Noise reduction for vocal pathologies.

A noise reduction scheme, particularly suited for the correction of vocal pathologies, is proposed. The filter makes use of concepts originated within the theory of dynamical systems and deterministic chaos. In particular, the idea of embedding scalar data in order to reconstruct a phase space is of fundamental importance here. Furthermore, the concept of an attractor as a result of dynamical constraints is exploited. In order to perform noise reduction one needs redundancy and the human voice provides it even within a phoneme, namely the smallest structural unit of speech. Due to several repetitions of a pattern called pitch inside a phoneme, separation between the pure voice signal and the noise is possible, provided the latter is uncorrelated with the former. With a proper parameter tuning, different kinds of noise can be removed. We describe the idea behind the noise reduction algorithm and present applications to vocal pathologies.

Analysis of vocal disorders in a feature space.

This paper provides a way to classify vocal disorders for clinical applications. This goal is achieved by means of geometric signal separation in a feature space. Typical quantities from chaos theory (like entropy, correlation dimension and first lyapunov exponent) and some conventional ones (like autocorrelation and spectral factor) are analysed and evaluated, in order to provide entries for the feature vectors. A way of quantifying the amount of disorder is proposed by means of a healthy index that measures the distance of a voice sample from the centre of mass of both healthy and sick clusters in the feature space. A successful application of the geometrical signal separation is reported, concerning distinction between normal and disordered phonation.

Denoising human speech signals using chaoslike features.

A local projective noise reduction scheme, originally developed for low-dimensional stationary deterministic chaotic signals, is successfully applied to human speech. This is possible by exploiting properties of the speech signal which resemble structure exhibited by deterministic dynamical systems. In high-dimensional embedding spaces, the strong inherent nonstationarity is resolved as a sequence of many different dynamical regimes of moderate complexity.

Coping with nonstationarity by overembedding

We discuss how nonstationarity in observed time series data due to pronounced fluctuations of system parameters can be resolved by making use of embedding techniques for scalar data. If a D-dimensional deterministic system is driven by P slowly time dependent parameters, a (D+P)-dimensional manifold has to be reconstructed from the scalar time series, which is done by an m>2(D+P)-dimensional time delay embedding. We show that in this space essential aspects of determinism are restored. We demonstrate the validity of the idea heuristically, for numerical examples and for human speech data.