RIGHTS.md

Rights, Licensing & Redistribution — SONAR packaging

Status: packaged and rehosted under a maintainer-asserted license. Read this before relying on the data for anything beyond non-commercial research.

What this repo is

A re-encoded, schema-normalized packaging of the SONAR synthetic-audio detection evaluation set (arXiv 2410.04324; upstream code/data: github.com/Jessegator/SONAR). 4,548 clips: 2,274 bonafide (LibriTTS clean-test) + 2,274 spoof from nine speech-synthesis systems.

The rights situation (be honest about it)

SONAR has no unified upstream redistribution license:

• The paper explicitly declines one: the constituent sources "may be subject to different distribution licenses and usage restrictions," and it directs users to "the usage policies of the respective API providers."
• The official upstream repo has no LICENSE file and no redistribution statement.
• The spoof set aggregates output from systems with their own terms — notably OpenAI TTS API output (600 clips), whose API Terms of Service restrict redistribution — alongside xTTS, Seed-TTS, VALL-E, VoiceBox, NaturalSpeech 3, PromptTTS 2, FlashSpeech, AudioGen. The bonafide side is LibriTTS (CC BY 4.0).

We therefore cannot point to an upstream grant that clearly permits rehosting these audio bytes.

What we asserted, and why

This packaging is published under CC BY-NC 4.0 (full text in LICENSE.txt) as the safest defensible choice for an aggregated research benchmark we can't relicense permissively:

• NonCommercial narrows exposure on the commercial-API-derived portion (OpenAI TTS).
• Attribution preserves credit to the upstream sources and the SONAR authors.

This is a packaging license applied by the Arena maintainers, not an upstream grant. The underlying sources retain their own terms. CC BY-NC 4.0 here is an assertion over an aggregate we do not hold full rights to — it reduces, but does not eliminate, redistribution risk.

For downstream users

• Use is intended for non-commercial research / benchmarking only.
• If you need stronger guarantees, regenerate the spoof audio yourself from the upstream recipes in github.com/Jessegator/SONAR rather than relying on this rehost, and consult each provider's usage policy (OpenAI, Coqui/xTTS, etc.).

For maintainers / takedown

If any upstream rights-holder (OpenAI or any synthesis-system owner) objects, treat this as a takedown candidate: depin the dataset from arena-manifest (remove the core_set entry + add a dataset_repin/note CHANGELOG event) and delete or gate the HF dataset repo. The labels/protocol are ours; the audio is the encumbered part. The cleaner long-term fix is a loader that regenerates spoof audio on the user's side instead of rehosting it.

Recorded 2026-06-07 alongside the SONAR add. See also the dataset card's "License & redistribution" section and docs/plans/2026-06-07-sonar-arena-dataset.md.