Manifold-Robust Speech Representations
Self-supervised approach to noise-robust speech using manifold-aware contrastive learning, preserving geometric structure without heavy augmentation.
View project →The core problem this project addresses is one that shows up constantly in real-world speech applications: models trained in clean conditions degrade badly when the audio is noisy. Most approaches respond to this with heavy data augmentation or complex pretraining pipelines.
This project takes a different angle. By building representations that are aware of the manifold structure of the data, the model learns to preserve the geometric relationships between speech samples even under noise. The approach is self-supervised, which means it does not require labeled data — a useful property for low-resource settings.
The work was never published due to some unresolved technical issues. The repository has the full implementation.