6 projects
rope-conformer
A clean, plug-and-play Conformer encoder with rotary positional embeddings.
mps-sdpa
Fast drop-in scaled_dot_product_attention for Apple silicon (MPS) — wraps Apple's native MPSGraph SDPA with a zero-copy C++ bridge. 5-7x inference, 2-2.5x training.
moises-light
Moises-Light: Resource-efficient Band-split U-Net for Music Source Separation
torch-l1-snr
L1-SNR loss functions for audio source separation in PyTorch
torch-log-wmse
logWMSE is an audio quality metric & loss function with support for digital silence target. Useful for training and evaluating audio source separation systems.
torch-log-wmse-audio-quality
logWMSE is an audio quality metric & loss function with support for digital silence target. Useful for training and evaluating audio source separation systems.
