README.md

---
license: other
license_name: trilogue-research-use-license
license_link: LICENSE
language:
  - en
  - ru
  - kk
pretty_name: "TRILOGUE: A Trilingual Spoken Dialogue Fact-Checking Benchmark with Evidence and Paired Audio"
size_categories:
  - "10K<n<100K"
task_categories:
  - automatic-speech-recognition
  - text-classification
tags:
  - audio
  - text
  - fact-checking
  - misinformation
  - spoken-dialogue
  - multilingual
  - evidence-retrieval
gated: true
extra_gated_heading: "Request access to TRILOGUE"
extra_gated_description: "TRILOGUE is available for permitted research use under the TRILOGUE Research Use License. Access is granted automatically after users provide the required information and accept the terms below."
extra_gated_button_content: "Agree and access TRILOGUE"
extra_gated_prompt: "By requesting access, you confirm that you have read and agree to comply with the TRILOGUE Research Use License."
extra_gated_fields:
  "Affiliation": text
  "Intended research use": text
  "Country": country
  "I agree to the TRILOGUE Research Use License": checkbox
  "I will not use the recordings for voice cloning, impersonation, speaker identification, re-identification, or biometric profiling": checkbox
  "I will not redistribute the dataset or provide access to unauthorized users": checkbox
---

# TRILOGUE

**[Dataset website: TRILOGUE, Lost in Speech, and MAD2](https://chaewanc-chun.static.hf.space/)**
— Compare the datasets' tasks, sizes, languages, and available audio, explore
their source relationships, and copy BibTeX citations. No login is required to
view the website.

TRILOGUE is a trilingual spoken-dialogue fact-checking benchmark with clean
text, turn-level ASR transcripts, word-level timestamp alignments, evidence
supervision, fixed article-disjoint experimental splits, and paired audio. The
complete benchmark contains 11,957 dialogues, 187,544 turns, and 390.3 hours of
audio in English, Russian, and Kazakh. The current package releases 8,036
Russian and Kazakh recordings (242.5 hours), including 4,998 human-read
recordings; 3,921 English TTS recordings are withheld as explained below.

Access to the dataset files is gated. Access is granted automatically after
users provide the required information and affirm their agreement to the
[TRILOGUE Research Use License](LICENSE) and the usage restrictions below. By
requesting, downloading, or using TRILOGUE, users agree to the License.

## English audio availability

English dialogue text, turn-level ASR transcripts, and word alignments are
included. The 3,921 English TTS recordings generated with Inworld are withheld
from this release because Inworld's current [Acceptable Use
Policy](https://inworld.ai/aup) prohibits using its output in a dataset that may
be used to train, fine-tune, develop, test, or improve machine-learning or AI
technology. The recordings will not be uploaded unless separate permission is
obtained or the applicable provider terms permit their release.

The English rows remain in `metadata/audio_manifest.csv` for cross-layer
provenance. They use `audio_release_status=withheld_provider_terms` and have a
blank `audio_path`. Russian and Kazakh audio rows use
`audio_release_status=released`.

## Dataset structure

- `data/`: one canonical clean-dialogue JSONL per language: 3,921 English,
  4,062 Russian, and 3,974 Kazakh records.
- `audio/`: 8,036 released lossless FLAC recordings. Russian and Kazakh have
  separate `human_read/` and `tts/` directories. English audio is withheld.
- `derived/asr/<language>/`: one turn-level ASR JSON per dialogue, stored
  directly under `en`, `ru`, or `kk`.
- `derived/word_alignments/<language>/`: one word/turn alignment JSON per
  dialogue, stored directly under `en`, `ru`, or `kk`.
- `splits/`: article-disjoint dialogue and article-family assignment manifests
  for the fixed seeds 13, 42, and 2026. `dev` is the validation split used in
  the experiments.
- `metadata/audio_manifest.csv`: the authoritative dialogue-to-audio, ASR, and
  alignment index, including audio-release status, speech origin, duration,
  generation version, TTS provenance, and voice-metadata availability.
- `metadata/dialogue_prompt_map.csv`: generation, annotation, and
  postprocessing identifiers for every dialogue.
- `metadata/dialogue_source_map.csv`: input-sample links to Lost in Speech
  for each dialogue, including availability and known source-issue flags.
- `metadata/article_family_source_map.csv`: links from internal article
  families to original-article record IDs in Lost in Speech.
- `metadata/source_linkage_provenance.json`: pinned upstream revision,
  input checksums, mapping checksums, and validation counts.
- `metadata/tts_voice_inventory.csv`: a deduplicated inventory of 46 synthetic
  voices, including voices associated with the withheld English audio. It
  contains no human-participant records.
- `docs/schema.md`: the detailed field, identifier, path, privacy, and split
  contract.

Dialogue text and available audio are stored once. The three seeded split
manifests point to the same canonical records; the release does not duplicate
data into physical train, validation, and test directories.

