AI and responsibility
Songs, voices, speech, covers and lyrics from Tonwerk-AI are generated by AI models. This page says which models those are, how the files are labelled and what you need to keep in mind when you publish any of it.
This is an orientation, not legal advice. If you make money with AI songs or use other people's voices, get advice from a lawyer when in doubt. The German original of this page is authoritative.
What is AI-generated
- Songs
- YuE2 and ACE-Step compose and sing from style, lyrics and score. Versions with another voice (Seed-VC), sung voices (SoulX-Singer) and stems are made or processed by models too.
- Speech
- The speech lab speaks text with speech models, if you like with a cloned voice, and cleans recordings.
- Pictures
- Painted covers come from image models (FLUX, Z-Image, Krea). Photos you choose yourself remain what they are.
- Text
- Lyrics, sketches and the answers on the Write page are written by a language model, locally in LM Studio or, with your own key, Claude by Anthropic. The app says there that the answers come from an AI model.
Labelling of the files
Since 2 August 2026 the EU's AI Act (Art. 50) requires AI-generated audio, image and video files to be recognisable as such in machine-readable form. Tonwerk-AI writes this into these files:
- Audio (exports as MP3, M4A, FLAC, the player's copies, downloads of songs and versions, a song's stems, spoken and sung takes, swapped voices): a field of its own
AI_GENERATEDwith the model and the fieldDIGITAL_SOURCE_TYPEwith the IPTC termtrainedAlgorithmicMedia, plus a sentence in the comment. - Music videos: the same information in the video's metadata, also in the small copy for sharing.
- Painted covers: the IPTC term
trainedAlgorithmicMediain the XMP data, which photo services and platforms read.
Not labelled are your own recordings that Tonwerk-AI only processes (cleaned speech, stems of an uploaded file), your photos, and MIDI and scores, which contain only notes. The recording in the Logic project does not carry the label yet; that will follow.
Metadata is lost when a file is re-encoded, trimmed or re-recorded. Tonwerk-AI does not yet have an inaudible watermark. Do not remove the label when you pass files on.
When you publish
- Disclose. Anyone who publishes AI songs, AI voices or AI pictures that may appear real must say that they were made with AI (Art. 50 (4) of the AI Act; purely private use is exempt). Streaming services, distributors and platforms such as YouTube also ask for their own declarations on upload.
- Other people's voices only with consent. Voice swap, speech lab and singing can imitate voices. Use only your own voice or that of a person who has agreed; a voice is part of a person's personality rights. Rebuilding the voices of well-known singers is off limits.
- Copyright. A song that a model made without a creative contribution of your own is, in Germany, generally not protected by copyright (§ 2 (2) UrhG requires a human creation). Your own lyrics, melodies and arrangements can be. Models can resemble existing works or reproduce lines of lyrics; check lyrics and melodies before you publish them.
- The models' licences. What you may do with a result depends on the model, see below. Some allow only non-commercial use or require attribution.
- Responsibility. The models run on your Mac and with your input. You are responsible for what you generate and publish, not me.
Models and licences
Tonwerk-AI ships no models; the server downloads them from their sources during setup, and their licences apply. This overview does not replace reading the licence: licences change, and the linked page is what counts.
| Model | Used for | Licence | Commercial use of results |
|---|---|---|---|
| YuE2 | Songs | CC BY-NC 4.0 with addendum | Expressly yes for individuals; credit YuE2. A LoRA of your own is a modification of the weights and may not be sold. |
| ACE-Step 1.5 | Songs | MIT | yes |
| Seed-VC | other voice | GPL-3.0 | the licence covers the software, not the results; the voice needs consent |
| SoulX-Singer | Singing | Apache 2.0 | yes |
| Higgs Audio v3 | Speech lab | Boson Higgs TTS 3 Research and Non-Commercial License | non-commercial only; exception “Creator Use Grant” for your own content such as podcasts and videos, if Boson AI is credited |
| Chatterbox, Qwen3-TTS, Higgs Audio v2, MOSS-TTS, MossFormer2, DeepFilterNet | Speech lab | different per model | check on the model's page before publishing |
| Stem models (mlx-audio-separator) | Stems | different per model, some without a clear statement | check |
| FLUX.2 Klein 4B, Z-Image Turbo | Covers | Apache 2.0 | yes |
| FLUX.2 Klein 9B | Covers | FLUX Non-Commercial License | no; content filter required (built in) |
| Krea 2 Turbo | Covers | Krea 2 Community License | up to USD 1 million in revenue; content filter (built in) and, where required by law, an AI notice |
| FLUX 3 Klein | Covers | FLUX Community License | up to USD 5 million in revenue; filter (built in), AI notice; not for training other image models |
| Whisper large-v3 turbo | Timestamps in the lyrics | MIT | – |
| Basic Pitch, SheetSage2 | Audio to MIDI, transcription | Basic Pitch Apache 2.0; SheetSage2 see project page | – |
| Language models in LM Studio | Text | that of the chosen model | see the model's page |
| Claude (Anthropic) | Text, translation | Anthropic Commercial Terms | under these terms the results belong to you |
Tonwerk-AI itself
The app and the server are under the PolyForm Noncommercial License 1.0.0: free for any non-commercial purpose, without warranty as far as the law permits. For commercial use of the software, ask me. The third-party libraries in the app and their licences are listed in NOTICE. The software's licence says nothing about what you may do with generated songs; the models above govern that.
As of 23 October 2026