Berlin University of Technology Electrical Engineering and Computer Science electronics and medical signal processing Department of Energy and Automation Technology Chair of Electronics and Medical Signal Processing Prof. Dr.-Ing. Reinhold Orglmeister Annual Report 2008 Electronics und Medical Signal Processing Berlin University of Technology Department of Energy and Automation Technology Office EN 3 Einsteinufer 17, 10587 Berlin Tel.: +49 (0)30 314-21391 http://www.emsp.tu-berlin.de Email: Reinhold.Orglmeister@tu-berlin.de
Berlin, January 2009 Dear colleagues and friends of the Chair of Electronics and Medical Signal Processing (EMSP), dear interested readers, at the beginning of this year, we are taking the opportunity to present a short overview of last year s activities, in research and teaching. In our research activities, we have seen a very good start to two new projects in the areas of biomedical signal processing and audiovisual speech recognition, where our results have led to a number of publications and a couple of ongoing, promising activities for new project acquisitions. Our focus in the ongoing work remains on the area of signal processing in theory, software and hardware design for applications in medical electronics, speech signal processing and acoustics. In medical electronics, we have intensified collaborations with the center for innovative health technologies (ZIG) and introduced a new specialization "Medical Electronics" for master students in collaboration with the Chair of Medical Engineering. Also, our classes offered for students of electrical engineering and computer engineering are in great demand. Having many interested and highly motivated students, our laboratory projects, master s and bachelor s thesis projects again led to very impressive results, both regarding electronics design and signal processing projects. Hoping to have caught your interest, we cordially invite you to a visit of our internet presence at http://www.emsp.tu-berlin.de as well. Thank you to all research partners, sponsors and friends for your many contributions. With many greetings to everyone in and connected with the Chair of Electronics and Medical Signal Processing we wish a successful, happy 2009, 3
Research Evaluation and Comparison of the Independent Components of Simultaneously Measured MEG and EEG Data Master Eng. Heriberto Zavala-Fernandez (DAAD scholarship), Partner: Physikalisch- Technische Bundesanstalt (PTB), Fachbereich Biosignale (8.2) Magnetoencephalography (MEG) and Electroencephalography (EEG) are widely used methods for the study of neural activity. This activity can be represented as a linear superposition of different signals generated by the processes active in the brain as well as, signals generated by sources not belonging to the brain. Given this framework, Independent Component Analysis (ICA) is a well positioned technique for the analysis and comprehension of brain behavior. ICA has been already successfully applied to many brain activity studies employing either MEG or EEG signals in unimodal approaches. However, the complementarity of MEG and EEG recordings, due to the fact that magnetic fields and electric potentials arise from the same source, is well-known. Taking this into account, it is started from the premise that simultaneous recordings of MEG and EEG data investigated in the context of bimodal ICA, have the potential to improve the source isolation and localization by exploiting their complementarity. Bimodal ICA of simultaneous recordings also eliminates environmental confounds inherent to separate analysis of MEG and EEG. In this project, a technique is presented to exploit the signal complementarity in the bimodal space formed by MEG and EEG recordings by applying different ICA methods. The bimodal ICA components are validated by computing the corresponding Equivalent Current Dipoles (ECD) of selected sources, e.g. Alpha and Mu oscillations, Auditory Evoked Responses (AER). Figure 1 compares the averaged auditory evoked response with the results of unimodal ICA and bimodal ICA. Clearly, the estimated patterns are more regular for bimodal ICA when they are compared with the average response. In general, bimodal ICA produces similar patterns compared to unimodal ICA. 4
Figure 1: Maps of the corresponding auditory evoked response components as result of unimodal and bimodal ICA. The top row shows EEG modality and the bottom row MEG modality. The averaged response is also plotted for comparison with resulting components (left column). 5
