MedDistractAudio

MedDistractAudio tests whether background speech from another consultation leaks into a transcript and then a generated clinical note. It contains 57 clean recordings and 456 mixed recordings, forming 456 clean–mixed pairs. Each foreground consultation has two donor segments at four levels: −20, −15, −10 and −5 dB. The prespecified primary level is −10 dB.

from datasets import load_dataset, Audio
pairs = load_dataset("NYU-OLAB/MedDistractAudio", "pairs", split="test")
recordings = load_dataset("NYU-OLAB/MedDistractAudio", "recordings", split="test")
# Access the lossless FLAC bytes without requiring an audio-decoding backend.
recordings = recordings.cast_column("audio", Audio(decode=False))

The test split is the full research evaluation corpus. The same foreground encounter and clean recording are reused across donors and levels. Resample or partition by foreground encounter, not independently by mixture.

Contents and keys

Configuration	Rows	Contents
pairs	456	Frozen clean/mixed Whisper transcripts, human reference transcript/note, donor, level, onset and exact segment parameters
recordings	513	Embedded lossless 16-kHz mono audio, stable recording ID and content hashes
transcripts	513	Saved Whisper large-v3 transcripts and chunk timestamps
notes	3,648	Saved notes, keyed by pair, model and clean/distracted condition
judgments	3,648	Paired automated contamination and quality judgments, including parse-failure flags
leakage	456	Frozen transcription-leakage and word-error measurements

Join recordings on item_id and distractor_type; use distractor_type="clean" for each pair's clean recording. Pair keys are source_dataset, item_id, distractor_type; add model and condition for note results. Legacy wav strings in the frozen tables are provenance paths, not downloadable URLs. The recordings configuration supplies the actual audio. transcripts.seconds is ASR processing time, not recording duration; use recordings.duration_seconds.

Audio was reconstructed for distribution from the original public PriMock57 WAV files and the exact saved donor/onset/segment/level parameters. The source WAVs were checked against their published Git LFS SHA256 hashes. FLAC preserves the reconstructed 16-bit PCM samples without lossy compression. All 513 reconstructed PCM waveforms were then hash-compared with the original experiment WAVs and match exactly. The container format is FLAC rather than WAV; no inference or judgments were rerun. The original transcripts, generated notes, scores and analysis inputs are unchanged. Source revision, waveform hashes and construction details are recorded in the release manifest.

Two of 1,824 note pairs have unparseable contamination judgments and are excluded by the analysis code, not assigned negative labels. The saved model scores are not clinician adjudications. The original run manifests verify that all four audio note models used raw prompts (no chat template), greedy decoding, a 512-token output cap and batch size two. These settings differ from the newer text-only Notes protocol; see protocol/note_generation.json. The frozen outputs remain the source for reproducing the reported results.

Sources and license

Derived from PriMock57, Papadopoulos Korfiatis et al. (2022), licensed under CC BY 4.0. PriMock57 consists of mock consultations involving clinicians and participants acting as patients. Original speaker channels are averaged, selected donor speech is overlaid at a recorded onset/level, and mixed audio is peak-normalized to 0.95. Source attribution is retained; this derivative release uses CC BY 4.0. Cite PriMock57 as well as MedDistractAudio.

Intended use and limitations

For research on background-speech contamination, ASR robustness and note attribution. These are controlled digital overlays of mock consultations, not recordings of natural overlap in deployed clinics. Saved transcripts use one ASR model without diarization. Results do not establish clinical deployment safety.

Code: https://github.com/nyuolab/llm_distract (the public toolkit release is being prepared).

Citation

Paper. The citation below uses the publicly posted v1 title; the authors have submitted a revised title. The scientific data are frozen to the submitted study.

@misc{vishwanath2026incidental,
  title={Incidental information contaminates patient notes and disrupts clinical reasoning in large language models},
  author={Vishwanath, Krithik and Ye, Brandon and Alyakin, Anton and Markert, John E. and Hsieh, Aaron and Mańkowski, Michał and Oermann, Eric K.},
  year={2026}, eprint={2610.08585}, archivePrefix={arXiv}, primaryClass={cs.CL},
  url={https://arxiv.org/abs/2610.08585}
}