|Abstracts of plenary talks|
|Statistical Methods for the Enhancement of Noisy Speech|
|Rainer Martin (Technische Universität Braunschweig)|
| With the advent and wide dissemination of mobile communications,
speech processing systems must be made robust with respect to environmental
noise. In fact, in noisy environments the performance of speech coders
or speech recognition systems is significantly degraded. As a result, speech
quality, speech intelligibility, or recognition rate requirements cannot
be met. However, improvements are obtained when the speech processing system
is combined with a speech enhancement preprocessor.
In this talk I will present algorithms for noise reduction which are based on statistics and optimal estimation techniques. The focus will be on estimation procedures for the spectral coefficients of the clean speech signal and on the estimation of the power spectral density of the background noise. While for the former I propose to use supergaussian priors, the latter is based on the recently developed "Minimum Statistics'' noise estimation technique. For the Minimum Mean Square Error estimation of speech coefficients I will present analytic solutions in the case of supergaussian priors and discuss their properties.
Furthermore, I will outline multi-microphone solutions and extensions which exploit properties of the human auditory system. The performance of these algorithms will be demonstrated with audio samples.
|The Automatic DJ: An appealing and instructive signal processing education project|
|Piet Sommen and Harrie van Meer (Eindhoven University of Technology (TU/e))|
| Classical courses are important to teach the individual student mathematical
and basic concepts. Laboratories can fill the gap between these concept
and their physical interpretation. These labs are mainly developed for
small groups, typically < 3 students.
An engineer rarely works alone to solve complex real world problems. Project education is a way to tackle such complex design problems in the education process. Students typically work together in groups of 6 - 8 students. These design problems are open-ended and allow for a wide variety of possible solutions involving engineering trade-offs between performance and cost. This makes the build up knowledge in courses and labs indispensable: basic courses, labs and education projects form a vital link in the engineering education process.
Since music and discotheques are appealing to young people, an education project has been defined around the theme "The Automatic DJ". The project groups have to design a programme on their notebook with which they are able to make a smooth transition from one music song, with a certain beat, to another music song, with a possible different beat, without degrading the quality of the music. The main signal processing challenge is to find and implement an efficient algorithm that performs time scaling (change the beat) without influencing the frequency scaling (preserve the pitch). The instructiveness of this education project lies in the fact that basic signal processing concepts have to be well understood and applied to a complex real world design problem.
|Independent Component Analysis and Its Application to Sound Signal Separation|
|Kiyotoshi Matsuoka (Kyushu Institute of Technology)|
| Independent component analysis (ICA) or blind source separation (BSS)
is a method for recovering a set of statistically independent signals from
the observation of their mixtures without any prior knowledge about the
mixing process. It has been receiving a great deal of attention from
various fields as a new signal processing technique.
In my experience, although most conventional methods for ICA are able to achieve separation for artificially synthesized data, they do not necessarily work well for real-world data. The results of separation are often unsatisfactory and, what is worse, they sometimes suffer from incomprehensible computational instability. In this talk I want to show a BSS algorithm that will gives us a desired separator in a relatively robust manner. The approach is fundamentally based on the minimal distortion principle, an idea for normalizing the demixing matrix with certain favorable properties.