How your music is used

1. 01

   You choose a file. It is hashed in your own browser before anything is sent, so what you signed for and what arrives can be compared later.

2. 02

   It uploads straight into private storage. It does not pass through this website’s server, and it never goes into a public folder or a shared link.

3. 03

   The stored file is hashed again and checked against what you signed. If the two disagree it is quarantined instead of used.

4. 04

   Existing transcription software is run over it, and its output is compared against your lyrics to find what it got wrong.

5. 05

   A person reads the result and corrects it by hand. Patois spelling and meaning are recorded in separate notes, never written over your words.

6. 06

   Combined results across all songs are reported with no audio, no lyric excerpts, and nothing identifying you.

Testing means running software that already exists over your recording to measure what it gets wrong. Nothing about the software changes. Training means using your recording to alter the software itself. The base permission covers testing only. Training is a separate box, it starts unchecked, and even if you check it no training happens: the terms that would have to apply to a trained model have not been settled, so it stays switched off across the whole project.

As it stands, no model training happens on any contributed track, whatever anyone has ticked. The website contains no training code and the project switch is off.

• Store the recording and the lyrics you provide in private storage that only the researcher can open.
• Run existing speech to text software on the recording to measure how accurately it transcribes your words.
• Align your lyrics to the recording and correct transcription errors by hand.
• Record notes about Patois spelling, pronunciation and meaning, kept separate from your original lyrics.
• Report combined results across all contributed songs in a form that does not identify you and contains no audio and no lyric excerpts.

• Adapting or fine tuning a speech to text model using this trackoff by default

  This is model training, not testing. Testing measures what existing software already does. Training changes the software using your recording. You can decline this and still take part in the evaluation.

• Publicly thanking you by artist nameoff by default

  Your artist name only. This does not publish your song name, your lyrics, any clip, or any result tied to you.

• No generating music, vocals or instrumentals.
• No voice cloning and no imitation of any artist.
• No selling, licensing or distributing your recording.
• No public dataset containing your audio or your lyrics.
• No public release of model weights trained on your track.
• No commercial use.
• No sublicensing to another company for their own purposes.
• No use of your track to promote anything.

There is no payment budget for this pilot. Taking part is unpaid and voluntary, and that is stated in the agreement rather than buried.

For a track that is accepted and gets through review, the intended result is a reviewed lyric transcript and a short summary of what the transcription software got wrong on your song.

Timed lyric files, the kind used for synchronised captions, are experimental. They are provided where it is feasible, after a person has checked them by hand.

There is no automated processing here and no instant turnaround. You get the signed agreement straight away. Anything else comes later, through manual work, with no promised date.