Add ITU-R BS.1770-4 Loudness

[PetrochukM]

🚀 Feature

Implement a standard algorithm for computing audio loudness, like ITU-R BS.1770-4.

Motivation

Audio loudness is an important feature for representing audio. For example, it's used in this popular library: https://github.com/magenta/ddsp

Pitch

Here is a numpy implementation of ITU-R BS.1770-4. It'd be helpful to have this translated to PyTorch and included in this library: https://github.com/csteinmetz1/pyloudnorm/blob/master/pyloudnorm/meter.py

Here is the official loudness implementation from the original paper: https://github.com/BrechtDeMan/loudness.py/blob/master/loudness.py

[mthrok]

Hi @PetrochukM

Sounds good idea. Couple of clarification questions;

1. What are the differences between the csteinmetz1 implementation and BrechtDeMan?
2. Are these implementations work for float dtype within the range of [-1.0, 1.0] or do they expect integer type?
3. What is the equivalent found in DDSP? Is the API there good to follow?

[PetrochukM]

Thanks for the questions:

1. "csteinmetz1" includes multiple loudness implementations, including "BrechtDeMan". "csteinmetz1" includes a parameter called filter_class to select which implementation to use.
2. Both "csteinmetz1" and "DeMan" expect floating dtype for the audio. See these tests for example: https://github.com/csteinmetz1/pyloudnorm/blob/6baa64d59b7794bc812e124438692e7fd2e65c0c/tests/test_loudness.py#L6
3. In DDSP, they implemented loudness here: https://github.com/magenta/ddsp/blob/5a06bced8e4b2f77aabe418d07f7279fe237daff/ddsp/spectral_ops.py#L184
   There are a couple of notable differences in their approach to calculating loudness as compared to "BrechtDeMan":
   • They use A-weighting, instead of K-weighting as recommended by ITU-R BS.1770-4.
   • They don't filter out quiet frames before calculating loudness as recommended by ITU-R BS.1770-4.
   • They use a spectrogram to calculate the loudness, instead of the audio signal. Mathematically, this doesn't matter. It's just an implementation preference.
4. I have not used their API in the past, so I don't have a good gauge of its effectiveness. Their library is only "used by" 5 people based on Github statistics.

Side-note: I do like the idea of being able to choose between A-weighting and K-weighting. I also like the idea of choosing to filter, or not filter quiet frames.