nufrnd/hyperscreen-enhance-voice

HyperScreen Enhance Voice (Core ML)

Core ML conversions of two open speech models, as bundled in HyperScreen's Enhance Voice audio effect (macOS 15+, run on the GPU).

File	Model	Source	Licence
EnhanceVoiceDenoiser.mlpackage	MossFormer2 SE 48K (speech enhancement)	ClearerVoice-Studio, weights alibabasglab/MossFormer2_SE_48K	Apache 2.0 (LICENSE-ClearerVoice-Studio.txt)
EnhanceVoiceRestorer.mlpackage	TF-Restormer baseline (speech restoration / bandwidth extension)	TF_Restormer, weights shinuh/tf-restormer-baseline	MIT (LICENSE-TF-Restormer.txt)

Inputs and outputs

• Denoiser — features [1, 512, 180]: 512 frames (40 ms, 8 ms hop, 48 kHz, samples scaled to 16-bit range) of Kaldi fbank (60 mel bands) plus first and second deltas. Returns mask [1, 512, 961], a gain per bin of the matching 1920-point symmetric-Hamming STFT (no centring).
• Restorer — two functions sharing weights: from16k takes spectrum [1, 321, 200, 2], from8k takes [1, 161, 200, 2] (200 frames of 40 ms / 20 ms hop, real and imaginary, divided by their mean magnitude). Both return restored [1, 961, 200, 2], the same frames at 48 kHz.

Changes from the originals

Weights and maths are unchanged apart from: rewrites the converter needs (fixed-length rotary tables, einsum and padding rewritten as matmuls and concatenations), power-of-two rescaling inside the denoiser's FSMN blocks so it runs in half precision, 8-bit palettized weights (denoiser) and 8-bit linear weights (restorer). Outputs match the PyTorch originals closely (denoiser waveforms within 52–59 dB, same DNSMOS; restorer high-band log-spectral distance 12.0 dB vs 12.1 dB for PyTorch).

Built by convert.py in the HyperScreen repository (products/apps/hyperscreen/macos/scripts/voice-models/), which pins the upstream commits and weight checksums.