[aguirreSL]

RoomAnalysis takes a new argument, edcMethod, which selects the noise compensation of the energy decay curve among the methods of ISO 3382-1 compared by Guski and Vorländer (2013):

edcMethod	Energy decay curve
'cutWithCorrection' (default)	truncated at the intersection with the background noise (Lundeby et al., 1995), plus the energy of the decay beyond it. This is the method RoomAnalysis used until now, so results do not change.
'justCut'	truncated at the intersection, with no correction
'noCut'	the whole response, noise included
'subtractNoise'	the background noise level subtracted from the squared response, then the whole of it integrated (Chu, 1978)

bypassLundeby=True raised UnboundLocalError, because the Lundeby parameters were returned without being set. It is now the 'noCut' method.

RoomAnalysis.uncertainty(param, numPositions=1, numDecays=10) gives the standard deviation of T20 or T30 by band from ISO 3382-1:2009, Equations (4) and (5). An impulse response counts as n = 10 decays, as clause 7.2 of the standard says. The bandwidth is fc (2^(1/(2 nthOct)) - 2^(-1/(2 nthOct))), which gives the 0.71 fc and 0.23 fc of the standard for octaves and thirds.

A second commit adds RoomAnalysis.PSNR, the peak-to-noise ratio of each band (the peak of the squared band response over the background noise level found by Lundeby, nan with 'noCut'), and RoomAnalysis.timbre_ratios(param='T30'), the bass and treble ratios of Beranek, BR = (T125 + T250) / (T500 + T1000) and TR = (T2000 + T4000) / (T500 + T1000). cumulative_integration now also returns the Lundeby parameters of each band. Its other caller, the deprecated pytta.rooms.analyse, already failed there, since it expects a single return value.

The tests in tests/test_room_edc.py use a synthetic exponential decay with T = 1 s over background noise 45 dB below its start. With it, T30 in the octaves from 500 Hz to 2 kHz is 1.01 to 1.03 s with 'cutWithCorrection', 0.98 to 1.00 s with 'subtractNoise', 1.00 to 1.02 s with 'justCut' and 5.1 to 5.5 s with 'noCut'. The tests check that the compensated methods find T within 5 %, that 'noCut' overestimates it and 'justCut' stays below 'cutWithCorrection', that bypassLundeby=True equals 'noCut', that the uncertainty follows Equations (4) and (5), that the PSNR of each band grows by 10 dB (9.91 to 10.02 dB measured) when the noise floor drops by 10 dB, and that BR and TR of a response with known decay per band come out within 5 % (1.288 against 1.267, 0.711 against 0.733).

Checked with Python 3.12, numpy 1.26 and scipy 1.17. The existing tests behave as on development: the 7 failures there (3 in test_h5io.py, 4 in test_pyttasave.py) are the same with and without this change.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com