starkdmi/MossFormer_GAN_SE_16K_CoreML

MossFormerGAN SE 16K (Core ML)

Speech enhancement at 16 kHz, compiled for Core ML on Apple silicon. Converted from alibabasglab/MossFormerGAN_SE_16K.

An MLX conversion of the same base model is at starkdmi/MossFormer_GAN_SE_16K_MLX.

Files

package	size
MossFormerGAN_256frames_FP16.mlpackage	7.6 MiB
MossFormerGAN_256frames.mlpackage	13.5 MiB

Core ML fixes precision when a package is compiled, so these are two separate models rather than one with a runtime flag. Both take 256 STFT frames per segment — 25500 samples, 1.594 s at 16 kHz.

Precision

precision	size	speed	peak footprint	quality vs FP32
FP16	7.6 MiB	4.2x	302 MiB	61 dB
FP32	13.5 MiB	3.7x	1656 MiB	reference

Bold marks the best value in each column. Quality is FP16 against the FP32 conversion's own output.

Resident memory is close to identical between the two; the difference in peak footprint is compressed pages, as the FP32 path allocates roughly 1.4 GiB of transients. That also shows in consistency — FP32's run-to-run standard deviation was 68x FP16's.

Measured on an Apple M1 Pro (16 GB), 30 s of audio, three timed runs per configuration; speed is realtime factor. These are single-machine figures — expect different absolute numbers on other hardware, and treat the ordering as indicative rather than guaranteed.

Used by

AudioToolSwift, which compiles the package once on first load and caches the result.