Tutorial I
-
Kalman Filters with Application to Speech Enhancement - Part 1
Prof. Dr.-Ing. Eberhard Hänsler
Institute of Telecommunications
Technische Universität Darmstadt
Darmstadt, Germany
Sunday, September 4th, 2010, 12:30 h - 13:30 h and 13:45 h - 14:45 h
Abstract:
To many engineers the Kalman filter is like a closed book. Reasons are manyfold. The strongest one might be, that the filter - consisting of a set of a equations - uses state space techniques, that might be unfamiliar to communication engineers. At the time of the development of the Kalman Filter, the already classical Wiener filter had been formulated by using impulse responses and transfer functions. With this techniques, however, the assumptions of stationary inputs and an infinite observation interval could not be overcome.
The purpose of this tutorial is to open this book. The emphasis is to explain the filter equations and to put them in perspective to results in system theory. Signal theoreticaly interesting steps in the development of the Kalman equations will be described in detail.
The tutorial starts with a brief discussion of the Wiener filter providing the background of the Kalman filter. The state space description of linear systems is covered in short. Then an extremely simple estimation problem is solved. The result can be written as a recursive algorithm exhibiting all the characteristics of the Kalman filter.
A state and a measurement equation form the starting point of the Kalman algorithm. The solution of the estimation problem requires three steps: Finding initial values, predicting the estimate based on already measured values and finally correcting the estimate according to the new measurement. A state diagram of the solution provides deeper insight into the Kalman estimation algorithm.
The tutorial closes with the example of a noise reduction problem solved by a Kalman filter.
-
Kalman Filters with Application to Speech Enhancement - Part 2
Dr.-Ing. Gerald Enzner
Institut für Kommunikationsakustik
Ruhr-Universität Bochum
Bochum, Germany
Sunday, September 4th, 2010, 15:00 h - 16:00 h
Abstract:
An important factor for the success of car hands-free telephones and speech dialog systems is the quality of the hands-free voice interface. In particular, we have to provide full-duplex ability and at the same time avoid deficiencies such as talker echo or unnatural background noise transmission.
In this tutorial, a relatively new model-based adaptive filtering approach is presented, which explicitly takes fundamental physical properties of realistic acoustic environments into account, e.g., the acoustic echo path variability, double talk, and ambient noise. For the time-varying acoustic echo path, a stochastic state-space model is introduced which is not found in textbooks yet. Based on the stochastic echo path model, a two-stage adaptive filter structure is derived for estimating the desired speech signal from the disturbed microphone signal in the minimum mean-square error sense. The key to the analytic derivation and accurate approximation of adaptive algorithms is the Kalman filter.
It turns out that state-space modeling leads to an outstandingly compact and robust signal processing solution for acoustic echo control. It does not require additional control mechanisms, such as double talk detection or echo path change detection, in order to achieve stable and high-end performance in realistic acoustic environments. The presented concept is thus suitable for today's voice communication systems, but it also provides a perspective for future research. Efficient implementations in time- and frequency-domain will be discussed. Finally, it is explained how the system properties were evaluated by a realtime prototype system in a car hands-free environment.
In summary, the tutorial presents the complete scientific development chain for car hands-free telephones, starting from the physical modeling of the acoustic environment, to the derivation of adaptive algorithms, as well as the realization and performance evaluation.
Tutorial II
-
Car Hands-Free Testing and Optimization - An Overview
Dr.-Ing. Hans-Wilhelm Gierlich
HEAD acoustics
Head of the Telecom Division
Herzogenrath, Germany
Sunday, September 4th, 2010, 16:30 h - 18:00 h
Abstract:
Although a lot of progress has been made in signal processing techniques which can be applied to car hands-free systems, the testing and optimization of such systems remains a challenging topic. Car hands-free systems include a variety of sophisticated non-linear and time-variant signal processing techniques intended to ensure good conversational quality in all different driving situations. All techniques used may impact the speech quality in sending, receiving, in conversational situations (double talk) and in the different background noise situations while driving.
The goal of all testing and optimization procedures developed over the recent years is to ensure and enhance (if needed) the speech quality for the driver but for the far-end conversational partner as well. This includes testing procedures on system level but also the optimization of the different subsystem such as acoustical components, wireless links used in the car, and the core hands-free algorithm.
The tutorial gives an overview about testing and optimization procedures which can be applied. Besides narrowband testing special requirements applicable for wideband testing will be introduced.