Lost in Speech is available for permitted research use under the Lost in Speech Research Use License. Access is granted automatically after you provide the required information and accept the terms. By requesting access, you confirm that you have read and agree to comply with the License.

Lost in Speech

A trilingual benchmark for reference-free classification of synthetically introduced factual and contextual alterations in English, Russian, and Kazakh. It contains 12,013 samples derived from news articles, with text, synthesized speech, and ASR transcript representations used in the study.

Altered samples are LLM-generated rewrites with a controlled alteration type—contradiction, fabrication, or context inconsistency—and severity level—mild, moderate, or severe. In the reference-free setting, the classifier receives the sample without its source article or external evidence.

v0.1 includes rewritten texts, labels, source URLs where available, ASR transcripts in all three languages, and Kazakh synthetic audio. Original article texts are withheld from the text files; their ASR transcripts are included, as is their Kazakh audio. Three Russian records have both text and transcript withheld pending review. English and Russian audio are not included in this release.

Access is gated and granted automatically after the required information is submitted and the Research Use License is accepted.

Release status

Component	v0.1 status
Synthetic subset: rewritten texts, labels, source URLs	Released, subject to the exceptions below
ASR transcripts (en, ru, kk)	Released, except three Russian records pending review
Train/dev/test split manifest	Released
Encoder predictions on the test split	Released
Kazakh TTS voice assignments	Released
Kazakh synthetic audio	Released for 3,978 samples, in FLAC format
Russian synthetic audio	Not released; planned for v1.0, subject to confirmation of distribution terms
English synthetic audio	Withheld pending clarification of distribution permissions
English TTS voice assignments	Planned for v1.0
Source article metadata: outlet and date	Planned for v1.0
Real-world evaluation metadata (ru, kk)	Released: identifiers, binary labels, origin metadata, and source/fact-check URLs where available; no texts, audio, or ASR transcripts
Audio manifest	Planned for v1.0

Generation and evaluation prompts are reported in the paper appendix and are not duplicated here.

Data availability and record status

Original news article texts are not included in the synthetic text files. Those files contain source references, labels, and LLM-generated rewrites where available.

Each synthetic record has a text_status field:

Value	Meaning
`available`	The rewritten text is included.
`withheld_original`	The record is an unaltered source article, labelled `No`; its `text` is `null`.
`pending_review`	Both text and ASR transcript are withheld pending verification against generation records.

The pending_review records are Russian IDs 8333, 8412, and 8456. Their stored labels remain in the release and in the aggregate counts below, but are not verified. Exclude these records from training and evaluation until their status is resolved.

Withholding the original text field does not remove the original content from all released representations. ASR transcripts of unaltered articles are included for all three languages, and their synthesized audio is included for Kazakh. These representations can preserve much of the original wording, with synthesis and recognition errors. Their availability and use are subject to the repository license and applicable third-party rights.

Dataset composition

No denotes an unaltered source article used as a truthful negative in the benchmark; Yes denotes a synthetically altered sample. These labels describe the benchmark construction and are not independent verification of every claim in a source article. Alteration type and severity counts apply to Yes samples only.

Language	Total	No	Yes	Released texts
English (`en`)	3,967	1,340	2,627	2,627
Russian (`ru`)	4,068	1,259	2,809	2,806
Kazakh (`kk`)	3,978	1,113	2,865	2,865
Total	12,013	3,712	8,301	8,298

Language	Contradiction	Fabrication	Context inconsistency	Mild	Moderate	Severe
`en`	903	938	786	875	877	875
`ru`	962	960	887	934	937	938
`kk`	947	956	962	957	953	955

Alteration types are not exactly balanced within each language. 127 records have no source URL; a missing URL does not establish that a source page has been removed.