## Dialogue identifiers and generation versions

A stable dialogue UID has the form `<language>::<version>::<id>`, for example
`english::v1::1`.

| Version | Paper setting | Generation model | Annotation model |
| --- | --- | --- | --- |
| `v1` | Base | Gemini 2.5 Flash | Gemini 2.5 Pro |
| `v2` | Diversity-controlled | DeepSeek-V3.1 via Azure | Gemini 2.5 Pro |
| `v3` | Style-randomized | Gemini 2.5 Flash | Gemini 2.5 Pro |

Prompt identifiers use `v1_artist`, `v1_judge`, and `v1_postprocessing`, with
corresponding `v2_*` and `v3_*` identifiers.

The complete generation and annotation prompt templates, runtime placeholders,
model settings, directive pools, and cleanup/normalization procedures are
reported in the paper appendix section **Generation and Annotation Prompts**.
The official paper URL will be added when the proceedings version becomes
available. Prompt text is not duplicated in this repository.

## Fixed experimental splits

The `splits/` directory contains dialogue-level and article-family-level
assignment files for seeds 13, 42, and 2026. Each dialogue manifest contains all
11,957 dialogue UIDs and a `split` value of `train`, `dev`, or `test`.

All dialogues derived from the same internal article family remain in the same
split for a given seed. To reproduce a paper experiment, select the appropriate
seed manifest and join it to the canonical data using `dialogue_uid`. The
`dev` label corresponds to the validation split used in the experiments.

## Accessing the gated dataset

Keep the repository authenticated after access is approved. Users may download
from the Hugging Face Files tab or with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

