Datasetpublic

maikezu/asr-context-induced-leakage

When Helpful Context Leaks: Privacy Risks in Domain-Adapted ASR Overview SpeechLLMs are increasingly deployed in professional settings where domain customisation is standard practice: users supply context in prompts, fine-tune on proprietary recordings, or both. We identify and systematically investigate an overlooked privacy risk of such customisation: a model adapted to recognise domain-specific terminology can be nudged into transcribing a phonetically… See the full description on the dataset page: https://huggingface.co/datasets/maikezu/asr-context-induced-leakage.

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes61downloads

Dataset Card

When Helpful Context Leaks: Privacy Risks in Domain-Adapted ASR

Overview

SpeechLLMs are increasingly deployed in professional settings where domain customisation is standard practice: users supply context in prompts, fine-tune on proprietary recordings, or both. We identify and systematically investigate an overlooked privacy risk of such customisation: a model adapted to recognise domain-specific terminology can be nudged into transcribing a phonetically similar word from its context or training data, even when a different word is spoken. We construct a controlled dataset and measure leakage rates across two customisation mechanisms, prompt context injection and data fine-tuning. Both mechanisms cause measurable leakage, compounding when combined. We evaluate a prompt-level mitigation strategy and find that fine-tuning without prompt context injection offers the best accuracy–leakage trade-off. We release our code and dataset publicly.

Paper: arXiv:2605.28211 Code: GitHub

Dataset Structure

The dataset contains five splits:

Split	Description
`test`	Evaluation pairs: original audio with all context sentence variants
`train_prompt_adapt`	FLEURS train data for prompt-adaptation fine-tuning
`train_context_word`	TTS audio for context word fine-tuning (leakage model)
`train_acoustic_word`	TTS audio for acoustic word fine-tuning (contextual biasing)
`train_both_words`	TTS audio for both-words fine-tuning (mitigation)

Test split columns

Column	Description
`audio`	Original audio (acoustic word spoken)
`reference`	Ground truth transcript
`acoustic_word`	The word spoken in the audio
`context_word`	The phonetically similar distractor word
`phoneme_distance`	Phoneme edit distance between the two words (1 or 2)
`dataset`	Source dataset (fleurs / acl6060 / voxpopuli)
`similarity`	Lexical similarity between context sentence and reference
`ctx_1_context`	1-sentence context containing the context word
`ctx_1_acoustic`	1-sentence context containing the acoustic word
`ctx_2_mixed`	2-sentence context containing both words
`ctx_5_context`	5-sentence context, context word
`ctx_5_acoustic`	5-sentence context, acoustic word
`ctx_5_mixed`	5-sentence context, both words
`ctx_10_context`	10-sentence context, context word
`ctx_10_acoustic`	10-sentence context, acoustic word
`ctx_10_mixed`	10-sentence context, both words

Source Datasets

Word pairs are derived from:

Citation

bibtex

@misc{züfle2026helpfulcontextleaksprivacy,
      title={When Helpful Context Leaks: Privacy Risks in Domain-Adapted ASR}, 
      author={Maike Züfle and Jan Niehues},
      year={2026},
      eprint={2605.28211},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.28211}, 
}