Qwen3-ASR for ComfyUI

Nodes on GitHub · Drama 3 for ComfyUI · The original model

Qwen's Qwen3-ASR speech recognition, one file per size, for the native nodes in ComfyUI-Drama3: speech to text in 30 languages and 22 Chinese dialects, singing included. In that pack it writes a voice clip's transcript straight into Drama 3's Encode Voice, and it works on any audio. Qwen3-ForcedAligner gives the time of every word, so the Drama 3 Editor's captions follow the voice word by word. The nodes run both on ComfyUI's own Qwen3 decoder with a native audio encoder: no extra Python packages.

Files

All go in ComfyUI/models/asr. A 6.6 s clip takes ~2 s with the 1.7B and ~1 s with the 0.6B on an RTX 3090, model load included.

hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_asr_1.7b_bf16.safetensors --local-dir ComfyUI/models/asr

Then add Load Qwen3-ASR and Qwen3-ASR Transcribe in ComfyUI: language detects the language or forces one, context takes names and terms that help recognition, and the node returns the text and the language.

For word timing, load the aligner with Load Qwen3-ASR and connect it to the Drama 3 Editor's aligner input:

hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_forced_aligner_0.6b_bf16.safetensors --local-dir ComfyUI/models/asr

How close it is to Qwen's code

Against Qwen's qwen-asr package on five clips and three prompt variants per size: the prompt ids are identical, the encoder output matches within 1e-6 in fp32, the first generated token is the same, and the transcripts match on clips up to 8 s. On longer clips the encoder attends in Qwen's 8 s windows, as their vLLM backend does (their transformers backend attends across the whole clip), which can move a comma.

The aligner, against Qwen's own on a 40 s take: 121 of 131 words get the same times, and every other word is within 0.16 s.

How these files were made

The original checkpoints' bf16 tensors, unchanged, in one file each, with the pack's tools/convert.py asr: the audio encoder under audio_tower. and the decoder under decoder.model.. The LM head is left out because it is the embedding table (checked tensor for tensor), and the model config is stored in the file's metadata. The aligner keeps its head, which classifies each word's time in 80 ms steps, as timestamp_head.weight.

License

Apache License 2.0, as the original models (LICENSE, NOTICE). Qwen3-ASR is by the Qwen team; this repository is not affiliated with or endorsed by it.