Blind Estimation of Reveberation Time from speech and music

Paul kendrick, University of Salford, Acoustics research centre, for more information please contact [email protected]

This program can compute the reverberation time from passivley recieved signals, such as speech and music. The advantage of this methodology is that it can be performed in-situ, with the room occupied. This code is free to use, under the condition that a reference is made to:

Paul Kendrick, Nicola Shiers, Robert Conetta, Trevor J. Cox, Bridget M. Shield, Charlie Mydlarz, Blind estimation of reverberation time in classrooms and hospital wards, Applied Acoustics, Volume 73, Issue 8, August 2012, Pages 770-780, ISSN 0003-682X, http://dx.doi.org/10.1016/j.apacoust.2012.02.010.

Description of algorithm

The algorithm is written in matlab, it contains a number of sub-functions, the following list describes the operation of the algorithm, for more details please consult the references at the end of this read-me.

• Choose wav file, this needs to be a recording of speech or music, for speech at least an hour, for music perhaps longer
• Read in section of wav file, and filter into octave band
• Maximum likelihood estimation of decays;
• Polyfit algorithm, this uses 0.5 s windows on the envelope of the signal, fitting a 1st order polynomial to the log envelope, the gradient of each fit is extracted, windows are moved along, using a very high overlap (0.002s), and regions which are continuously decaying identified from the gradient (gradient equivalent for RT==100s used as maximum gradient) - these decay phases are then fine tuned (start maximum, end min of LP filtered envelope)
• Maximum Likelihood fit to every decay phase, a model of sound decay in a room is where the envelope, alpha*a.^I+(1-alpha)*b.^I, modulates Gaussian white noise. This is a model of non-diffuse sound decay, utilising two exponential decays, with decay rates determined by a and b, added together using a convex sum where the balance between the two is controlled by the parameter alpha, all parameters are constrained between 0 & 1. The fit procedure first computes the likelihood over a coarse grid of values for a and b (alpha optimised for each grid point), then performs a fine search using the coarse result as a starting point, the function fmincon is used. This is constrained minimisation.
• Once all ML parameters are computed, the dynamic range of each decayphase is computed from the ML decay curve.
• Post-process results, for every octave band, a framework for estimating the RT is as follows, find the length of signal required to ensure that at least 40 decay phases with at least 25 dB of dynamic range are present, (this is just a rule of thumb), I found for speech 3-5 (segLen is this parameter - in secs) mins works. In each 3-5 min section, compute the RT from the decay phases with at least 25dB dynamic range, select the minimum RT as the estimate for that section. Compute the minimum RT for multiple sections. Average over all sections provides the blind RT estimate, the 95% CL from the standard error indicates the confidence in that result.

References

• Paul Kendrick, Nicola Shiers, Robert Conetta, Trevor J. Cox, Bridget M. Shield, Charlie Mydlarz, Blind estimation of reverberation time in classrooms and hospital wards, Applied Acoustics, Volume 73, Issue 8, August 2012, Pages 770-780, ISSN 0003-682X, http://dx.doi.org/10.1016/j.apacoust.2012.02.010.
• P Kendrick, TJ Cox, FF Li, Y Zhang, JA Chambers, Monaural room acoustic parameters from music and speech - The Journal of the Acoustical Society of America, 2008
• P Kendrick, FF Li, TJ Cox, Y Zhang, JA Chambers, Blind estimation of reverberation parameters for non-diffuse rooms, Acta Acustica united with Acustica, 2007