Datasets:
MAD2 Source and Method Notices
Prepared September 14, 2026 for the dataset-only MAD2 distribution. These notices identify documented sources and method licenses, not a certification that every possible use is cleared. The custom MAD2 license covers only interests the authorized releasing party controls. Original third-party license texts prevail for their respective material.
Dataset Content and Voice References
| Source | Role and documented terms | Treatment in MAD2 |
|---|---|---|
| Alba MacKenna, Casual Voice Lines | Reference recording: CC BY 4.0 | Retain creator credit, original source URL, license and modification/use notice |
| EARS participants p015 and p016 | Reference recordings: CC BY-NC 4.0 | Preserve applicable attribution and noncommercial terms; do not replace the upstream grant |
| Expresso performer ex01 | Reference recording: CC BY-NC 4.0 | Preserve applicable attribution and noncommercial terms; do not replace the upstream grant |
| Kyutai tts-voices | Distributor of the selected reference excerpts | Credit as distributor, not owner of all underlying recordings; preserve pinned paths and hashes |
| LIAR, William Yang Wang and original sources | Seed claims, inherited labels and source IDs; original notice permits research use and reserves source copyright | Retain the unchanged LIAR notice and citation; no claim of unrestricted relicensing |
The original reference WAVs and LIAR TSVs are not bundled. The selected references were converted to mono 24 kHz PCM16 without enhancement, pitch, speed, or gain changes, then used to synthesize unrelated dialogue text. These synthetic performances do not represent actual statements or endorsements by the speakers. See voice attribution, exact reference records, CC BY 4.0, CC BY-NC 4.0, EARS license, and LIAR notice.
Source links: Alba, EARS, Expresso, pinned Kyutai distribution, LIAR paper.
The EARS authors report that speakers consented to open release of their recordings and research use (paper, section 2). This is affirmative publisher evidence; the underlying participant agreements have not been independently reviewed for MAD2's public cloned-speech context.
Source-recording licenses and any applicable personality rights are separate. These notices do not claim that the performers separately approved public cloned speech or that a new individual permission is always required. They do not assert that every generated waveform is necessarily a copyright adaptation. No separate identity, endorsement, or further-cloning permission is granted by MAD2 alone.
Synthesis and Audio Quality Processing
| Method | Actual use and source | Documented license scope |
|---|---|---|
| MoonCast | Synthesis of the 1,000 replacement dialogues; modified inference and runner are preserved privately | Assembled source has MIT notice, copyright 2025 Zeqian Ju; README has separate demonstration-prompt restrictions and a research-purpose statement |
| MoonCast pretrained resources | Text-to-semantic, audio tokenizer/detokenizer, tokenizer and bundled vocoder used by the preserved runtime | The exact archived model card declares MIT; this is distinct evidence from the GitHub source license |
| Meta W2v-BERT 2.0 | Default reference-audio encoder in the preserved MoonCast tokenizer | Publisher model card identifies MIT; exact cache revision or environment override is not established by this notice |
| NVIDIA BigVGAN implementation | Vocoder implementation loaded with MoonCast's bundled vocoder checkpoint | Upstream implementation is MIT, copyright 2024 NVIDIA CORPORATION; this does not substitute a different pretrained checkpoint for the actual MoonCast resources |
| faster-whisper / Systran faster-whisper-large-v3 | ASR for new-audio quality reports | Software and converted model are identified as MIT; this is not a declaration that underlying dialogue text is MIT |
| pyannote.audio / speaker-diarization-community-1 | Speaker diarization and reference matching for new-audio QA | pyannote.audio software is MIT; Community-1 model is CC BY 4.0, a distinct model license |
| WhisperX and alignment components | Used in historical processing; replacement-audio sentence alignment is not yet complete | Historical alignment model resolution is unpinned; no future alignment software/model license is represented as already verified |
The new-audio QA runner records Systran/faster-whisper-large-v3 and
pyannote/speaker-diarization-community-1; the latter is pinned to revision
3533c8cf8e369892e6b79ff1bf80f7b0286a54ee. Predictions and times are automatic
measurements, not human gold annotations.
Method sources: MoonCast code license, MoonCast README, W2v-BERT model card, BigVGAN license, faster-whisper license, ASR model card, pyannote.audio license, pinned diarization model card.
The exact assembled MoonCast source license is also retained in MOONCAST_CODE_LICENSE.txt. Keeping this notice does not mean source code or weights are included. Its restricted demonstration prompts and old generated audio are excluded from the dataset release.
The exact archived MoonCast model card is retained
unchanged. Its SHA256 is
d8d7a46d41a1a37fe4f0a5f637bf55c649310185329127d8a2204632e480be17.
It contains a model-license declaration, not a separate assurance concerning every
upstream training right. The runtime supports the default encoder identifier
facebook/w2v-bert-2.0 and an environment override, and uses MoonCast's own bundled
vocoder resource rather than a separately identified NVIDIA checkpoint.
Text Generation, Screening, and Auxiliary Outputs
Gemini chat generated dialogue text according to the project provenance; earlier records identify Gemini 2.5 Pro. The creator confirmed personal-account use. ChatGPT screening is described in project records, but the exact service version, generation dates and extent of incorporated rewrites are not fully established. DeepSeek-related scripts identify DeepSeek-V3.1 through an Azure endpoint for summary/relabel workflows; script presence alone does not prove that every current label was generated by that model. Auxiliary summaries and current labels are not represented as human-verified gold annotations.
The compact public package contains original dialogues, replacement audio, ASR, and current labels. Context summaries, full generation/QA reports, old labels, experimental splits, and local preparation tools are archived outside the release. Method descriptions above include historical construction and quality-review use; they do not imply that all resulting diagnostic artifacts are distributed here.
DeepSeek-V3.1 identifies its model license as MIT. Hosted-account terms are separate from a model's weight license. This does not relicense all generated text or third-party inputs as MIT.
Applicable generation-time service terms remain a creator-side provenance and compliance matter. They are not pasted into the MAD2 license as though every recipient signed the creator's account agreement. These notices neither certify unconditional downstream training rights nor characterize dataset-only publication as automatically prohibited. Any actual output-specific restriction remains applicable according to its scope.
Google's May 2024 terms and current US terms restrict use of service-generated content for ML development. The applicable generation-time agreement and its consequences for this training-benchmark release remain unresolved. This is not stated as a general prohibition on publishing generated text.
OpenAI's consumer terms address output ownership, review, and competing-model development. The publication policy also addresses disclosure and publication. These are distinct from the MIT license of locally run Whisper. DeepSeek's hosted Azure use likewise depends on the actual deployment agreement, not just its MIT model label; see Microsoft's deployment distinctions.
Scope Boundaries
No experimental code, runtime environment, pretrained weights, or experiment checkpoints are part of this release. This is therefore not a redistribution license bundle for every runtime dependency. A future code/model release would need its own exact-version dependency and notice review. Using a method is not by itself a reason to label all MAD2 output with the method's software license.
Preserve attribution and third-party conditions where they apply. Do not add blanket restrictions that narrow recipients' independent upstream CC permissions. Nothing in the MAD2 license supplies an absent right in third-party material or limits public-domain uses or applicable legal exceptions.