Blind source separation for dependent source signals in evoked MEG Dipl.-Ing. Florian Kohl (DFG, OR 99/4-1), Partner: Physikalisch-Technische Bundesanstalt (PTB), Arbeitsgruppe Messdatenanalyse und Messunsicherheit s 1 s 2 s 3 s 4 s 5 Epoch 1 Epoch 2 in the first epoch of Fig. 1a. If all epochs are equal to the first epoch the resulting signals in the 2D subspace are highly dependent. In order to lower their dependence, a jitter parameter is introduced. More specifically, in each epoch the fixed response latencies are changed by two independent Gaussian distributed random numbers with zero mean and standard deviation σ leading to a jitter in the response latencies. Furthermore, an ordering assures that one source always responds prior the other. For convenience, the standard deviation σ is referred to as jitter in the following discussion. Hence, one may gradually change the dependence of the source signals by gradually changing the value of the jitter. The more jitter the less dependent are the signals and vice versa. A setting with 40 ms jitter is depicted in Fig. 1a. We refer to this model as Synthetic Stimulus Evoked Jittered Response (SSEJR) model, which uses a single parameter to gradually change the mutual information of synthetic MEG source signals. The 3D subspace is modeled likewise. In Fig. 1 the whole 2D+3D setting is depicted. The same jitter parameter is used to adjust the dependence within the subspaces. These are made almost independent by choosing the source specific latencies such that signals from different subspaces do not overlap. The dependence of a set of SSEJR generated signals will be evaluated in terms of pairwise mutual information, given by 2D 3D beginning of the responses and not their shapes vary. In contrast, the joint distribution p( s, s ) becomes more flat with i j increasing jitter as more sample combinations occur. Hence, the mutual information decreases and the joint distribution approaches the product of the marginals. This is why the generated SSEJR signals are assumed to become more independent by increasing the jitter. A homogeneous conducting sphere model serves as an approximation of the human head, where stimulus evoked currents in the brain are modeled by equivalent current dipoles. 25 such dipoles are placed, five of them being the source dipoles with 2D and 3D stimulus evoked source signals generated by the SSEJR model. The remaining 20 dipoles are randomly placed and represent interfering noise processes. Each noise dipole follows a 7th order AR process activated by Laplacian distributed white noise with coefficients that are obtained from real prestimulus MEG data. Subsequently, the induced magnetic field is calculated solving the quasistationary Maxwell equations for the dipole currents and respective volume currents [10]. Sensor coordinates from the PTB 93-channel MEG system are used. Furthermore, sensor noise is introduced with an average signal to noise ratio of In this project, we created p( s1, s2a) new simulation I( s, s ) = 1 2 p( s, s ) log ds ds, (2) 30 db. tool for the generation of synthetic 1 2 1 2 p( s1 ) p( s2 ) Following (1), the mixing matrix A emerges from the MEG, namely the Synthetic Stimulus Evoked Jittered Response (SSEJR) model (cf. contribution of each source to each sensor and their Figure which 2 equals ). Source the amount signals of uncertainty canreduction be generated of one superposition at random gives the by synthetic a mixture observed ofmeg Gaussians, data signal gained from knowing the other. Following (2), the vector x (t). Fig. 1a,b illustrate source and noise current while SSEJR the model signal seems dependency plausible. The marginal is gradually distributions adjustable. Furthermore, the number and processes and their associated field maps, while Fig. 1c,d p( s i ) and p( s j ) of any pair of signals within a subspace stay dimensionalities of subspaces can be chosen illustrate freely. the current Hence, dipole amodel flexible and typical and realistic sensor unchanged regardless of the amount of jitter, as only the observations x generated. model of evoked MEG serves for the development and assessment of blind source separation in the dependent source signal scenario. 