Room impulse response estimation by iterative weighted L<inf>1</inf>-norm

Abstract

This paper presents a novel method to solve for the challenging problem of acoustic Room Impulse Response estimation (RIR). The approach formulates the RIR estimation as a Blind Channel Identification (BCI) problem and it exploits sparsity and non-negativity priors to reduce illposedness and to increase robustness of the solution to noise. This provides an iterative procedure based on a reweighted l 1-norm penalty and a standard l 1-norm constraint. The proposed method guarantees the convexity of the problem at each iteration, it avoids drawbacks related to anchor constraints and it enforces sparsity in a more effective way with respect to standard l 1-norm penalty approaches. Experiments show that our approach outperform current state of the art methods on speech and non-speech real signals.

Uncalibrated 3D Room Reconstruction from Sound

“…[32,33]) or using unknown natural signals (e.g. exploiting the methods for blind RIR estimation [23,24]).…”

“…Finally, it is worth mentioning a set of blind method that relies on the intrinsic sparsity of the first part of the room impulse response (RIR) [21]. More recent approaches also include non-negative constraints to increase the accuracy of the solution [22,23,24]. These methods firstly estimate the RIR without the knowledge of the transmitted signal using an iterative sequence of convex problems.…”

“…Subsequently N M J matchings are evaluated for each position i of the grid in (17) and N M J score functions are computed by (18) and combined in (19). Once the j − th reflector has been estimated through (20) and (refg6) a variable number of delays < j is removed from each of the N M delay sets T n m,j+1 , as described in (22), (23). Finally a new set of N image sources pn j is considered for the computation of geometry delays in (16) at iteration j + 1.…”

Are Multiple Cross-Correlation Identities better than just Two? Improving the Estimate of Time Differences-of-Arrivals from Blind Audio Signals

“…Additionally, requiring i h i 2 2 = 1 as in ( 4) makes the AIRs to be co-prime [15] and constraint each of them to have a fixed norm -each of such requirements are likely to introduce numerical instabilities and artifacts during the optimization process. As a remedy for this, sparsity priors have been successfully applied to a broad spectrum of prior work in TDOAs estimation [1]- [6] [15], while also encompassing speech enhancement [16] and de-revereberation [8].…”

“…Error metrics. Once the AIRs have been computed through (8), we apply the peak finding method of [10] and we evaluate performance by means of two standard error metrics: the Average Peak Position Mismatch (A PPM ) and the Average Percentage of Unmatched Peaks (A PUP ) [15]. To ensure statistical robustness towards the random generation of reflections using [16], we performed Z = 50 random repetitions of the experiments using Monte-Carlo simulation [10].…”