HeartMuLa Transcription Loader

The HeartMuLa Transcription Loader is a one-input loader whose only job is to bring up the Whisper-based HeartTranscriptor model so the HeartMuLa Lyrics Transcriber can run. If that sounds like a lot of machinery for one model, you're right - but it mirrors how the generation side works, where loaders and runners are separate nodes, and it keeps the model loaded once instead of re-reading it per transcription.

You'll reach for it in two situations. The obvious one is transcribing an existing track to get lyrics back out. The sneaky one is more fun: transcribe a reference song, feed the resulting text into the Music Generator as lyrics, and get a track that lands on the same words with different everything else. For that loop, this loader is the front door.

What it does

One input, one output:

• base_path - default HeartMuLa, pointing at ComfyUI/models/HeartMuLa/. Same 📁 folder-browser button the other loaders get.

Output is a HEARTMULA_TRANSCRIPTOR socket that connects to the Lyrics Transcriber's transcriptor input. That's genuinely all it does - no version dropdown, no options. The model runs fp16 on CUDA if available, fp32 otherwise, and it loads from disk once per execution (it isn't a session singleton like the MuQ loader in this pack; the code moves it to CPU after each transcribe run).

The model file

This one's optional in the pack's default layout - the README lists it under "optional extras" - so nothing auto-downloads and nothing errors if you skip it until you actually want transcription:

cd ComfyUI/models/HeartMuLa
git clone https://huggingface.co/HeartMuLa/HeartTranscriptor-oss

The loader resolves models/HeartMuLa/<base_path> and throws a FileNotFoundError if the folder isn't there, so a missing clone is the #1 cause of "the transcription loader won't load."

Install

Same as every node in this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Crono141/ComfyUI-JK-HeartMuLa.git
cd ComfyUI-JK-HeartMuLa
pip install -r requirements.txt

Restart ComfyUI and you'll find it under JK-HeartMuLa (a recent ComfyUI with the V3 node API is required - this is a V3-API pack). The transcription stack is part of the bundled heartlib, so the usual torchao/torchtune/accelerate requirements cover it; nothing extra to install on top.

Honest expectations

Whisper-grade transcription means it's good at words and indifferent to nuance. It doesn't timestamp anything, it won't tell you who sang what, and vocals buried under dense production can come back as cheerful nonsense - which is a Whisper trait, not this node's bug. If you're transcribing your own generated HeartMuLa track, the words are usually clean because they're the exact lyrics you fed the generator in the first place. Feed in someone else's dense mix and read the output with a grain of salt before you build a lyric sheet on it.