RT HeartMuLa Transcriptor Loader

RT HeartMuLa Transcriptor Loader loads the pack's optional second model: HeartTranscriptor-oss, a speech-recognition model that turns audio back into text. It exists to feed the RT HeartMuLa Transcribe runner, so if you don't plan to transcribe anything, you can skip this node - and more importantly, skip the multi-gigabyte download - entirely.

The README is explicit about it: the transcriptor is the one model file marked "optional only download if you want." That's refreshingly honest for a custom-node pack, and worth taking literally. The core HeartMuLa pipeline (loader + sampler + preview) needs nothing from this branch.

How it works

The loader is a thin wrapper around Hugging Face's transformers ASR pipeline. Given the model name and a device, it looks for the model at ComfyUI/models/HeartMuLa/HeartTranscriptor-oss, builds pipeline("automatic-speech-recognition", model=path, chunk_length_s=30), and hands the resulting pipeline object out as a single transcriptor_pipe output.

The only two inputs are model_name (a dropdown with the single choice HeartTranscriptor-oss) and device (cuda or cpu, default cuda). There's nothing else to tune here - the interesting part is all in the runner node that consumes the pipe.

What it wires into

The transcriptor_pipe output feeds the transcriptor_pipe input of RT HeartMuLa Transcribe, which pairs it with an AUDIO input. Together they form the round-trip half of the pack: generate a song with the sampler, transcribe it back to text, and compare what the model actually sang against the lyrics you wrote.

Installing it

Install the pack normally (Manager search "ComfyUI-RT-HeartMuLa", or clone into custom_nodes and pip install -r requirements.txt). Then, only if you want transcription, clone the extra model into your models folder:

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

Everything the node needs beyond that - transformers, the pipeline machinery - comes from the pack's requirements.

Common issues

• Model not found: <path> error - you tried to load the transcriptor without downloading it. Clone HeartTranscriptor-oss into ComfyUI/models/HeartMuLa/ exactly as above. This is the loader's one real failure mode and it's a loud, clear one.
• Slow on CPU - the ASR pipeline is chunked at 30 seconds and genuinely chugs on CPU. If you're on a laptop, consider whether you actually need transcription before eating the load time.
• First load downloads nothing - unlike some wrapper packs (a sibling CRT-HeartMuLa pack auto-downloads models), this one never fetches weights for you. No download in the logs is normal; it's looking for files on disk.