Research project

Full-night Tracheal Audio Event Modelling

Full-night acoustic representation learning for localising apnea–hypopnea events and estimating burden.

Problem

Short isolated clips can miss temporal context and event burden across a full sleep study.

Why it matters

Event-level localisation can make model outputs more interpretable and clinically aligned than only producing a single whole-night label.

Pipeline

The pipeline remains readable without animation and progressively activates while scrolling.

01

Full-night audio

02

Chunking & QC

03

SSL representation

04

Temporal localisation

05

Event burden

Signals

Tracheal audio

Models

Self-supervised speech/audio encodersTemporal event models