Skip to Content
Documents⚡ Core FeaturesQwen3-ASR Engine

Qwen3-ASR Engine

In addition to Whisper, OnceScribe ships with the Qwen3-ASR audio model. It is optimized for Chinese, dialects, and Chinese-English code-switching, and like Whisper it runs entirely on your device — no uploads.

When to Use It

  • Mandarin, Chinese dialects, or Chinese-English mixed audio/video
  • Scenarios where Chinese accuracy matters more than Whisper provides
  • Live recording transcription (the workbench live mode uses the Qwen3 engine by default)
  • Everyday transcription with a smaller local model

Model Selection

Switch the ASR engine to Qwen3-ASR in Settings → Model and download the model you want:

ModelSizeNotes
Qwen3-ASR 0.6B~1.7 GBLightweight and fast, recommended for most devices and daily use
Qwen3-ASR 1.7B~4.5 GBHigher accuracy on complex audio, uses more disk space, Pro only

The first download requires an internet connection. After that, the model works offline.

Language Support

Qwen3-ASR is officially trained for Chinese and English:

  • Chinese content: set language to zh for best results
  • English content: set language to en
  • Other languages: try auto, but accuracy may be lower than Whisper

The language list adjusts automatically when you switch engines.

Prompts & Vocabulary

Qwen3-ASR supports a decoding prompt to guide recognition:

  • Prompt: product names, person names, domain terms, up to 512 characters
  • Vocabulary: add frequently used proper nouns to the prompt path to improve brand and technical-term recognition

Prompts apply to both file transcription and live recording.

Frame-Accurate Subtitles

Qwen3-ASR outputs text first; subtitle timestamps are estimated by default. For frame-accurate timing:

  1. Go to Settings → Model
  2. Download the optional Qwen3-ForcedAligner-0.6B (~1.3 GB)
  3. SRT/VTT exports will be automatically aligned

Without the Forced Aligner, exported subtitles still work but use estimated segment boundaries.

Limitations

Qwen3-ASR does not have the exact same capabilities as Whisper. Keep these differences in mind:

  • No speaker diarization: switch back to Whisper when you need speaker labels
  • No translate-to-English mode: Qwen3 outputs in the original language
  • Unsupported parameters: temperature, context window, Beam search, word-level timestamps
  • Live recording: supports fixed-interval or VAD-based streaming modes

FAQ

Q: Which is better, Qwen3 or Whisper? A: Qwen3-ASR is usually more accurate for Chinese. Use Whisper for multilingual content or speaker diarization.

Q: Why is the 1.7B model Pro only? A: The 1.7B model is larger and more resource-intensive, so it is reserved for Pro users. The 0.6B model is free to use.

Q: Can I install both Whisper and Qwen3 locally? A: Yes. They are stored in separate folders, and you can switch engines at any time in Settings.

Last updated
TranscriptionRecording