Clinical speech sharing drops one-size-fits-all anonymisation: CASA tunes per task and clears privacy and diagnosis on all 14 tasks
Synopsis
Clinical speech carries both speaker-identifying voiceprints and sensitive spoken content, while different clinical tasks rely on different speech cues, so a fixed anonymisation pipeline struggles to keep both privacy and diagnostic value. CASA recasts anonymisation as an optimisation driven by a task description, repeatedly evaluating and adjusting strategies across voiceprint transformation and sensitive-content masking, and storing cross-task experience in memory. Across 14 clinical tasks in five languages, CASA met speaker-anonymisation and downstream classification requirements on all of them, versus 8 for the strongest fixed method and 4 for SecureSpeech.
Interpretation
Anonymisation of clinical speech moves from a fixed preprocessing step to a task-tailored optimisation, with one system choosing voiceprint and content strategies for different diseases, languages and elicitation paradigms. Earlier anonymisation split into two families: general-speech methods mainly suppress speaker identity and often degrade clinically relevant speech information, while disease-specific methods target a single disorder or narrowly defined task and do not carry across clinical settings. Evaluation covered 11 clinical speech datasets, 14 tasks or subsets and five languages spanning mental health, cognitive and voice disorders; success required a speaker-verification equal error rate of at least 20%, sensitive-content localisation F1 of at least 0.75, and a downstream balanced-accuracy drop of less than three percentage points versus original speech, and CASA met all three on all 14 tasks.
The system plans its own anonymisation workflow from task requirements and iterates on evaluation feedback until both privacy and utility thresholds are met. Available operations comprise five voiceprint methods, namely pitch shifting, linear predictive coding based formant transformation, spectral formant scaling, neural voice conversion and transcription-resynthesis, plus large-language-model localisation and masking of sensitive spoken content. The agent receives a natural-language description of the clinical objective, speech paradigm and user-set privacy and utility requirements, scores each execution in an evaluation module, and replans when requirements are unmet for up to seven iterations; sensitive-content localisation calls a locally deployed Qwen2.5-32B-Instruct so transcripts are not sent outside.
Experience accumulated across tasks lets the agent find acceptable strategies faster on new tasks and cuts repeated trial and error. Without memory, the agent could only adjust step by step using feedback from the current task. Across 140 requests, ten for each of the 14 datasets or task subsets, enabling memory raised first-evaluation success from 46.4% to 62.9% and success within two evaluations from 57.1% to 81.4%, while the mean number of strategy evaluations needed to reach an acceptable configuration fell from 2.27 to 1.56 steps, a 31.3% reduction.
Perspective
Evaluation covers 14 tasks or subsets across 11 clinical speech datasets, five languages and mental health, cognitive and voice disorders, including single-speaker and multi-speaker recordings; clinical expert assessment was conducted only on psychological crisis, suicide risk and cleft-palate speech tasks.
Sensitive-content localisation and transcription-resynthesis depend on automatic speech recognition, whose performance may decline for low-resource languages, dialectal speech, child speech and severely pathological speech, and recognition errors can propagate to later anonymisation steps; only five voiceprint methods and one LLM-based content method are currently integrated; human review may remain necessary for high-risk applications where residual privacy leakage is unacceptable.