transformers

Fun-ASR-Nano with Transformers

Transcribe Chinese, English or Japanese with the native Hugging Face processor and generation API. No FunASR toolkit, repository-local model.py, or remote Python code is needed. The first model load downloads about 1.66 GB of weights.

Online demo · Notebook · 中文指南

First transcript

Clone the QwenAudio/Fun-ASR code repository, not a Hugging Face weight repository. Run all commands below from this repository's root:

git clone https://github.com/QwenAudio/Fun-ASR.git
cd Fun-ASR

Use an isolated environment. This reproducible recipe was exercised on Linux x86-64, Python 3.12 and CPU; it does not upgrade your existing serving environment.

python3.12 -m venv .venv-native
. .venv-native/bin/activate
python -m pip install --index-url https://download.pytorch.org/whl/cpu 'torch==2.10.0+cpu' 'torchaudio==2.10.0+cpu'
python -m pip install -r examples/transformers/requirements.txt
python -m pip check
python examples/transformers/transcribe.py

The default input is the pinned official English sample. A successful result contains text and reached_eos: true. No access token is required for this public checkpoint. The model itself also runs through the short, no-clone Python recipe.

Use the Transformers pipeline API

After the CPU installation above, this Python example also works without a repository clone. Explicitly select any-to-any: it uses the native processor and the checkpoint's structured transcription chat template.

from copy import deepcopy
import torch
from transformers import pipeline

torch.set_num_threads(4)
transcriber = pipeline(
    "any-to-any",
    model="FunAudioLLM/Fun-ASR-Nano-2512-hf",
    revision="d93b302ee7fd505e1b3576120fc142fc6f7820e1",
    device="cpu", dtype=torch.float32, trust_remote_code=False, token=False,
)
messages = [{"role": "user", "content": [
    {"type": "audio", "url": "https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512/resolve/272c57b82523ada6fd87095e955f8e29100979ab/example/en.mp3"},
    {"type": "language", "language": "英文"},
]}]
generation_config = deepcopy(transcriber.model.generation_config)
generation_config.update(max_new_tokens=128, do_sample=False, num_beams=1)
result = transcriber(
    text=messages, return_full_text=False,
    generate_kwargs={"generation_config": generation_config},
    processor_kwargs={"audio_kwargs": {"sampling_rate": 16000}},
)
print(result[0]["generated_text"])

For a local recording, replace the audio entry with {"type": "audio", "path": "recording.wav"}. In this checkpoint's chat template, the language field uses 中文, 英文 or 日文; the higher-level apply_transcription_request helper additionally accepts ISO codes. return_full_text=False excludes the input conversation from the generated text.

On Transformers 5.17.0, pipeline("automatic-speech-recognition", ...) is a different processing path: our pinned-checkpoint test fails with a floating-point token-index error. Do not replace any-to-any with that task or rely on automatic task inference. The example above was checked on the public English sample with CPU float32; it is not a pipeline GPU/batching, accuracy or capacity evaluation. The 128-token limit is for this short sample, not a completeness guarantee. Use the CLI below for explicit audio limits and EOS diagnostics.

Your recordings and batches

python examples/transformers/transcribe.py recording.wav --language zh --keywords 开放时间
python examples/transformers/transcribe.py chinese.wav english.wav --language zh en
python examples/transformers/transcribe.py english.wav --language en --prompt 'A conversation about opening hours.'

Files are read as float32, stereo/multichannel audio is averaged to mono, and non-16 kHz audio is explicitly resampled with soxr_hq. Original files are not modified. This example accepts 1-4 files and at most 60 seconds total. Longer recordings are rejected, not silently trimmed. Segment long audio deliberately or choose a serving recipe.

Output order matches input order. One language applies to all files, or give one per file. keywords and prompt are hints, not enforced vocabulary. Missing EOS or empty text produces a nonzero exit after printing the diagnostic result. EOS is a generation boundary, not proof that every word was recognized.

CUDA: an isolated, tested recipe

Use a separate environment for this GPU recipe; do not install the CPU requirements into it. It was functionally tested on Linux x86-64, Python 3.12, an NVIDIA H100 80 GB and driver 550.127.08, with PyTorch/torchaudio 2.11.0+cu128 and Transformers 5.17.0. Other GPU/driver combinations need their own validation.

python3.12 -m venv .venv-native-gpu
. .venv-native-gpu/bin/activate
python -m pip install --index-url https://download.pytorch.org/whl/cu128 'torch==2.11.0+cu128' 'torchaudio==2.11.0+cu128'
python -m pip install -r examples/transformers/requirements-gpu.txt
python -m pip check
python examples/transformers/transcribe.py --device cuda --dtype float32
python examples/transformers/transcribe.py chinese.wav english.wav --language zh en --device cuda --dtype bfloat16

The last command uses your local files. Without device flags, the CLI still uses CPU float32. CUDA requests fail if CUDA is unavailable; they never silently fall back to CPU. BF16 also requires device support. Model weights and processor tensors are moved to the selected device without converting integer token IDs to a floating dtype. JSON output records the actual device, dtype and versions.

On 2026-09-10, both float32 and BF16 passed English, Chinese, Chinese with keywords, and padded Chinese/English batch inference using the pinned public samples. Outputs were nonempty and reached EOS. The Chinese sample still transcribed 开放时间 as 开饭时间, including with the keyword hint; these checks do not establish accuracy or keyword benefit. They are not throughput, minimum VRAM or concurrency benchmarks. Float16 and other accelerators were not tested.

Transformers emitted an attention-implementation warning in this environment. Successful inference does not verify the attention kernel of every component; this recipe makes no Flash Attention or all-SDPA performance claim. The online demo and notebook linked above do not imply a verified hosted GPU environment.

Select a backend, not just a suffix

Need	Artifact and entry point
Native Transformers	FunAudioLLM/Fun-ASR-Nano-2512-hf, AutoProcessor + AutoModelForSpeechSeq2Seq
FunASR pipelines and existing services	FunAudioLLM/Fun-ASR-Nano-2512, funasr.AutoModel
Native vLLM serving	FunAudioLLM/Fun-ASR-Nano-2512-vllm, native vLLM guide
C++ / edge	Converted GGUF, llama.cpp guide

The native -hf export produces transcription text. It does not add word timestamps, speaker identities, a streaming protocol or an HTTP server. The 31-language MLT checkpoint is a separate model, not an alternate name for this zh/en/ja export. The CUDA recipe above covers the stated functional cases only; attention kernels and serving throughput require separate hardware evaluation.

Verification

Transformers 5.17.0 is a released package containing fun_asr_nano. The examples pin native checkpoint revision d93b302ee7fd505e1b3576120fc142fc6f7820e1 and set trust_remote_code=False. Chinese/English public-sample inference, keywords and padded batching were functionally exercised on CPU on 2026-09-09; these checks are not CER/WER or capacity evaluation.

Upstream model documentation and model card describe the native interface. Keep the runtime, model revision and raw output with your own evaluation. Never publish private recordings in a bug report.