fast-melroformer

About

fast-melroformer runs the Mel-Band RoFormer voice separation model on Apple Silicon using the Neural Engine and Metal GPU kernels, reaching about 4x real time on an 8 GB M3 MacBook Air versus 1.08x for the stock PyTorch GPU run. It uses an unmodified public checkpoint and reports 56.5 dB SDR against the fp32 output on a 3 minute clip. The author notes one quiet passage matched at only 40 dB because the Neural Engine runs in fp16.