local_path = snapshot_download(
    repo_id="chaewanC/TRILOGUE",
    repo_type="dataset",
)
```

See `docs/schema.md` for field definitions and release paths.

## Evidence and source-article linkage

The release includes claim-level `evidence_snippet` annotations needed for
evidence-grounded fact-checking. It does not include the source-article Excel
workbook, full source articles, or a duplicate catalogue of article titles,
publishers, and URLs.

The source resource is
[Lost in Speech](https://huggingface.co/datasets/maristombayeva/lost-in-speech).
`metadata/dialogue_source_map.csv` links 11,956 of the 11,957 dialogues to
upstream input samples. `metadata/article_family_source_map.csv` links 3,665
of the 3,672 internal article families to upstream original-article records.
The links are pinned to upstream revision
`a6e5feed8552233ba47c5b2bb34b5e4e294e8279`. Dialogue IDs, content, and fixed
TRILOGUE split assignments are unchanged.

An input sample can be a hallucinated rewrite; it must not be used as the
original factual reference. Join by `article_family_id` to obtain the
`source_sample_ids` of original records. Lost in Speech withholds original
article text, so these links identify records and their available metadata;
they do not make the original reference texts available for reproducing the
full evidence-retrieval experiment.

Dialogue `4485` has no upstream input record. Seven families (nine dialogues)
have no original record in the local source workbook and remain unresolved.
Upstream pending-review records `8333`, `8412`, and `8456` and prior
dialogue/source issues for `4485`, `4553`, and `6098` are explicitly flagged.
See [source linkage](docs/source_linkage.md) for matching methods, duplicate
original records, limitations, and a join example. The upstream dataset has
its own access conditions and experimental splits; this mapping does not
establish cross-dataset train/test separation. Third-party evidence and source
materials remain subject to their applicable rights and access conditions.

## Synthetic-audio provenance

Synthetic recordings are identified as TTS in the audio manifest. The released
Russian TTS audio was generated using a paid Azure Speech tier and the standard
prebuilt neural voices `ru-RU-DariyaNeural`, `ru-RU-DmitryNeural`, and
`ru-RU-SvetlanaNeural`. The generation code uses named Azure Speech voices and
does not configure a custom-voice endpoint or deployment. Microsoft lists these
names among its [supported prebuilt Russian
voices](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support),
and its [Product
Terms](https://www.microsoft.com/licensing/terms/en-US/productoffering/MicrosoftAzure/allprograms)
grant paid-tier customers use rights in prebuilt-neural-voice output, subject to
the applicable terms.

The English Inworld TTS audio is not released, as described above. Provider and
voice metadata are retained so the synthetic status and provenance of all
benchmark recordings remain explicit.

## Intended uses

TRILOGUE is intended for internal or published research and education on:

- Spoken and multilingual fact-checking
- Misinformation detection
- Evidence retrieval and localization
- Automatic speech recognition and ASR-robust reasoning
- Multilingual language and speech processing
- Evaluation of speech-native, audio-language, and long-context systems

Researchers employed by commercial organizations may use TRILOGUE for a
permitted research purpose. This does not authorize incorporation into a
deployed commercial product or service.

Users may train, evaluate, and publish research model checkpoints and other
learned artifacts when they comply with the License and do not intentionally
reproduce, redistribute, or provide practicable access to Dataset records or
human-participant recordings.

Controlled generation of false or misleading content is permitted when it is
reasonably necessary for a research purpose, clearly identified as generated or
false, and handled consistently with the License.

## Prohibited and out-of-scope uses

The License prohibits, among other things:

- Voice cloning, voice imitation, or impersonation
- Speaker identification, speaker verification, re-identification, or biometric
  profiling
- Attempts to discover or infer participant identities or demographics
- Redistribution of the Dataset or sharing access credentials
- Use in deployed voice-synthesis, biometric, surveillance, advertising, or
  identity-related products
- Generation, amplification, deceptive presentation, or repackaging of false
  claims for deceptive, harmful, or non-research purposes
- Removal or misrepresentation of factuality labels, evidence, provenance, or
  generated-data status
- Reuse outside the stated research purpose
- Use of third-party synthetic audio to create, train, or improve a TTS or
  synthetic-voice product or service where the applicable provider terms
  prohibit that use

The [LICENSE](LICENSE) controls if this summary and the full terms differ.

## Human recordings and privacy

The project did not collect or release human-participant names, voice IDs,
gender, or demographic attributes. Human-read rows in
`metadata/audio_manifest.csv` use
`voice_metadata_status=not_collected_irb`; their voice fields are intentionally
blank.

Human speakers read assigned dialogue content rather than providing personal
opinions, autobiographical statements, or private information. Adult speakers
participated under an institutionally approved protocol and gave informed
consent for research use and dataset release. Access is gated because voice
recordings may constitute identifiable biometric data.

Human validation was conducted with consenting adult annotators proficient in
the relevant language. Only aggregate validation findings are reported; the
release does not contain annotator identities or individual-level metadata.

## Limitations

TRILOGUE is a controlled benchmark. The dialogues were generated from source
articles and controlled hallucinated rewrites rather than collected from
naturally occurring misinformation conversations. Results should not be
interpreted as measuring performance on every real-world social-media,
private-messaging, or adversarial misinformation setting.

Evidence retrieval is source-conditioned within a known associated article
family rather than open-domain web source discovery. The current release does
not evaluate conflicting sources or temporally changing evidence.

Kazakh speech technology remains a bottleneck, particularly for synthesis,
recognition, and alignment quality. The paper baselines use ASR transcripts as
the official model input and do not fully evaluate prosody-aware,
speaker-aware, or speech-native reasoning over the released audio.

The benchmark includes generated false claims. Examples must not be presented
without their generated-data status, factuality labels, and evidence context.

## Known data notes

- Kazakh TTS dialogue `9574` has audio but no recoverable voice assignment. It
  is marked `missing_voice_assignment`.
- The source WAVs for Kazakh human-read dialogues `5262`, `5294`, `5297`,
  `7299`, `7322`, `7338`, and `7340` end before their header-declared durations
  and before the full scripted dialogues are complete. The released FLAC files
  preserve all physically available samples. These recordings are included for
  consistency with the paper-version dataset, but their missing final audio and
  any alignment timestamps beyond the physical endpoints should be treated as
  incomplete. We plan to repair these records in a later dataset revision if
  complete source recordings can be recovered.
- Repository paths and Hugging Face language metadata use the ISO codes
  `en`, `ru`, and `kk`.

## Deliberately excluded

- English Inworld TTS audio, pending permission or provider terms that permit
  dataset release
- Source-article workbook and full source-article texts
- Human-participant identities and individual demographic metadata
- Duplicate JSONL/Parquet dialogue copies
- Duplicate synthesis-input text folders
- Seed-specific copies of train, validation, and test data
- Separate public and human-audio repositories
- `corrections.jsonl`, a top-level `quality/` directory, and a separate
  `checksums/` directory

## Versions

- **v1.0** — Initial release accompanying the Findings of EMNLP 2026 paper.
- **Source-linkage update (2026-10-03)** — Added version-pinned Lost in Speech
  input-sample and original-article mappings, availability and source-issue
  flags, and validation provenance. Existing dialogues, audio, and splits are
  unchanged.

## Citation

Until the official ACL Anthology record is available, cite:

**Chaewan Chun**, Meruyert Aristombayeva, Jiyoung Choi, Mahjabin Nahar, Delvin Ce Zhang, and Dongwon Lee.  
“TRILOGUE: A Trilingual Spoken Dialogue Fact-Checking Benchmark with Evidence and Paired Audio.”  
*Findings of the Association for Computational Linguistics: EMNLP 2026*, 2026.

```bibtex
@inproceedings{chun-etal-2026-trilogue,
  title     = {{TRILOGUE}: A Trilingual Spoken Dialogue Fact-Checking
               Benchmark with Evidence and Paired Audio},
  author    = {Chun, Chaewan and Aristombayeva, Meruyert and
               Choi, Jiyoung and Nahar, Mahjabin and
               Zhang, Delvin Ce and Lee, Dongwon},
  booktitle = {Findings of the Association for Computational Linguistics:
               EMNLP 2026},
  year      = {2026}
}
```

## License and contact

TRILOGUE is distributed under the custom [TRILOGUE Research Use
License](LICENSE).

Contact: Chaewan Chun, `czc5884@psu.edu`