20 0-20 20 2 0 0.5 1 x 1 0 n 20 x 1D i -20 0 time 0.5 / s 1 0 1 time / s 2 a) Time signals b) Field maps c) Current dipole model d) Synthetic MEG data Figure 1: Synthetic MEG data. (a) 2 epochs of weakly dependent source signals are generated by using the SSEJR model with a high jitter value. Lowering the jitter, all epochs will eventually have responses as depicted in the first epoch leading to two independent signal subspaces (2D and 3D) with highly dependent components. The last row depicts one representative AR noise process. (b) The field maps emerge from the associated source and noise dipoles, respectively. The noise field map is a superposition of all noise field maps emerging at the first time instant. (c) Homogeneous conducting sphere model of the human head with 5 source dipoles (black) and 20 randomly placed noise dipoles (gray). Sensor coordinates (diamonds) correspond to PTB 93-channel MEG system. (d) 2 typical sensor observations exemplify the generated MEG data. Figure 2: Synthetic evoked MEG with gradually adjustable source signal dependencies: The Synthetic Stimulus Evoked Jittered Response (SSEJR) model. Magnetoencephalography (MEG) is a non-invasive technique to record brain activity. In order to better understand the functioning of our brain, stimulus experiments are often conducted. The location and the time dynamics of an evoked neural B. Synthetic MEG data current are of interest. However, the resulting magnetic fields from different evoked sources superimpose at each MEG channel. Hence, the observer has access only to a mixture of signals. Independent component analysis (ICA) is often used to separate the MEG data. Nevertheless, the key assumption of ICA is source independence, which may not be given in stimulus experiments. For example, different sources responding to the same stimulus may have a similar activation and termination time leading to a correlation in their energies.... x 93 MBEC_939.doc 6
All tested ICA methods that we applied to SSEJR MEG data fail for depe critically, for nearby spaced dipoles, even physiologically meaningful resu significantly wrong. This is depicted in the figure where the dipolar struct sources does not indicate a false d However, these field maps differ f field maps to a great extent. Our findings ask for non-independ new methods are being investigate (ISA) and to unmix dependent sou Original Infomax Figure 3: Typical field maps from simulations with 2 cm spaced dipoles. For dependent SSEJR source signals, Infomax recovers a clear dipolar pattern as an estimate suggesting successful unmixing. Nevertheless, this physiologically apparently meaningful ICA result is significantly wrong. Applied to simulated MEG data, all tested ICA methods fail for dependent sources. Most critically, for nearby spaced dipoles, even apparently physiologically meaningful results turned out to be significantly wrong. This is depicted in Figure 3 where the dipolar structure of the recovered sources does not indicate a false decomposition. However, these field maps differ from the original field maps to a great extent. Our findings ask for non-independent blind source separation. Currently, new methods are being investigated in order to group (i.e. independent subspace analysis) and to unmix dependent source signals. 7
Wireless Body Sensor Network and Signal Processing for Low-Power Motion-Tolerant Vital Sign Measurement Dipl.-Ing. Achim Volmer Personal Healthcare Systems are expected to be revolutionary in many applications ranging from prophylaxis to rehabilitation of cardiovascular disease. Considering the demographic change leading to a higher quota of middle aged and elderly people the advantages of Healthcare Systems for domestic use are outstanding as they enable patients to remain in their familiar environments with only few constrictions. Figure 4: Body Sensor Network In this project a variety of biological signals like the electrocardiogram (ECG), photoplethysmogram (PPG), phonocardiogram (PCG) and oxygen saturation are acquired to allow the estimation of the health state by seperate or combined evaluation. For 24/7 monitoring good signal coverage can be identified as the gravest problem because motion artefacts continuously degrade the signal quality. Wearable sensors must be robust against disturbances, in particular, to the motion of the wearer. However, reliability is by far inferior to their traditional counterparts, leading to frequent false alarms. The goal of this project is to enhance the signal quality of each biosignal and to increase the robustness und reliability of the extracted vital signs to estimate the global health state. 8
