#473·VidBee

[Bug]: Qwen3-ASR extract can finish with incomplete tokenizer, then transcription fails with Please check your config!

Author: Designer-AweiCreated Sep 8, 2026Updated Sep 8, 2026

Observed behavior and screenshots

On Windows, downloading / extracting the Qwen3-ASR (quality) model can leave the model directory looking ready (ONNX files present) while tokenizer/ is incomplete. Transcription then fails immediately.

UI shows 转稿异常 / Please check your config!

This is not a VidBee config.json problem. The message comes from sherpa-onnx failing to initialize the recognizer because tokenizer files are missing or truncated.

Expected

After the quality model is marked ready, tokenizer/ should contain the full official set, for example:

  • vocab.json (~2.6–2.8 MB)
  • merges.txt (~1.6 MB)
  • tokenizer_config.json (~12 KB)

Transcription should start.

Actual

  • Audio download succeeded.
  • Qwen3 ONNX files were present.
  • tokenizer/ only had a truncated vocab.json (~850 KB).
  • merges.txt and tokenizer_config.json were missing.
  • Retrying transcription always failed the same way.
  • Whisper Tiny in the same app was complete and usable.

The leftover archive in .downloads suggests the tarball download finished, but tokenizer extract stopped early / was treated as success:

%APPDATA%\vidbee\models\transcription\.downloads\sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25.tar.bz2 (~838 MB)

Model dir:

%APPDATA%\vidbee\models\transcription\sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25

Workaround

Manually replacing tokenizer/ with the complete files from the official package unblocks Qwen3-ASR.

Temporary alternative: switch Settings → transcription model to Whisper Tiny.

Suggested fix

  1. After extracting a Qwen3-ASR (or any non-Whisper) model, verify the tokenizer directory (vocab.json size + required sibling files) before marking the model ready.
  2. If extract is incomplete, keep the model in a failed/retry state and do not treat ONNX-only extract as success.
  3. Surface a clearer error than sherpa-onnx's Please check your config! (e.g. "ASR model files incomplete: missing tokenizer/merges.txt").
  4. Consider deleting or retrying a leftover .downloads/*.tar.bz2 when extract is incomplete.

Logs or screenshots

main.log reports VidBee 2.1.0. Transcription fails at recognizer init; the UI error is sherpa-onnx Please check your config!. Happy to attach more logs if useful. Please do not request cookie files.

App version

2.1.0

OS version

Windows 10 (10.0.19045)