SupaVoice HF Whisper Transcriber

If you're feeding a voice-cloning pipeline like VoxCPM, your transcript is your training data - a mistranscribed label quietly teaches the model the wrong thing. That's the gap this node exists for. While the sibling Local Whisper Transcriber is the cheap daily driver, SupaVoice HF Whisper Transcriber is the one you reach for when the audio is Burmese and you want a shot at accurate Myanmar script. It's the only node in the SupaVoice pack with access to the chuuhtetnaing Whisper fine-tunes trained for Burmese.

What it does

It takes a ComfyUI AUDIO in, runs it through a Hugging Face transformers speech-recognition pipeline, and hands you a transcript string plus saved .txt and .json files. Under the hood it converts your audio to a mono WAV, loads the model you picked, and chunks the file every chunk_seconds (default 20, range 5–60) so long clips don't blow up. When language is set to my, it passes language="myanmar" to the model and asks for task="transcribe" - no translation, just straight script.

The model_id dropdown is the whole story:

• chuuhtetnaing/whisper-small-myanmar and whisper-medium-myanmar - the sweet spot to try first (they're real HF repos; the author admits the cards report high WER overall but says small/medium are worth local testing).
• chuuhtetnaing/whisper-large-v3-myanmar - the big Burmese bet, slow and heavy.
• openai/whisper-large-v3-turbo and openai/whisper-large-v3 - generic Whisper when you don't need Burmese specifically.

language defaults to my; set auto to let the model guess or en for English. output_prefix just names the saved files.

Outputs: transcript (plain text, wire it straight into a VoxCPM prompt/training field), txt_path, and json_path (the full result payload with timestamps).

Installing it

Same as the rest of the pack - ComfyUI Manager (search comfyui-supavoice-transcriber) or:

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

Restart ComfyUI after. First transcription downloads the chosen model from Hugging Face, so budget some time (and disk) - large-v3 is a multi-GB pull. The node caches under E:\models\huggingface by default, and all transcripts land in E:\SupaVoiceData\comfy_transcripts unless you set the SUPAVOICE_DATA_ROOT env var - which you should if you're not on the author's Windows box, because on Linux that literal E:\ string becomes a weird relative directory.

Where people get burned

The heaviest node in the pack: it pulls transformers plus a real model into your ComfyUI env, and the Myanmar fine-tunes are as inconsistent as the author warns. Test small on your actual clips before assuming it beats generic Whisper. Also remember this is the accurate-but-slow path - if you just need a fast English first pass, the Local Whisper Transcriber gets you 80% of the way for free. For stubborn Burmese, though, the OpenRouter/Gemini transcriber is the author's own expected winner; this node is the no-API-key middle ground.