Figure 5: ECG belt - Block Diagram Wired systems would not be acceptable as they reduce mobility. This implicates the use of battery-powered systems. Moreover size and weight have to be minimized which is conflicting to long operating times. Wireless communication is essential to transmit biosignals to a base station and to provide the patient with feedback about the health state. Therefore a Body Sensor Network (BSN) was already developed that is capable of collecting vital signs by the help of wearable sensor systems attached to adequate locations all over the body. Each sensor system consists of a MSP430 low-power microcontroller by Texas Instruments that acquires at least one biosignal and is combined with tri-axis accelerometer to estimate the motion and posture. Figure 6: Chest located ECG belt with textile electrodes The upcoming low power wireless link according to the physical layer (PHY) and Media Access Control (MAC) of the IEEE 802.15.4 standard was implemented to 9
archive a robust low-latency communication with a coordinator and to establish a secure real-time monitoring. The motion artefact reduction is separated into two stages. Weak artefacts are compensated directly on each sensor system by help of adaptive filtering followed by vital sign extraction. This reduces the amount of data that has to be transmitted via the wireless link whenever the biosignal is quite good. With strong motion artefacts the raw data is transmitted and data fusion algorithms are used to reconstruct the biosignals within the base station. Figure 7: PPG Fingerclip Sensor Some work has also been done for the development of data fusion algorithms for a base station (PC or PDA) to reconstruct corrupted sensor data and to apply pattern recognition. To reconstruct the highly corrupted PPG signals the influence of motion has been investigated and different algorithms using mutual information of the biosignals have been implemented. One important goal of future works is the implementation and testing of further reconstruction algorithms especially with help of statistical signal processing. It is also intended to implement an expert system to extract simple vital parameters supporting medical diagnosis. The implementation of further embedded algorithms on the sensor systems also is a challenge. 10
Network controllers in education, development and research Manuel Borchers, Dipl.-Ing. Achim Volmer Partners/Sponsors: Hilscher GmbH, Hitex Development Tools The future of automation technology lies in individuality, in distributed systems with local intelligence. Systems solving automation tasks completely independent are interconnected by uniform communication structures. This enables a consistent data access from management to the field level. The use of IT technology and the Internet are already well established. Worldwide communication is carried out via open interfaces with standard tools such as Ethernet with TCP / IP and Web browser. In the past two years several interesting projects have been implemented together with groups of students during their participation in the microcontroller project course. The following sections show projects that were realized by students during the microprocessor project course using the innovative network processor family netx. Streaming Audio Network Participants: B. Erik, M. Fritz, Ch. Leiste, J. Pfeil, N. Raehse, H. Rinke et al. Based on netx starter boards, a streaming audio system was developed. The aim was to stream audio data from several sources (e.g. analogue sources and SD cards) throughout an Ethernet network to different client systems. The clients receive the audio data as data frames from the network and transmit them to an output device. These outputs include speakers or a series of LEDs which visualize the spectrum of the transmitted audio stream. A control unit was designed to establish the connection between the audio sources and distinct clients. This way, several audio streams can be transferred to different clients over the same Ethernet even at the same time. For a better usability a graphical user interface controlled by a touch screen was implemented. The interface is easy to use: Clicking and connecting a source with a sink is enough. File based sources also have buttons for skipping tracks or stopping the player. A sample image of the user interface can be seen in Figure 8. The system has been presented on the international trade fair Electronica 2008 in Munich very successfully. The whole system was realized using just netx 500 boards. 11