Files and fields

Path	Contents
`data/synthetic/{en,ru,kk}.jsonl`	Sample records and labels
`data/synthetic/PROVENANCE.md`	Export provenance and input checksums
`derived/asr/{en,ru,kk}.jsonl`	ASR transcripts
`derived/asr/wer_cer.csv`	Corpus-level WER and CER by language
`derived/predictions/synthetic_test_predictions.csv`	ReMBERT and mDeBERTa test predictions for text (O) and ASR transcript (T) conditions
`metadata/splits/binary_split.csv`	Train/dev/test assignments
`metadata/tts_voices_kk.csv`	Kazakh TTS voice assignments by sample
`audio/kk/`	Kazakh synthetic audio in FLAC format
`metadata/real_world_ru.csv`	Russian real-world evaluation metadata: 580 records, comprising 290 misinformation items and 290 truthful negatives
`metadata/real_world_kk.csv`	Kazakh real-world evaluation metadata: 580 records, comprising 290 misinformation items and 290 truthful negatives

Synthetic JSONL fields are id, title, text, hallucination, hallucination_type, hallucination_level, language, url, and text_status. ASR JSONL fields are id, language, asr_system, and text; in these files, text is the ASR transcript.

Use sample identifiers and language to link records across files. Language codes are en, ru, and kk.

Splits and reproducibility

Use metadata/splits/binary_split.csv to recover the train/dev/test membership used in the study. Assignments were produced from the source workbook before original article texts were withheld. Reconstructing splits after filtering the public export may change sample membership.

Compute class counts and majority-class baselines from the actual evaluation split after any exclusions, rather than from the corpus totals above. Report the evaluated sample count, input condition, and excluded records alongside metrics.

The public release does not contain every text input needed to rerun the O condition, because original article texts for truthful negatives are withheld. Published predictions allow inspection of the stored outputs but do not replace those missing inputs. The T condition uses the separately released ASR transcripts, subject to the record availability described above.

Audio availability

Kazakh. Audio generated with KazakhTTS (Mussakhojayeva et al., 2021) is included in audio/kk/ for all 3,978 samples, in FLAC format. Kazakh voice assignments are provided in metadata/tts_voices_kk.csv. See LICENSE for the repository's terms and applicable third-party notices.

Russian. Audio generated with Silero TTS is not included in v0.1. A release is planned for v1.0, subject to confirmation of the applicable distribution terms. Russian ASR transcripts are included except for the three records pending review.

English. Audio generated with Coqui XTTS-v2 is withheld while distribution permissions for this grant-supported research are clarified with the rightsholder. The model and its outputs are governed by the Coqui Public Model License. English ASR transcripts are included. English voice assignments are planned for v1.0; any local regeneration must independently comply with the provider's terms and requires the corresponding input texts.

Pipeline and model provenance

Stage	Model or system
Alteration generation: English	`gpt-3.5-turbo`, `gpt-4`, `gemini-2.0-flash-lite`
Alteration generation: Russian and Kazakh	`gemini-2.0-flash-lite`
Faithful rewrites for provenance analysis	`gemini-2.0-flash-lite`
Independent quality judges	`claude-haiku-4-5`, `deepseek-chat`
Text-to-speech	Coqui XTTS-v2 (en), Silero TTS (ru), KazakhTTS (kk)
Automatic speech recognition	Whisper-large-v3 (en, ru), fine-tuned wav2vec2 (kk)
Detection encoders	ReMBERT, mDeBERTa-v3-base
Translation for the real-world evaluation split	Google Translate

Real-world evaluation metadata

Real-world evaluation metadata is provided in metadata/real_world_ru.csv and metadata/real_world_kk.csv. Each file contains 580 records: 290 misinformation items (label = 1) and 290 truthful negatives (label = 0).

The misinformation component comprises 290 distinct stories collected from factcheck.kz: 225 originally in Russian and 65 in Kazakh. Cross-translation provides a version of each story in both languages; these are not 580 independent misinformation stories. Each language-specific evaluation set also contains 290 truthful negatives drawn from the synthetic benchmark's unaltered source articles.

The released fields are item_id, story_id, language, origin, label, factcheck_url, source_id, and source_url. Five misinformation stories have missing factcheck_url values in both language files.

Only identifiers, labels, origin metadata, and URLs where available are released for this component. News and misinformation texts, synthesized audio, and ASR transcripts are not redistributed.

Intended uses and restrictions

Intended uses, within the scope of the Research Use License, include:

Reference-free classification of factual and contextual alterations in text and speech.
Multilingual speech processing and research on underrepresented languages, particularly Kazakh.
Studying classification robustness to TTS and ASR errors.
Evaluating audio-language models and cascaded speech pipelines, subject to available inputs.
Studying provenance confounds in synthetic detection benchmarks.

