SupaVoice Local Whisper Transcriber

This is the node you'll actually run most of the time. SupaVoice Local Whisper Transcriber turns any AUDIO into a transcript using faster-whisper - fully local, free, no API key, no network beyond the one-time model download. If you're building voice-cloning training labels (the README's whole reason for existing) and your audio is mostly English, this is your workhorse. It's also the author's recommended first pass for Burmese, with the caveat that the Gemini path is the accuracy upgrade when local Whisper chokes.

How it works

The node writes your incoming AUDIO to a temp mono WAV, then loads a faster-whisper model - on CUDA if available, CPU otherwise - and transcribes with VAD filtering on and a configurable beam_size. Segments get joined into one plain-text transcript, and the full per-segment payload (with timestamps and language probability) goes to a saved .json alongside a .txt.

The settings that matter:

• whisper_model - small, medium, large-v3, or large-v3-turbo (default). Turbo is the sane default: near-large quality at roughly half the speed of the big one.
• compute_type - float16 (default) for GPU, int8_float16 when you want to cut VRAM, int8 for CPU. If you're on CPU, don't fight it - go int8 and grab a coffee.
• language - my (default), en, or auto. If you're not doing Burmese, flip this to en or auto; forcing my on English audio will do exactly what you'd expect.
• beam_size - default 5, higher is slower-but-surer. Leave it alone until you have a reason.

Outputs: transcript (plain text - wire it into VoxCPM prompt/training fields), txt_path, and json_path.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/mexxmillion/ComfyUI-SupaVoice-Transcriber

Restart ComfyUI. First run downloads the Whisper model from Hugging Face - a couple of GB for turbo - cached under E:\models\huggingface by default, with transcripts saved to E:\SupaVoiceData\comfy_transcripts. Neither path exists if you're not the author, so set SUPAVOICE_DATA_ROOT to something sane on your machine; on Linux the literal E:\ default becomes a weird relative directory.

Where people get burned

The README warns the ComfyUI env has needed a Python 3.12 compatibility shim for older setuptools - the pack ships that shim internally, so you shouldn't have to do anything, but it's worth knowing the error you'd otherwise see (pkgutil.ImpImporter gone in 3.12) has a fix baked in. Memory-wise, large-v3 in float16 will comfortably eat VRAM alongside a diffusion model - that's what the int8_float16 option is for. And keep expectations honest: the author says this is "good for local/offline English and first-pass Burmese," not a Burmese specialist. When it mangles your Myanmar clips, that's the cue to switch to the HF transcriber's Burmese fine-tunes or pay for Gemini via the OpenRouter node.