Figure 8: Screenshot of the control unit displaying two sources (one connected) and three output devices Figure 9: Photo of the Audio Network presented at Electronica 2008 12
Wireless Network Webcam Robot Participants: B. Fischer, M. R. Pourhashemi, R. Kavalakkatt, T. Winter, T. Förster, X. Wang et al. In this project a tracked vehicle (see Figure 10) was equipped with a CMOS camera and a wireless network link. The link using standard IEEE802.11b WLAN is used to transfer control information for the movement of the robot as well as sending the captured video data to a base station. Figure 10: Photo of the completely equipped Webcam Robot The processing of the image data is done using a netx 50 board which features a direct interface for CCD and CMOS video sensors. The video data is then compressed on the same board before sending the live data over the wireless link. Motor control is also done on this board. A second netx board (using a netx 500 processor) equipped with a touchscreen display is used to display the video data. The touchscreen is used as a simple method of interacting with a human to control the movement of the robot. 13
Robust Speech Recognition Dr.-Ing. Dorothea Kolossa, Dipl.-Ing. Eugen Hoffmann, Dipl.-Ing. Ramon Fernandez Astudillo, Dipl.-Ing. Alexander Vorwerk, Master Eng. Georgios Tsontzos Speech recognition in reverberant conditions and under the influence of noise and interfering speech remains a difficult task. The aim of the robust speech recognition group is to investigate a range of solutions to this problem, which includes multichannel statistical signal processing, uncertain speech recognition and the inclusion of video information in the recognition process. In statistical speech processing, a focus of the work lies in independent component analysis. This approach utilizes higher order statistical information in order to recover the speech signals from noisy and reverberant mixtures of multiple speakers. Especially in combination with a nonlinear time-frequency mask, this method has proven successful for enhancing the signal quality as well as the recognition results, as described in the subsequent section on multi-channel signal processing. However, in order for speech recognition to be successful after time-frequency masking, the use of uncertainty information in a so-called missing-feature recognizer is vital. This task and its recent solutions, especially for speech recognition under noisy conditions, are described in the section "Uncertainty Estimation and Propagation for Speech Recognition". Finally, audiovisual speech recognition can be a good solution even under extremely difficult acoustical conditions, since a video signal of the speaker is invariant to acoustical disturbances. Therefore, audiovisual speech recognition has been included as a new area of research, and the promising results shown in the section "Audiovisual Speech Recognition" have led to a recently submitted project proposal and to a number of ongoing masters and bachelors theses to improve and better employ visual information within the missing feature recognition framework. 14
Multi-channel Signal Processing Dipl.-Ing. Eugen Hoffmann, Dr.-Ing. Dorothea Kolossa Figure 11: Overview of the algorithm structure. The work deals with the reconstruction of disturbed speech and aims at an improvement of the speech quality and intelligibility in noisy environments and in environments in which several speakers are active simultaneously. To make disturbed or noisy speech signals usable for automatic speech recognition systems, statistical properties of the speech signals and noise signals can be used. An example of the application of statistical methods is the Independent Component Analysis (ICA) which can be used for the extraction of signals of different speakers from a signal mixture. The basic idea of ICA consists in disassembling multi-channel signals into statistically independent components. At the moment, essential problems exist during the application of the acoustic blind source separation. On the one hand, the available ICA solutions are extremely time-consuming which makes a practical use of the algorithms difficult. On the other hand so called permutation and scaling problems appear which possibly cannot be solved unambiguously. So the main research interests of this group lie in (real time) implementation, investigation and test of already existing and new ICA and beamforming algorithms in real environments (e.g. cars, business premises etc.). The actual research point is to deal with the permutation problem using information theoretic distance measures on the approximated probability density functions of the subsequent frequency bins of the separated signals. Another research aspect is the improvement of the separation results by use of the time-frequency masks. The main problem here is to find the time frequency points in the signal where remaining interferences or noise signals are active and then to minimize them with a suitable time frequency mask. For this purpose, different 15
