Licences and attribution

Last updated 9 October 2026

rythmify is built on open-source software and freely licensed models. This page names them and their licences. Full licence texts are available on request.

Models used for every song

• MVSep Mega 53-stem BS-RoFormer (voice and instrument separation) — weights and model code © Roman Solovyev (ZFTurbo), MIT License, released under the MIT License by the author on 25 September 2026 (statement).
• Beat This! (beat tracking) — Foscarin, Schlüter and Widmer, ISMIR 2024; code and weights © CPJKU, MIT License.
• Whisper large-v3 (transcription, when no words are supplied) — © OpenAI, MIT License; CTranslate2 conversion © SYSTRAN, MIT License.
• wav2vec 2.0 large (LV-60k, 960 h) (English word timing) — © Facebook AI Research, MIT License.
• Conformer-CTC Large, Croatian (Croatian word timing) — © NVIDIA, CC BY 4.0, huggingface.co/nvidia/stt_hr_conformer_ctc_large; trained on ParlaSpeech-HR (Ljubešić et al., 2022).
• pYIN (sung pitch) — as implemented in librosa, ISC License.

Software

PyTorch (BSD-3), torchaudio (BSD-2), Transformers (Apache-2.0, © Hugging Face), faster-whisper and CTranslate2 (MIT, © SYSTRAN), librosa (ISC), SoundFile (BSD-3) with libsndfile (LGPL-2.1+), NumPy (BSD-3), python-soxr (LGPL-2.1+) with libsoxr, einops and rotary-embedding-torch (MIT), Pyphen (MPL-1.1) with its hyphenation dictionaries (en_US/en_GB © László Németh et al., BSD-style; hr_HR © Igor Marinović, LGPL), Unidecode (GPL-2.0+), mutagen (GPL-2.0+), httpx (BSD-3), FastAPI (MIT), Uvicorn (BSD-3), FFmpeg (GPL-2.0+, Debian build; source at ffmpeg.org). The web page is built with Vite (MIT) and TypeScript (Apache-2.0). Fonts Fraunces, Instrument Sans and DM Mono are used under the SIL Open Font License 1.1.

Apache-2.0 components: licensed under the Apache License, Version 2.0; you may not use these files except in compliance with the License (apache.org/licenses/LICENSE-2.0).

Optional components

Some accounts may have extra features switched on, which use further components: Demucs (© Meta Platforms, code MIT; weights for research use only), Meta MMS forced aligner (CC BY-NC 4.0), ParlaSpeech-HR wav2vec 2.0 (CLASSLA), yt-dlp (Unlicense), and lyrics data from LRCLIB. These are not used for the public service.