Researchers at commercial organizations may use the dataset for research permitted by the license. This does not authorize incorporation into a deployed product or service.

The license prohibits deceptive, harmful, or other non-permitted use of fabricated claims; presenting samples without their factuality labels and generation status; removing or misrepresenting those labels; redistributing the dataset or sharing access credentials; and uses prohibited by applicable audio-provider terms. Preserve the distinction between unaltered source articles, synthetic rewrites, and synthesized or recognized speech representations.

This section summarizes the repository terms. The full LICENSE governs, together with applicable third-party rights and component-specific terms.

Limitations

Synthetic speech and recognition errors. The benchmark uses read speech synthesized with language-specific TTS systems, not spontaneous or participant-recorded speech. Transcripts come from a TTS-to-ASR cascade, so differences can reflect both synthesis and recognition errors.

Kazakh audio quality. Reported corpus-level WER is 34.0% for Kazakh, 7.3% for English, and 9.1% for Russian. Targeted inspection of 120 Kazakh audio samples found 94.2% mostly intelligible, while 38.3% showed major semantic discrepancies relative to the source text. Thus, intelligibility does not ensure preservation of label-relevant meaning. These inspection percentages describe the inspected sample, rather than establishing prevalence across the full corpus. For affected items, the text-level label may not be recoverable from audio alone. Different recognition systems and language characteristics also limit direct comparison of WER across languages; consult CER alongside WER.

Provenance–label confound. Truthful negatives are human-written articles, whereas altered samples are LLM-generated rewrites. A classifier may partly detect generation style rather than factual or contextual alterations. The paper examines this issue through a provenance-by-veracity analysis on Russian.

Generator–language confound. English samples were generated primarily with GPT models, while Russian and Kazakh samples were generated with Gemini. Cross-lingual differences cannot be attributed exclusively to language or resource availability.

Release and evaluation coverage. Text and audio availability differ by language and label. The three pending-review records have unverified labels. Comparisons must account for the evaluated samples, excluded records, and available input representations.

Scope. The synthetic task concerns controlled alterations to news articles. It does not establish general-purpose fact-checking ability. Findings cover one domain and three languages; conclusions about Kazakh should not automatically be extended to other underrepresented languages.

Privacy

Source material comes from publicly available news; the real-world evaluation split uses published fact-check reports. Such sources can contain names and information about people, and public availability should not be interpreted as a guarantee that the content is free of sensitive information.

No human-participant speech recordings were collected for this benchmark; all benchmark audio is synthetic. Validation findings are reported in aggregate, without annotator identities or individual annotation records in the release.

Versions

v0.1 — Initial release: synthetic records, available rewritten texts and source URLs, labels, ASR transcripts in three languages with the stated exceptions, split manifest, encoder test predictions, Kazakh TTS voice assignments, and Kazakh synthetic audio.
2026-09-24 update — Added Russian and Kazakh real-world evaluation metadata in metadata/real_world_ru.csv and metadata/real_world_kk.csv. This update does not include real-world texts, audio, or ASR transcripts.

Citation

Meruyert Aristombayeva, Jason S. Lucas, Chaewan Chun, and Dongwon Lee. 2026. Lost in Speech: Trilingual Spoken Hallucination Detection Across Audio and Transcripts. In Proceedings of the Second Workshop on Speech and Audio Language Models (SALMA).

@inproceedings{aristombayeva-etal-2026-lost,
  title     = {Lost in Speech: Trilingual Spoken Hallucination Detection
               Across Audio and Transcripts},
  author    = {Aristombayeva, Meruyert and Lucas, Jason S. and
               Chun, Chaewan and Lee, Dongwon},
  booktitle = {Proceedings of the Second Workshop on Speech and Audio
               Language Models (SALMA)},
  year      = {2026}
}

License and contact

Distributed under the Lost in Speech Research Use License, with applicable third-party rights and terms as described above.

Contact: Meruyert Aristombayeva — m.aristombayeva@satbayev.university (primary), m.aristombayeva@gmail.com.

Downloads last month: 19

Total file size:

15.9 GB

You need to agree to share your Hugging Face information to access this dataset