Figure 12: Time Frequency Masking. The spectrograms of the demixed signals (left column) the generated masks (middle column) and the masked signals (right column). features of signals can be used, for instance the direction of arrival (DOA), amplitude difference or multi-channel voice activity detection and a Bayesian integration of multiple such criteria. Besides the application of the ICA and a time-frequency masking building up on it other methods of statistical signal processing are also examined, amongst other procedures like the Ephraim and Malah filter and similar methods, as for example the IMCRA procedure which can also adapt the noise signal model, while the signal of interest is active. 16
Uncertainty Estimation and Propagation for Speech Recognition Dipl.-Ing. Ramon Fernandez-Astudillo One of the main problems that automatic speech recognition still has to solve is robustness against noise. A little amount of noise, which would only slightly annoy a human being, still is fatal for most automatic speech recognizers (ASRs). To cope with this limitation, many statistical noise suppression procedures have been developed to partially eliminate interfering signals before the recognition takes place. Unfortunately, due to the unpredictable nature of noise and the limitations of the models used to describe the noise and speech processes, these methods often also degrade the quality of speech, especially in instationary-noisy environments. To deal with this degradation we have studied the use of supergaussian models [1] to describe the noise process more accurately trough a data-driven technique. Figure 13: Left: Histogram (blue), Gaussian (green) and data driven (red) prior distributions of the real component of noise D. Right: Corresponding posterior distributions of clean amplitude A given noisy signal Y and Minimum Mean Square Error - Short-Time Spectral Amplitude (MMSE- STSA) estimations for the Gaussian (green) and supergaussian (red) models. Another research aspect was the development of a black-box approach for the estimation of the uncertainty derived from the noise suppression process. This estimation of uncertainty establishes bounds in form of a probability density function for the position of the true clean signal X given the estimation ˆX obtained from the noise suppression process. 17
Artificialy added Noise Uncertainty Estimation Estimated uncertainty variance (a) Speech Noise + Suppression - (b) True uncertainty Figure 14: Left: Obtainment of true uncertainty and estimated uncertainty variance by adding noise to speech. Right: True (blue) and complex Gaussian (red) uncertainty distributions given the estimator output for (a) high uncertainty and (b) low uncertainty. Only real part of complex uncertainty is displayed. This probabilistic description can be later used for re-estimation of the clean signal in the feature domain by using uncertainty propagation techniques [2]. The black-box approach is aimed to judge the goodness of a given uncertainty estimator and uncertainty model without taking into consideration the internal structure of the noise suppression procedure. This was achieved by constructing the histogram of the joint probability distribution of uncertainty δ = X ˆX and the estimated uncertainty variance λ from a set of experiments in which noise was artificially added to speech. By using the uncertainty obtained from a given estimator λ rather than the true uncertainty λ we take into consideration the imperfection of the uncertainty estimation process and thus we are able to correct any bias created. References [1] Martin, R.: Speech Enhancement Based on Minimum Mean-Square Error Estimation and Supergaussian Priors. In: IEEE Trans. on Speech and Audio Processing, Vol. 13, September 2005, pages 845 856. [2] Astudillo, R.F.; Kolossa, D.; Orglmeister, R.: Propagation of statistical information through non-linear feature extractions for robust speech recognition. In: Proceedings MaxEnt07, 2007, pages 245 252. 18
