Whisper Model Loader

The unglamorous half of a two-node setup

Whisper Model Loader is the node that earns its keep before you ever look at it. Its whole job: fetch a Whisper model from Hugging Face on first use, cache it in ComfyUI/models/whisper, and hand it to the pack's Whisper Transcribe node. You'll never preview its output, and you'll definitely notice when you try to transcribe without it - the transcribe node demands a WHISPER_MODEL value, and nothing else in ComfyUI can provide one.

It's part of the Salt AI Whisper Transcriber pack, three nodes that turn audio and video into text from inside ComfyUI. The pack is a community mirror of Salt AI's original repo - the company pulled its public GitHub down in 2024, and the code survives because someone PR'd it into the Comfy Registry. That's the whole story behind the odd PR-SaltAI_Whisper_Transcriber-e087b361 name; "PR" here is registry-mirroring convention, not a review flag.

How it works

A thin wrapper around Hugging Face. The node calls AutoModelForSpeechSeq2Seq.from_pretrained(model, cache_dir=models/whisper, use_safetensors=True).to(device) and loads a matching AutoProcessor. First run downloads the weights - a few hundred MB to roughly 3 GB depending on the model - and every run after that reads from disk. The output is one WHISPER_MODEL value that only Whisper Transcribe accepts, so wiring is a single line: loader → transcribe. (The socket is labelled "model, processor," but it's one output - don't go hunting for a second one.)

The two dropdowns

model lists eight choices: the openai/whisper-* line from tiny up through large-v2 and large-v3, plus distil-whisper/distil-large-v3. device is cuda (default) or cpu.

Which model? Same sizing rule as everywhere: accuracy scales with size, and so does VRAM.

• tiny, base, small - fast, run on weak cards, but they mangle accents and background noise.
• medium, large, large-v2, large-v3 - the ones that transcribe reliably; large-v3 is the best all-rounder for English.
• distil-whisper/distil-large-v3 - the one I'd reach for first. A distilled Whisper, roughly half the size of large-v3 and several times faster at near-equal accuracy. That's the standard distillation trade - speed for a sliver of quality - and for transcription it's usually worth taking.

Install

ComfyUI Manager → search "SaltAI Whisper", or:

cd ComfyUI/custom_nodes
git clone https://github.com/haohaocreates/PR-SaltAI_Whisper_Transcriber-e087b361

Then install requirements.txt into the Python ComfyUI uses (portable Windows: ComfyUI\python_embedded\python.exe -m pip install -r requirements.txt), restart, and refresh the browser tab. ffmpeg and ffprobe also need to be on your PATH because the transcribe node shells out to them.

Gotchas

First run downloads into models/whisper, so a hang there is almost always a Hugging Face connectivity issue - retry or fix your proxy. Picking cuda with no working GPU is a hard crash; use cpu if that's your situation (slow, but fine for short clips). OOM on large-v3? Drop to medium or distil instead of fighting your card.

And the same mirror quirk as the rest of this pack: the code does from . import logger, but the repo is missing logger.py, which can make the whole pack fail to load with a ModuleNotFoundError. Fix: create logger.py in the pack folder containing import logging / logger = logging.getLogger(__name__), then restart.