torch-log-wmse
Copyright 2024-2026 Christopher Landschoot
Copyright 2026 SoundFoxLabs

This product includes software originally developed by Nomono:

    log-wmse-audio-quality
    https://github.com/nomonosound/log-wmse-audio-quality
    Copyright 2023 Nomono
    Licensed under the Apache License, Version 2.0

This project is a PyTorch port of that work and is licensed under the same terms. See LICENSE.

Portions derived from the original work, and the modifications made to them:

  * torch_log_wmse/{metric,freq_weighting_filter,utils,constants}.py - ported from numpy to PyTorch;
    FFT convolution in place of scipy.signal.oaconvolve; batched [batch, channel, stem, time] tensor
    API; differentiable loss support; changes to aggregation, impulse-response handling, and silence
    handling as described in README.md and CHANGELOG.md.

  * torch_log_wmse/filter_ir.f32 - the bundled hearing-sensitivity impulse response, derived from the
    filter design in the original work and stored as raw little-endian float32 rather than a pickle.

  * dev/create_freq_weighting_filter_ir.py - vendored from the original work with modifications, so
    that the bundled impulse response can be regenerated and verified rather than trusted as an
    opaque binary.

Where this port's numerical results deliberately differ from the original, the differences are
documented in README.md and CHANGELOG.md.