Probabilistic Graphical Models in Speech Recognition Dipl.-Ing. Georgios Tsontzos Statistical applications in fields such as bio-informatics, speech processing, image processing and communications often involve large-scale models in which thousands or millions of random variables are linked in complex ways. Graphical models (GM) provide a general methodology for approaching such problems in a modular, flexible and structured way. The field of speech recognition falls into the same category, especially multimodal speech recognition (MSR), which combines different kinds of information, such as audio, visual, and linguistic knowledge, for a common goal. The first task was to investigate of how an HMM can be translated to a GM. There were many ways to attain such a translation, since many structures of GM can provide this kind of behaviour. The choice of applying causal GMs was made since the influence between the different kinds of input sources can vary between modalities in an MSR system, so causal relations are needed to specify how the modalities influence each other (see Figure 15). Subsequently, a first approach Figure 15: Structure of a GM for multimodal speech recognition. was implemented with limited complexity, using only discrete variables in a static GM. The implementation was carried out in Java, in order to benefit from the reusability and extensibility that this object oriented programming language can offer. In the upcoming year, the task is to expand the framework with continuous and deterministic variables and to add dynamic behaviour. Finally, learning and classification of a complex model will be carried out for evaluation and assessment of the system. 19
Audiovisual Speech Recognition (AVSR) Dipl.-Ing. Alexander Vorwerk, Dr.-Ing. Dorothea Kolossa The inclusion of features extracted from a video stream into a speech recognizing system is considered to be helpful in increasing the robustness above that of audio-only systems. Several aspects have to be taken into account while realising a combined audiovisual recognition system. Calculating useful features from a video stream requires the availability of appropriate training data as well as algorithms to find the position of the face combined with the extraction of a region of interest (ROI, e.g. the mouth region). After the calculation of different features, the possibility to evaluate the relevance of the extracted feature values is desired in order to be able to weigh the video stream while combining it with the audio stream. Depending on the fusion strategy for the streams, interpolation of one of them may be needed to synchronize both the video and the audio data. In a first step, the robust extraction of the mouth region based on color information combined with geometric assumptions has been implemented and tested on the so called GRID database, that includes audio and video recordings of 34 speakers [1]. Figure 16 shows some example video frames and the associated results of the mouth region finding. Figure 16: Sample frames of different speakers from the GRID database (upper row), their mouth regions extracted by the AVSR system (middle row) and the corresponding DCT coefficents (bottom row) In order to evaluate the influence of added video features, the extracted mouth regions have been transformed by a 2D-discrete cosine transformation (DCT). A subset of the DCT coefficients was presented to the Java Audiovisual SPEech Recognizer (JASPER) as video feature without any further adjustments. As can be 20
seen in Figure 17, the integration of this quite simple video stream has improved the result of the speech recognition at low signal-to-noise ratios significantly. Recognition Rate Figure 17: Example for the improvement of recognition results attained by use of the 2D-HMM-recognizer JASPER, employed on audiovisal data. The curves show the recognition rate, obtained from (N Del Sub Ins)/N where Del denotes the number of deletion errors, Sub the substitutions and Ins the insertions and N is the total number of words. Test results were obtained on 1000 sentences with a 52 word vocabulary, after the speech signal had been artificially corrupted by additive white noise at the shown signal to noise ratios (SNRs). Further aspects in terms of video processing for AVSR are presently investigated. Especially using stereo recordings of a speaker and utilizing depth maps obtained from stereo imaging for video feature extraction is a main goal of this project, wherein a knowledge on the uncertainty of the provided features is also of great interest. References [1] Cooke, M.; Barker, J.; Cunningham, S.; Shao, X.: An audio-visual corpus for speech perception and automatic speech recognition. In: Acoustical Society of America Journal, Vol. 120, 2006, pages 2421-2424. 21
