OmniVoice Whisper Loader

A tiny utility node with one job: load a Whisper speech-to-text model once and hand it to the other OmniVoice nodes, so they stop re-downloading and re-loading Whisper every time you clone a voice. That's it. No synthesis, no knobs that matter, no mystery. If you only use Voice Design (which needs no transcription), you never touch it. The moment you're cloning voices - Voice Clone, Longform with a reference clip, or Multi-Speaker dialogue - you want it in the workflow.

Here's the annoyance it solves: when you clone a voice and leave the reference transcript empty, the TTS node auto-transcribes the clip with Whisper first, because it needs the transcript to pair with the audio. Do that repeatedly and a fresh Whisper model gets pulled from HuggingFace and loaded on each run - slow, and on a metered connection it adds up. The loader fixes both: connect its whisper_model output to the whisper_model input on the clone, longform, or multi-speaker node, and the ASR model stays cached between runs. The README literally lists "Whisper re-downloads every run" as a troubleshooting entry, and this node is the fix.

Setup is three dropdowns. model picks the ASR model: whisper-large-v3-turbo (the default, ~1.5GB VRAM, the sensible starting point), whisper-large-v3 (~3GB), whisper-medium (~1GB), whisper-small (~0.5GB), or whisper-tiny (~0.4GB) - all marked "(auto-download)" and fetched to ComfyUI/models/audio_encoders/ on first use, or drop your own model folder there and it shows up in the list. device (auto/cuda/cpu) and dtype (auto/bf16/fp16/fp32) follow the same auto-logic as the TTS nodes: bf16 on Ampere+ GPUs, fp16 on older ones, fp32 on CPU. That's the entire input surface. The one output is whisper_model (WHISPER_ASR).

One thing worth knowing before you panic: this node is deliberately the one that always loads. If you install the pack and see only OmniVoice Whisper Loader in your node list while every TTS node is missing or red, that's not this node being special - it's omnivoice failing to import at startup, so the pack's other nodes don't register. Check your ComfyUI log for the [OmniVoice] omnivoice import failed: line; the usual culprits are a transformers version too old for the model (needs 5.3+; look for the HiggsAudioV2TokenizerModel import error) or a missing soxr, both fixed with pip and a full restart. Install the pack itself via ComfyUI Manager (search "OmniVoice") or git clone https://github.com/Saganaki22/ComfyUI-OmniVoice-TTS plus python install.py, and let the installer do its --no-deps thing so the omnivoice package can't downgrade your PyTorch. Once it's up, connect the loader, clone away, and you'll never watch Whisper re-download again.