MAD2

MAD2 contains 1,000 synthetic English conversations, their original written dialogues, ASR transcripts of the replacement recordings, and current sentence check-worthiness labels. Audio was generated with MoonCast using alba, ears-p015, ears-p016, and expresso-ex01 reference voices.

The package contains 11.04 hours of mono 24 kHz PCM16 audio and 8,192 labeled sentences: 1,230 check-worthy and 6,962 negative. Each dialogue has two distinct voices; each reference voice occurs in 500 dialogues.

This is the compact local release package, not a record of public-release approval. Sentence forced alignment is not included: sentence start/end times remain null. The dataset release version remains to be finalized. The paper citation is provided below.

Audio Release and Re-Synthesis

The paper's experiments and WER evaluation use the original recordings generated with MoonCast's two default demonstration voices. MoonCast's speech-generation notice prohibits redistribution of its bundled demonstration audio prompts and audio generated from those prompts. We therefore regenerated the recordings for public release using the four alternative reference voices listed above. The original default-voice recordings are not distributed in this package.

The released audio is a replacement corpus: the paper's reported model performance and WER do not evaluate these replacement recordings.

Researchers are welcome to use the scripts in dialogues/original.jsonl to generate their own audio with other TTS systems and voices for noncommercial research, subject to the MAD2 license, applicable third-party terms, and the permissions required by their chosen TTS models and voices. Please identify the synthesis system and voices used and distinguish any new recordings from the supplied MAD2 audio. New recordings require their own ASR transcripts and timing alignments; timestamps from the supplied audio should not be reused.

Files

MAD2_release/
  README.md
  LICENSE
  THIRD_PARTY_NOTICES.md
  VOICE_ATTRIBUTION.md
  licenses/
  dialogues/
    original.jsonl
  audio/
    <dialogue_id>.wav
    metadata.jsonl
  asr/
    transcripts.jsonl
  annotations/
    labels.jsonl

The four JSONL files each contain 1,000 rows, joined by the string dialogue_id. There is one WAV per dialogue. JSONL is UTF-8 with one JSON object per line. No train/dev/test partition is prescribed; corpus names the complete collection, not a training split. Researchers must specify their own evaluation partition and keep all sentences from a dialogue together when splitting.

Original Dialogues

dialogues/original.jsonl stores the written dialogue text once, preserving the sentence strings, speaker roles and turn boundaries from the hash-verified server scripts. It does not include the old script labels. Fields:

• dialogue_id, source_dataset, source_claim_id: corpus and LIAR source identity.
• scenario_type, spread_style: corrected server categories.
• source_script_sha256: hash of the original source JSON, not this packaged row.
• turns: original_turn_idx, role, and sentences. Each sentence has sentence_id, original_sent_idx, and text.
• audio_script: compact text edits and turn spans that map the original dialogue to the exact text used to synthesize the released recording, without storing another full dialogue copy.

Indices are zero-based. A sentence ID is <dialogue_id>:<original_turn_idx>:<original_sent_idx>. Roles 0 and 1 refer to the voice assignments in audio metadata, not automatically inferred diarizer IDs.

Audio-Script Differences

Nine dialogues have text edits: seven added acknowledgments and two apostrophe corrections. Twelve inputs merge adjacent same-role turns. The original text and current sentence IDs are preserved; generated and original turn indices differ.

To interpret audio_script, first join the original sentence strings in order with one ASCII space. text_edits contains ordered, nonoverlapping char_start, char_end, replacement patches against that string. Apply them from last to first. The resulting string is the exact joined synthesis text. Then each turn_spans entry selects a generated turn using generated_turn_idx, role, char_start, and char_end in the resulting string.

All character spans are half-open Python Unicode-character offsets, not UTF-8 byte offsets. These are text mappings, not audio timestamps.

Audio

audio/metadata.jsonl contains file_name, dialogue_id, duration_seconds, sample_rate, audio_sha256, synthetic, engine, role_0_voice, and role_1_voice. file_name is relative to the audio directory. No full script, ASR transcript, or sentence-label table is duplicated in this metadata.

All 1,000 WAVs are the unchanged approved replacement takes. 979 passed automatic QA; the remaining 21 were accepted by the dataset owner, including seven exact takes reviewed by listening and 14 accepted as minor differences. This acceptance does not mean every ASR word or diarization boundary is correct.

