Qwen3-ASR Loader

The Qwen3-ASR Loader is the least glamorous node in the SynVow pack and the one most likely to make you think something's broken. The first time you hit "run," it does nothing visible for a while, and then a multi-gigabyte model lands in your models folder. That's by design. This node pulls one of Alibaba's Qwen3-ASR speech-recognition models off HuggingFace and parks it in VRAM so the rest of the graph can actually transcribe. Everything else in the pack depends on it; nothing runs without it.

What it does

One input, one output, one decision. The loader has a single model_name dropdown:

• Qwen3-ASR-1.7B (default) - the bigger, more accurate model. Realistically ~8GB VRAM once loaded.
• Qwen3-ASR-0.6B - noticeably lighter (~4GB VRAM), faster to download and to run. Plenty for clean audio.

That's the whole node. Out the other side comes a single model connection of type QWEN3ASR_MODEL, which only wires into the pack's Qwen3-ASR Transcribe node. It's a loader in the same boring, essential sense a checkpoint loader is: there's nothing to configure and nothing to admire, but the downstream node is a paperweight without it.

How it works

First run is the interesting part. The loader checks ComfyUI/models/Qwen3-ASR/<name> and, if it's empty or missing, calls snapshot_download to pull the matching repo (Qwen/Qwen3-ASR-1.7B or Qwen/Qwen3-ASR-0.6B) from HuggingFace into that folder. Later runs just load from disk. The model then opens with Qwen3ASRModel.from_pretrained in bfloat16 pinned to cuda:0 - and both of those are hardcoded in the source.

That's the gotcha of this whole pack. You need a CUDA GPU, and a modern-ish one: bf16 on a GTX 16-series card is a wall of half-working precision, and there's no CPU fallback baked in. If you're not on a recent NVIDIA card, this node simply won't start.

Installing

The pack ships four nodes; this is one of them. Install the whole thing:

ComfyUI Manager → search "Comfyui_SynVow_Qwen3ASR" → Install → Restart

or, if Manager can't find it:

cd ComfyUI/custom_nodes
git clone https://github.com/shumoLR/Comfyui_SynVow_Qwen3ASR
cd Comfyui_SynVow_Qwen3ASR
pip install -r requirements.txt   # qwen-asr, huggingface_hub, torchaudio

Then restart ComfyUI and load the model once. The 1.7B download is several gigabytes, so budget some patience on the first queue - or pre-seed the folder yourself by downloading the HF repo into ComfyUI/models/Qwen3-ASR/Qwen3-ASR-1.7B/.

Common issues

• The run just sits there. That's the download, not a freeze. Watch your network, not the node.
• CUDA errors at load. Remember: cuda:0 and bf16 are non-negotiable here. No CPU mode.
• VRAM bloat. Loading the 1.7B ASR model and the 0.6B forced aligner in the same session is ~12GB before your video pipeline even starts. On an 8GB card, stick to the 0.6B ASR model or run alignment separately.

One honest aside: at 0 impressions on comfy.icu, this is a fresh, lightly-adopted pack - expect rough edges. DarioFT's Qwen3-ASR nodes are the better-known alternative in the community if this one fights you.