Skip to content

models.md

Latest commit

 

History

History
83 lines (66 loc) · 3.96 KB

models.md

File metadata and controls

83 lines (66 loc) · 3.96 KB
Raw

Model licensing notes

Records the license basis for ML checkpoints StemDeck downloads at runtime. Not exhaustive -- only entries where the decision wasn't a simple upstream license file are documented here.

Demucs (htdemucs_6s)

MIT, published by the demucs PyPI package (Meta/Facebook Research). No audit needed -- an unambiguous upstream license.

All-In-One (automatic song sections)

StemDeck runs this model on CPU after separation and passes its existing stems. The checkpoint is not bundled in StemDeck installers.

Whisper (lyrics transcription)

  • Runtime: openai-whisper (PyPI, MIT, openai/whisper), on the torch StemDeck already ships for Demucs.
  • Checkpoints: large-v3-turbo (1.6 GB) on a GPU with room for it, small (0.5 GB) on a CPU or a smaller GPU. Overridable with STEMDECK_WHISPER_MODEL_GPU / STEMDECK_WHISPER_MODEL_CPU.
  • License: MIT for both the code and the published weights.
  • Upstream: https://github.com/openai/whisper

Runs only when no lyrics were found for a job and the transcribe_lyrics setting allows it (by default on a GPU only). Weights download on first use into <TORCH_HOME>/whisper, or during desktop setup when transcription would run, and are verified by Whisper's own SHA-256 check. Not bundled in StemDeck installers.

UVR-MDX-NET Karaoke 2 (on-demand lead/backing vocal split, #275)

  • File: UVR_MDXNET_KARA_2.onnx
  • Distributed via: audio-separator (PyPI, MIT, nomadkaraoke/python-audio-separator), which bundles/downloads models trained as part of the Ultimate Vocal Remover (UVR) project by Anjok07.
  • License: MIT + attribution, per the audio-separator README:

    "If you choose to integrate this project into some other project using the default model or any other model trained as part of the UVR project, please honor the MIT license by providing credit to UVR and its developers."

  • Credit: Ultimate Vocal Remover (Anjok07) -- https://github.com/Anjok07/ultimatevocalremovergui

This is the shipped default (VOCAL_SPLIT_MODEL in app/core/config.py, overridable via STEMDECK_KARAOKE_MODEL).

Rejected alternative: mel_band_roformer_karaoke (aufr33/viperx)

audio-separator's community-trained roformer checkpoint (mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt) has meaningfully better reported SDR than the MDX-Net Karaoke 2 model above, and was the originally preferred choice while scoping this feature. It was rejected after directly verifying:

  • No LICENSE file was ever published for this checkpoint, nor a stated license anywhere in its distribution.
  • It was originally released through UVR's Boosty supporter-paywall page, not as a public open release.
  • The public Hugging Face mirror (jarredou/aufr33-viperx-karaoke-melroformer-model) now returns 401 (gated/removed) -- confirmed directly, not secondhand.

A cleanly-licensed roformer alternative (Kimberley Jensen's Mel-Band-Roformer-Vocal-Model) was also checked and found to have no LICENSE file despite a claim to the contrary surfacing in a web search.

Since StemDeck is free/non-commercial, the practical risk of using an unlicensed-but-freely-shared community checkpoint is low -- but a Boosty-paywall origin is a step past "unlicensed," suggesting the author did not intend it for free redistribution at all, and its already-dead HF mirror makes it an unreliable thing to depend on regardless of the licensing question. STEMDECK_KARAOKE_MODEL remains available as an env override for a deployment that wants to accept that risk itself.