multislope

Multi-slope sound energy decay estimation for room impulse responses.

Rooms rarely decay with a single reverberation time. Coupled volumes, non-diffuse fields, and strongly absorbing surfaces produce energy decay curves that need two or three exponentials to be described well. multislope estimates those decay times, their amplitudes, and the noise floor, per octave band, from a measured RIR.

pip install multislope

The core install is NumPy, SciPy, and ONNX Runtime — no PyTorch required.

Quickstart

import multislope
import soundfile as sf

rir, fs = sf.read("my_rir.wav")

net = multislope.DecayFitNet(sample_rate=fs)
fit = net.estimate(rir)

print(fit)
# DecayFit(method='DecayFitNet', sample_rate=48000)
#      125 Hz  T = [0.690] s  A = [0.963]  N = 1.9e-08
#      250 Hz  T = [0.530, 2.303] s  A = [0.787, 0.0634]  N = 1.15e-09
#   ...

fit.t          # (n_bands, n_slopes) decay times in seconds; 0 marks an inactive slope
fit.a          # (n_bands, n_slopes) amplitudes, linear scale
fit.n          # (n_bands, 1) noise floor, linear scale
fit.n_slopes   # (n_bands,) number of active slopes per band

Reconstruct the fitted EDC to compare it against the measured one:

import numpy as np

pre = multislope.PreprocessRIR(sample_rate=fs)
measured_edc, _ = pre.schroeder(rir)              # (n_channels, n_bands, n_samples)
time_axis = np.arange(measured_edc.shape[-1]) / fs

fitted_edc = fit.edc(time_axis)                   # (n_bands, n_samples)

Methods

	DecayFitNet	BayesianDecayAnalysis
Approach	trained neural network (bundled, ~10 MB of ONNX weights)	slice sampling over a discrete parameter grid
Speed	milliseconds per RIR	seconds to minutes per RIR
Slopes	1–3, either fixed or estimated by the network	1–3, either fixed or selected by BIC
Needs training data	already trained	no
Determinism	deterministic	stochastic; pass seed= to reproduce

Both take the same arguments and return the same DecayFit:

net = multislope.DecayFitNet(n_slopes=0, sample_rate=fs)          # 0 = estimate the count
bda = multislope.BayesianDecayAnalysis(n_slopes=0, sample_rate=fs, n_iterations=100, seed=0)

fit_net = net.estimate(rir)
fit_bayes = bda.estimate(rir)

Options

Fixing the number of slopes. n_slopes=2 fits exactly two slopes and skips model-order selection. n_slopes=0 lets the estimator decide, and inactive slopes come back with T = A = 0.

Frequency bands. The default bands are 125 Hz to 4 kHz. Pass your own centre frequencies; a 0 adds a lowpass band below the lowest octave, and sample_rate / 2 adds a highpass band above the highest one.

net = multislope.DecayFitNet(sample_rate=fs, filter_frequencies=[0, 125, 250, 500, 1000, 2000, 4000, fs / 2])

Direct sound. By default the whole RIR is analysed. Pass analyse_full_rir=False to detect the direct-sound onset and discard everything before it.

Analysing an EDC directly. If you already have an energy decay curve, pass input_is_edc=True to skip filtering and backwards integration.

Example data

The example RIRs from the DecayFitNet repository are downloaded on first use and cached locally, so they are not part of the wheel:

from multislope import data

rir, fs = data.example_rir("doubleslope")
print(data.available_examples())

Optional extras

pip install multislope[plot]    # matplotlib helpers in multislope.plotting
pip install multislope[train]   # PyTorch training code in multislope.training

multislope.training re-exports the original DecayFitNet training pipeline (dataset, model definition, and EDC loss) for anyone who wants to retrain or fine-tune the network. It is not needed for inference.

Attribution

The original implementation was developed by Georg Götz. Both estimators in this package, and the bundled network weights, come from his DecayFitNet toolbox, written with Sebastian J. Schlecht and Ville Pulkki and released under the MIT license. The network architecture, the training procedure that produced the weights, and the Bayesian analysis are all his work.

What this package adds is packaging, not method: the Python toolbox is ported from PyTorch to NumPy/SciPy, published to PyPI with the weights bundled, and given a common API that further estimators can plug into. The numerics are unchanged, and regression tests against the reference implementation are what verify that.

If you use this software, please cite:

@article{goetz2022decayfitnet,
  title   = {Neural network for multi-exponential sound energy decay analysis},
  author  = {G{\"o}tz, Georg and Schlecht, Sebastian J. and Pulkki, Ville},
  journal = {The Journal of the Acoustical Society of America},
  volume  = {152},
  number  = {2},
  pages   = {942--953},
  year    = {2022},
  doi     = {10.1121/10.0013416}
}

The Bayesian analysis follows Xiang et al., "Bayesian characterization of multiple-slope sound energy decays in coupled-volume systems", JASA 129(2), 741–752, 2011, and Jasa & Xiang, "Efficient estimation of decay parameters in acoustically coupled-spaces using slice sampling", JASA 126(3), 1269–1279, 2009.

Development

git clone https://github.com/artificial-audio/multislope
cd multislope
pip install -e ".[dev]"

python -m multislope.data download   # the regression tests analyse the example RIRs
pytest                               # full suite, including torch parity and slow Bayesian tests
pytest -m "not slow"                 # fast subset
ruff check .

The regression tests compare every estimate against tests/data/golden_upstream.npz, captured by running the reference PyTorch implementation. scripts/make_golden_fixtures.py regenerates it; see its docstring for how.

License

MIT. See LICENSE.