A real-time framework for identifying and removing speech from recorded audio while retaining non-speech acoustic information that remains useful for sound event detection.
Audio recorded in homes can contain valuable information for machine listening but also highly sensitive personal information in speech. The project treats speech removal as a release-stage privacy mechanism rather than assuming domestic recordings can simply be published as captured.
The framework brings together PANNs and E-PANNs convolutional audio models, the Audio Spectrogram Transformer, Silero VAD and WebRTC VAD in a common workflow for speech versus non-speech decisions.
A software GUI exposes the comparison in real time so that model behaviour can be inspected interactively rather than only through offline metrics. The work connects directly to the privacy requirements behind releases such as The Sounds of Home.
