Skip to content

Qwen3-ASR Pipeline

Qwen3-ASR is an alternative ASR engine based on Alibaba's Qwen architecture, offering strong Japanese text quality with a different approach to speech recognition.

Models

Model Size Notes
Qwen3-ASR 1.7B ~4 GB Full model, best quality
Qwen3-ASR 0.6B ~2 GB Smaller, faster, slightly lower quality

How to Use

GUI (Ensemble Tab)

  1. Go to the Ensemble tab
  2. Set Pipeline to Qwen3-ASR
  3. Select model size
  4. Click Start

CLI

whisperjav video.mp4 --mode qwen

As Ensemble Pass 2

Qwen3-ASR pairs well with Whisper-based pipelines:

  1. Pass 1: Balanced (Whisper — good timing)
  2. Pass 2: Qwen3-ASR (good text quality)
  3. Merge Strategy: Smart Merge

Requirements

  • HuggingFace extra must be installed: pip install whisperjav[huggingface]
  • Requires transformers and accelerate packages
  • First run downloads the model from HuggingFace

Strengths and Limitations

Strengths:

  • Excellent Japanese text quality
  • Good handling of casual/colloquial speech
  • Strong punctuation and sentence structure

Limitations:

  • Timing can differ from Whisper-based pipelines
  • Apple Silicon: currently CPU-only (MPS not yet supported for forced aligner)
  • Requires the HuggingFace extras installed