EraX WoW Run

EraX WoW Run is a speech-to-text node. Feed it an audio clip and it hands you back a plain string of what was said - locally, on your own GPU, no API key, no cloud. The name might sound like a video-game plugin, but it's transcription: people keep building audio pipelines inside ComfyUI (voice to subtitle, speech into prompts, dubbing workflows) and the graph has no native way to go from sound to words. This pack is that missing piece.

The model behind it is EraX-WoW-Turbo-V1.0, a fine-tune of Whisper Large-v3 Turbo from the erax-ai team on Hugging Face, with extra training on eleven languages: English, Chinese, Japanese, Korean, Vietnamese, Hindi, Russian, German, French, Dutch, and Ukrainian. Whisper Large-v3 Turbo is already OpenAI's fast, distilled take on the big Whisper model; this is that model tuned to nail a handful of languages more reliably. If your audio is in one of those eleven, it's worth a spin. Anything else, look elsewhere.

How it works

On first run it loads the Hugging Face model folder through the transformers library, resamples whatever sample rate your audio arrives at to 16 kHz mono (handled internally, so you don't need to pre-normalize), then generates a transcript with your chosen language hard-forced into the decoder. Note the "no auto-detect" part: unlike vanilla Whisper, this node makes you pick the language up front, and the dropdown defaults to Chinese. Forget to change it and a German clip comes back as gibberish.

The inputs that matter

• audio - any node that outputs an AUDIO socket: an audio file loader, or the audio arm of a video pipeline.
• language - the dropdown. Set it. It doesn't guess.
• num_beams - 1 is greedy decoding (fast, the default); 2–5 is beam search (slower, usually cleaner text).
• max_length - a cap on generated tokens, default 200. Leave it unless long clips keep getting their tail chopped.
• unload_model - true by default: after each run it drops the model and clears VRAM. Flip it to false if you're transcribing a batch and want the model to stay warm.

The single output is text (a STRING). Wire it into a Save Text node, show it on screen, or feed it further down the graph - it's just a string.

Install

ComfyUI Manager is the easy route: search the pack title, MW-ComfyUI_EraX-WoW-Turbo. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/billwuhao/ComfyUI_EraX-WoW-Turbo.git

Restart ComfyUI. Then the model: grab EraX-WoW-Turbo-V1.0 and drop the folder into ComfyUI/models/TTS/EraX-WoW-Turbo-V1.0. Yes, that's this pack's convention - it lives under models/TTS, not a whisper folder. And it's a full Hugging Face model directory (config plus safetensors), not a single .pt file.

Common issues

• The pack won't load at all. The requirements file only installs openai-whisper, but this node hard-imports Hugging Face transformers at module load - and since both nodes in the pack load together, one missing library takes the whole pack down. Fix: pip install transformers, then restart.
• "Model not found"-style errors. Check the folder name matches exactly: models/TTS/EraX-WoW-Turbo-V1.0.
• Transcription is garbage. You left the language on the default chinese. Set it properly.

One more thing: the README announces training on "eight languages" and then lists eleven. That's the README being sloppy, not the model - trust the dropdown. This is a small, quiet pack with essentially no community write-ups behind it, so don't expect a wiki of solutions. But it's a genuinely useful trick: speech-to-text that lives entirely inside your graph.