NOTICE — the demo audio in this directory

mix.m4a, speaker1.m4a and speaker2.m4a were produced with OpenAudio S2-Pro
(HuggingFace `fishaudio/s2-pro`, 4B parameters, run offline through the
fish-speech 2.0.0 repository). mix.m4a is the synthesised two-speaker source
clip; speaker1.m4a and speaker2.m4a are that same mix put through this
project's own separation pipeline, untouched.

The voices are synthetic. They were cloned from the model's own reference-free
output, so no real person was recorded and no real person's voice was cloned.

mix.m4a additionally carries a real recorded noise bed at +10 dB SNR, from the
"Background noises" subset of OpenSLR SLR28 `pointsource_noises`
(noise-free-sound-0086.wav).

NO ATTRIBUTION IS REQUIRED FOR THE NOISE, AND THAT SILENCE IS DELIBERATE — do
not read it as an oversight and do not add a credit line "to be safe". That
subset's LICENSE reads, in full:

    All selected recordings were marked as in the Public Domain
    on Free Sound (url: https://www.freesound.org/).

Public domain: no attribution obligation, no share-alike, no licence-version
ambiguity, and nothing to maintain on the page forever. Crediting "OpenSLR
SLR28" anyway would be harmless and honest about provenance, but it is a
choice, not a condition. (Second-hand caveat, stated once: that is the JHU
packagers' assertion about the upstream Freesound state.)

TWO NOISE SOURCES WERE USED EARLIER AND DELIBERATELY REPLACED. Do not
reintroduce either without re-reading this.

  DEMAND (10.5281/zenodo.1227121) — its Zenodo record contradicts itself. The
  metadata licence field reads `cc-by-4.0`, while the record's own description
  states "This work, the audio data and the document describing it, is licensed
  under a Creative Commons Attribution-ShareAlike 3.0 Unported License." Both
  permit commercial use, but under the ShareAlike reading a work incorporating
  that noise could itself have to be offered share-alike — an unresolved
  obligation on a commercial page.

  MUSAN (openslr.org/17) — no ambiguity, and fine to ship: its README states
  every file is Creative Commons or US Public Domain and that content
  forbidding commercial use was excluded. It was replaced only because the file
  used was CC BY 3.0, which means a credit line someone has to keep correct
  forever. The public-domain set above needs none.

The stems are the pipeline's own output from that noisy input; the noise is
largely removed from them, but they remain derived from it.

THE WEIGHTS ARE NOT UNDER A PERMISSIVE LICENCE. OpenAudio S2-Pro is released
under the Fish Audio Research License:

    https://huggingface.co/fishaudio/s2-pro/blob/main/LICENSE.md

Research and non-commercial use is free. COMMERCIAL USE REQUIRES A SEPARATE
WRITTEN LICENCE FROM FISH AUDIO (Section III; business@fish.audio) — the
weights grant no commercial rights on their own. Section IV(a) further requires
that anyone distributing these files, or a product built on them, retains this
notice and prominently displays "Built with Fish Audio". This site does so in
the footer on every page.

The required attribution notice, verbatim:

    This model is licensed under the Fish Audio Research License,
    Copyright © 39 AI, INC. All Rights Reserved.

Before regenerating these files, reusing them elsewhere, or swapping in audio
from another model, check that the terms above still describe what ships. The
obligations travel with the audio, not with this directory.