Open-Sourced Fun-ASR-Nano Model Lacks Trained Weights for Reliable Timestamp Prediction

[GeorgeCaoJ]

FunASR Version

1.3.3

Python Version

3.10

Device

cuda (GPU)

Bug Description

Fun-ASR-Nano timestamp support seems incomplete in the current open-sourced release.
Although the implementation introduced in #70 returns timestamp outputs, the released model.pt checkpoint does not include trained ctc_decoder.* / ctc.* weights. As a result, the CTC forced-alignment branch appears inactive or untrained, making the generated timestamps unreliable.

From the current artifacts, timestamp inference can run, but the alignment quality does not seem to reflect a properly trained CTC decoder.

If correct timestamp is not yet supported in the open-source release, it may help to document this limitation explicitly to avoid confusion.

Steps to Reproduce

See issue #70

Error Message / Traceback

[LauraGPT]

Thank you for the detailed report. You are correct.

Current status of timestamps in Fun-ASR-Nano:

The open-source model.pt checkpoint does not include trained CTC decoder weights (ctc_decoder.*, ctc.*). The CTC forced-alignment branch exists in the code architecture but the corresponding weights are not in the released checkpoint. This means:

• The model loads and runs without errors (missing keys are logged as warnings)
• Timestamp output may be returned but is not reliable (untrained CTC alignment)

Workaround for accurate timestamps:

If you need accurate character-level timestamps now, use Paraformer which has a fully trained timestamp model:

from funasr import AutoModel
model = AutoModel(model="paraformer-zh", vad_model="fsmn-vad", device="cuda")
res = model.generate(input="audio.wav")
# res[0]["timestamp"] contains accurate character-level timing

For Fun-ASR-Nano, the vLLM offline service (serve_vllm.py) provides segment-level timestamps via VAD boundaries, which are accurate:

# Segment-level timing from VAD (accurate)
{"text": "...", "start": 1.7, "end": 5.5}

We will update the documentation to explicitly note this limitation. Trained CTC weights may be released in a future update.

[GeorgeCaoJ]

Got it. Thanks for the clarification and for open-sourcing the great model.

FunASR-Nano combined with VAD and a vLLM-served Qwen3 forced aligner can provide accurate word-level timestamps.