AUDIS Miniproject: Environment Classification

Location: Sound and Image Processing Lab, Prof. Arne Leijon, KTH School of Electrical Engineering, Sweden
May 2010

Miniproject objectives

1. Classification of spatial noise types for HA algorithm
2. To better understand principles of standard classifiers

Problem definition

Design an environmental classifier that distinguishes between different spatial types of noise (directional, wind and diffuse noise) for a hearing aid scenario, utilizing two microphones to acquire input signals. The decision of the classifier is valuable for switching between different noise-reduction algorithms (e.g. a beamformer, wind cancelation or noise tracking methods).

Proposed algorithm

The first step is to extract features from the signal which represent characteristics of the different noise. In this project, the following features have been chosen: LPC cepstrum (in order to describe the overall spectral shape of the input signal); Fourier transformation of an interaural coherence function (describing the spatial properties of the noise); and a 4Hz modulation spectrum (to detect pauses between speech signals). The LPC cepstrum and the modulation spectrum extractors operate on monaural data, the interaural coherence uses input from two microphones

Two models have been considered: a Hidden Markov Model (HMM) and a Gaussian Mixture Model (GMM), both trained on a database of binaural signals obtained in different noise situations. The decision function of the classifier can be either a forward algorithm (for the HMM output) or a maximum likelihood method (for the GMM output). The feature extractors operate on a different timescale, therefore, supposing their stochastic independency, a separate model is trained on each feature set. As a training method the expectation-maximization (EM) algorithm was employed.

Participants:

• Anneke Lenssen
• Hongmei Hu
• Jalal Taghia
• Vaclav Bouse
• Yuva Vadlamudi