ComfyUI-QwenASR v1.1.0: official Qwen3-ASR models, smart ITN, and long-form forced alignment

Narrow-Particular202 · reddit · 2026-09-12

ComfyUI-QwenASR v1.1.0 is out, closing the gap between raw speech recognition and usable text/subtitles in ComfyUI:

Transformers 5 + official native models: legacy checkpoints deprecated; now uses Qwen3-ASR-1.7B-hf, Qwen3-ASR-0.6B-hf, and Qwen3-ForcedAligner-0.6B-hf with transformers >= 5.13.0 — pure upstream PyTorch, lower VRAM and faster on NVIDIA GPUs and Apple Silicon.
Production-ready ITN: spoken numbers, percentages, and decimals auto-converted to numerals; spaced acronyms merged ("A S R" → ASR); whitelist protection keeps idioms intact.
Hot-reloadable custom dictionary: rules live in itnrules. for custom acronyms and brand names across English, Chinese, Japanese, Korean, French — effective on the next run, no restart.
Three nodes: ASR for lightweight transcription, Subtitle for sentence chunking with one-click .srt export, and Forced Align for long-form audio (podcasts/lectures) using iterative speaking-rate windowing, with automatic transcribe-and-align when transcript is left empty.

Install steps and sample workflows in the GitHub README.

More from Multimodal