Efficient GCI Detection for Efficient Sparse Linear Prediction

Abstract

Abstract. We propose a unified non-linear approach that offers an efficient closed-form solution for the problem of sparse linear prediction analysis. The approach is based on our previous work for minimization of the weighted l2-norm of the prediction error. The weighting of the l2-norm is done in a way that less emphasis is given to the prediction error around the Glottal Closure Instants (GCI) as they are expected to attain the largest values of error and hence, the resulting cost function approaches the ideal l0-norm cost function for sparse residual recovery. As such, the method requires knowledge of the GCIs. In this paper we use our recently developed GCI detection algorithm which is particularly suitable for this problem as it does not rely on residuals themselves for detection of GCIs. We show that our GCI detection algorithm provides slightly better sparsity properties in comparison to a recent powerful GCI detection algorithm. Moreover, as the computational cost of our GCI detection algorithm is quite low, the computational cost of the overall solution is considerably lower.

Spectral Zero-Crossings: Localization Properties and Applications

“…The identification rates based on Hilbert envelope are poorer, because the Hilbert envelope contains significant energy near the glottal opening instants. The problem of multiple peaks near epochs has been alleviated to some extent by enforcing sparsity constraints on the LP residue [37] or by using a weighted LP where the weighting allows for more energy concentration near epochs [38], [39].…”

Computational analysis of a fast algorithm for high-order sparse linear prediction

“…In other words, we are not interested in a solution of (5) but we are seeking a good enoughα ≈ α that captures the essence of our endeavor. For example, using (5) for speech coding purposes might require different accuracy than using it for modeling the speech glottal flow [44]. So, since we have used several approximations to formulate (5) it is likely that the performance as a function of the accuracy f (α (k) ) − f (α ) ≤ shows a saturation effect as a function of k (like in [11]).…”