Speech Restoration Demo

Sebastian Braun, Microsoft Research, 2026.

This demonstrates our speech restoration model using mean flow-matching aimed for real-time processing.

▼ Audio examples

Click a spectrogram to play / pause. NFE = Number of Function Evaluations.

Model Type Compute Algorithmic latency
NCSN++ acausal Flow Matching /w data prediction upper bound 66.4 GMACs/sec > 600 ms
NCSN++ causal Flow Matching /w data prediction Baseline 142.8 GMACs/sec 20 ms
ConvGLU1D Mean Flow Matching /w data prediction older model 0.1 GMACs/sec 20 ms
RMFSR Mean Flow Matching /w data prediction Proposed 1.2 GMACs/sec 20 ms
Click spectrogram to play
Currently playing
Loading audio examples...