ASR

asr/transcripts.jsonl contains dialogue_id, the recognized transcript, and words. Word entries preserve word, start, end, and probability from the new-audio faster-whisper large-v3 reports. Times are seconds relative to the WAV. The transcript and word sequence are two views of ASR output, not corrected script text. Word times/confidences are automatic, not human gold annotations; words are not assigned speaker identities in this table.

Labels

annotations/labels.jsonl stores one row per dialogue with:

• source_claim_veracity: inherited binary LIAR seed label (true or false), not a truth judgment about every utterance.
• source_claim_original_label: original LIAR true, false, or pants-fire. false and pants-fire map to binary false; true remains true.
• annotation_version: server_active_aligned_20260913, identifying the current server label layer. The name does not assert alignment to the new recordings.
• sentences: sentence_id, check_worthy, generated_text_spans, audio_start, audio_end, and alignment_status.
• audio_only_turns: the seven synthesis-only acknowledgments, identified by generated_turn_idx, with check_worthy: null and annotation_status: synthesis_only_addition_unlabeled.

check_worthy is a selection label, not factual correctness. The current 1,230 positive labels are retained, not the old 3,368 positives. Current does not imply human-verified. Audio-only acknowledgments are not counted among the 8,192 labeled sentences and must not silently be scored as negative.

generated_text_spans selects text within the reconstructed generated turn: generated_turn_idx, char_start, and char_end. Matching ignores case and punctuation, enforces the speaker role, and can omit boundary punctuation. It is not forced alignment. audio_start and audio_end are null, with status pending_new_audio_sentence_alignment; no old-audio sentence times are reused.

Loading

The default AudioFolder view loads audio and minimal metadata only. Load the other tables separately and join on dialogue_id; sentence text and labels join on sentence_id.

from datasets import load_dataset

audio = load_dataset("./MAD2_release", split="corpus")
dialogues = load_dataset("json", data_files={
    "corpus": "./MAD2_release/dialogues/original.jsonl"
}, split="corpus")
asr = load_dataset("json", data_files={
    "corpus": "./MAD2_release/asr/transcripts.jsonl"
}, split="corpus")
labels = load_dataset("json", data_files={
    "corpus": "./MAD2_release/annotations/labels.jsonl"
}, split="corpus")

No custom loader or experimental code is required in the repository. Local validation was performed with datasets 3.6.0; decoding dependencies depend on the installed library version.

Scope and Limitations

MAD2 supports research on conversational check-worthiness and misinformation. These are synthetic dialogues, not authentic conversations or evidence that a reference performer made the statements. Content can be false, political, misleading or offensive. Inherited seed labels and automatic annotations are not independent fact-checking evidence. The four-voice synthetic setting and source bias limit generalization. Earlier results on the original audio do not measure this replacement corpus.

Experimental splits, old labels, context summaries, diagnostic QA/diarization reports, generation logs, validation scripts, and release-preparation records are kept locally, outside this dataset package. Original restricted MoonCast audio, reference WAVs, LIAR TSVs, credentials, model weights and runtime code are not distributed here.

License and Attribution

MAD2-controlled contributions use the MAD2 Noncommercial Research License 1.0, permitting the specified noncommercial research uses only within the licensor's controlled rights. Upstream rights remain separate. Alba's source recording is CC BY 4.0; EARS and Expresso references are CC BY-NC 4.0; LIAR retains its source-copyright and research-use notice. Neither the custom license nor this repackaging certifies public cloned-voice scope, source/service compliance, or authority over any institutional/collaborator rights. Future license revisions do not retroactively replace validly granted permissions.

See source and method notices, voice attribution, and the unchanged texts in licenses/. LIAR: William Yang Wang (2017), "Liar, Liar Pants on Fire": A New Benchmark Dataset for Fake News Detection. Synthesis: MoonCast.

Contact/licensor for controlled or authorized rights: Chaewan Chun, czc5884@psu.edu.

Citation

@inproceedings{chun-etal-2026-context-aware,
  author    = {Chun, Chaewan and Zhang, Delvin Ce and Lee, Dongwon},
  title     = {Context-Aware Multimodal Claim Verification in Spoken Dialogues},
  booktitle = {The 2nd Speech and Audio Language Models Workshop
               ({SALMA}), {EMNLP}},
  year      = {2026}
}