription across the 31 languages supported by the MLT checkpoint.
- Developers exploring Chinese dialect and regional-accent recognition.
- Implementers seeking documented model inference or serving routes, including a llama.cpp CPU or edge path.
Documented uses
- Convert supported-language audio into text.
- Transcribe Chinese dialects and regional accents with the Nano checkpoint.
- Add Fun-ASR inference or serving to an application through FunASR.
- Explore CPU or edge inference for Nano using the documented llama.cpp and GGUF route.
Requirements & limits
- Checkpoint choice determines language coverage: Nano lists Chinese, English and Japanese; MLT lists 31 languages.
- The native Transformers quickstart is specifically for the Nano Hugging Face checkpoint.
- FunASR is documented for inference and serving with both checkpoints.
- The llama.cpp instructions describe a CPU or edge route for Nano using GGUF files.
- Speaker diarization requires separate FSMN-VAD and CAM++ models; it is not emitted by the Fun-ASR checkpoints alone.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Transcribe audio with Fun-ASR-Nano using Transformers
Set up the documented native Transformers quickstart and run its example transcription.
Before you start
- Python environment with Transformers 5.17.0, PyTorch 2.10.0, torchaudio 2.10.0, librosa 0.11.0 and soundfile 0.13.1.
- The Fun-ASR-Nano-2512-hf checkpoint and the example audio URL used in the repository recipe.
Steps
Install the versions of Transformers, PyTorch, torchaudio, librosa and soundfile listed in the repository quickstart.
Load the checkpoint processor and model with Hugging Face Transformers.
Pass the example audio to the processor with the desired language and generate transcription tokens.
Decode the generated tokens to obtain the transcription text.
Expected outcome
A text transcription from the documented Fun-ASR-Nano example.
Guide limitations
- This quickstart is for Fun-ASR-Nano, whose listed languages are Chinese, English and Japanese; the 31-language MLT checkpoint is separate.
- The cited older Hugging Face Nano artifact is text-only and omits timestamps.
Run Fun-ASR-Nano through the documented llama.cpp route
Use the repository’s CPU or edge command-line path to transcribe an audio file with GGUF models.
Before you start
- A compatible prebuilt llama.cpp binary or the documented llama.cpp runtime setup.
- Fun-ASR-Nano encoder, language-model and FSMN-VAD GGUF files.
- An audio file to pass to the command-line tool.
Steps
Use the repository downloader to fetch the Nano GGUF files and FSMN-VAD model into a local directory.
Run llama-funasr-cli with the encoder, language model, audio file and VAD model paths.
Expected outcome
The documented command-line inference path processes an audio file with Fun-ASR-Nano.
Guide limitations
- The repository names this as a CPU or edge route, but the supplied instructions do not specify hardware requirements or performance.
- This path describes Nano; the documentation provided here does not establish the same route for the 31-language MLT checkpoint.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- QwenAudio/Fun-ASR · Official homepage title · retained in its source language