Load faster-whisper

This is the node that actually pulls a Whisper model into memory - everything else in the faster-whisper side of this pack depends on it. It wraps SYSTRAN's faster-whisper, the CTranslate2 reimplementation of OpenAI's Whisper that trades nothing on accuracy (same weights) for a real speed win, especially with quantization turned on.

How it works

CTranslate2 is an inference engine built for exactly this kind of workload - it re-implements the transformer forward pass with aggressive kernel fusion and supports quantized weights (int8 and friends) natively, which is where most of the speed comes from versus the reference PyTorch Whisper implementation. This node configures and instantiates that engine, downloading the converted model weights from Hugging Face on first use if they're not already cached locally.

The inputs and outputs that matter

• model_size - which checkpoint to load. The dropdown here is scoped to large-v3, so if you're expecting a tiny/base/small ladder like stock Whisper, this pack's loader is set up around the large model specifically - makes sense for a toolkit aimed at building clean datasets rather than fast live captioning, where you'd usually trade accuracy for speed with a smaller model instead.
• device - auto, cpu, or cuda. auto is the safe default; pick cpu deliberately if you want to keep the GPU free, or cuda to force it and fail loudly if no GPU is available (better than a silent CPU fallback if you're benchmarking).
• compute_type - the quantization/precision CTranslate2 runs at: default, auto, int8, int8_float32, int8_float16, int8_bfloat16, int16, float16, float32, bfloat16. This is the actual speed/quality/memory knob. int8 variants are dramatically faster and lighter, especially on CPU, at a small accuracy cost; float16 is the standard GPU choice; float32 if you want maximum fidelity and have the VRAM to spare.
• cpu_threads (default 0 = let CTranslate2 decide) and num_workers (default 1) - tune these if you're running multiple transcriptions in parallel on CPU; most single-job setups can leave them alone.
• model (output, FASTER_WHISPER) - feeds directly into SDT_FasterWhisperTranscribe.

How to install it

ComfyUI Manager: search ComfyUI-speech-dataset-toolkit, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kale4eat/ComfyUI-speech-dataset-toolkit
cd ComfyUI-speech-dataset-toolkit
pip install torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

This node is the heavy one in the pack: large-v3 is a multi-gigabyte download the first time you run it, and it needs real VRAM (or patience, on CPU) to run at a decent clip.

Common issues & troubleshooting

First run hangs or takes forever. That's the model download, not a stall - large-v3 is big, and there's no progress bar surfaced in the node UI. Watch your console/server logs the first time.

Out of memory on GPU. Drop compute_type to an int8 variant - int8_float16 is a solid middle ground on CUDA - before you assume you need more VRAM. The quantized modes exist specifically to make large-v3 fit on smaller cards.

CUDA device selected but it's clearly running on CPU (slow). Confirm your onnxruntime/torch CUDA install actually matches your driver - if the underlying CTranslate2 CUDA build can't initialize, some setups silently fall back rather than erroring. Check device is actually respected in your logs.