Doctoral Theses [1] Dorothea Kolossa: Independent Component Analysis for Environmentally Robust Speech Recognition. PhD-Thesis, reviewer: Prof. Dr.-Ing. Reinhold Orglmeister, Prof. Dr.-Ing. Te-Won Lee (University of California, San Diego), Berlin, (May 2008). [2] David Dmitry Polityko: Physikalischer Entwurf für die vertikale SIP-Integration. PhD-Thesis, reviewer: Prof. Dr.-Ing. Herbert Reichl, Prof. Dr.-Ing. Reinhold Orglmeister, PD Dr. habil. Karl-Heinz Küfer (Fraunhofer ITWM Kaiserslautern), Berlin, (June 2008). [3] Tobias Lorenz: Advanced Gateways in Automotive Applications. PhD-Thesis, reviewer: Prof. Dr.-Ing. Otto Manck, Prof. Dr.-Ing. Reinhold Orglmeister, Berlin, (July 2008). [4] Heriberto Zavala-Fernandez: Evaluation and Comparison of the Independent Components of Simultaneously Measured MEG and EEG Data. PhD- Thesis, reviewer: Prof. Dr.-Ing. Reinhold Orglmeister, Prof. Dr.-Ing. habil. Jens Haueisen (TU Ilmenau), Dir. u. Prof. Dr. Lutz Trahms (Physikalisch-Technische Bundesanstalt), Berlin, (submitted November 2008). Publications [1] Zavala-Fernandez, H.; Sander, T.H.; Burghoff, M.; Trahms, L.; Orglmeister, R.: Occipital alpha components found in a joint ICA of combined MEG and EEG data. In H. Malberg, T.H. Sander, N. Wessel, and W. Wolf (Hrsg.), DGBMT- Workshop Biosignalverarbeitung 2008. Potsdam, Germany, (July 2008), p. 199 202. [2] Volmer, A.; Orglmeister, R.: Wireless Body Sensor Network for low-power motion-tolerant synchronized vital sign measurement. In Proc. 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society EMBS 2008. Vancouver, Canada, (August 2008), p. 3422 3425. [3] Astudillo, R.F; Kolossa, D.; Orglmeister, R.: Uncertainty Propagation for Speech Recognition using RASTA Features in Highly Nonstationary Noisy Environments. 8. ITG-Fachtagung Sprachkommunikation (ITG08), (October 2008). [4] Kolossa, D.; Araki, S.; Delcroix, M.; Nakatani, T.; Orglmeister, R.; Makino, S.: Missing Feature Speech Recognition in a Meeting Situation with Maximum 22
SNR Beamforming. In Proc. ISCAS 2008, Seattle, USA, (May 2008), p. 3218 3221. [5] Kolossa, D.; Hoffmann, E.; Orglmeister, R.: ICA-Based Bayesian Time- Frequency Masking. 8. ITG-Fachtagung Sprachkommunikation (ITG08), (October 2008). [6] Vicinus, P.; Kitzenmaier, P.; Orglmeister, R.: Erfassung von Sprache im KFZ mit Störgeräuschreduktion im Nahfeld eines Mikrofon-Sparse-Arrays. In Prof. Dr. phil. Ute Jekosch; Prof. Dr.-Ing. habil. Rüdiger Hoffmann (Hrsg.) Fortschritte der Akustik - DAGA 2008, Deutsche Gesellschaft für Akustik e.v., Dresden, (March 2008). [7] Kohl, F.; Wübbeler, G.; Sander, T.; Trahms, L.; Kolossa, D.; Orglmeister, R.; Elster, C.; Bär, M.: Performance of ICA for dependent sources using synthetic stimulus evoked MEG. In DGBMT-Workshop Biosignalverarbeitung 2008, Potsdam, Germany, (July 2008), p. 32 35. [8] Kohl, F.; Wübbeler, G.; Kolossa, D.; Orglmeister, R.; Elster, C.; Bär, M.: Performance of ICA for MEG data generated from subspaces with dependent sources. In 4th European Congress for Medical and Biomedical Engineering, Antwerp, Belgium, (November 2008). Diploma Theses [1] Hoang Tung NGUYEN PHAM: Integration eines Mikrocontroller-Softcores in die Architektur eines Video-Formatwandlers und Implementierung eines realzeitfähigen Steuerprogramms. External diploma thesis at Frauenhofer Institut für Nachrichtentechnik (Januar 2008) [2] Quyet VU TRUNG: Auslegung, Inbetriebnahme und Erprobung einer potenzialfreien Zwischenkreis-Spannungserfassung für die Regelung und Grenzwertüberwachung eines Antriebssystems für Hybridfahrzeuge. External diploma thesis at Siemens AG (Februar 2008) [3] Taner YILMAZ: Erprobung von Ultraschall- und Wirbelstromlösungen zur Absicherung und Detektion von Sicherheitsmerkmalen für den Marken- und Produktschutz. External diploma thesis at Fak. V, PTZ 5 (Februar 2008) [4] Markus HAAG: Einsatz berührungsloser Temperatursensoren in Personenfahrzeugen zur Fußgängerlokalisierung. External diploma thesis at Daimler AG Ulm, (Februar 2008) [5] Michael MUTHIG: Entwicklung eines Programms zur interaktiven und automatisierten Auswertung anisotroper Streudaten von nanostrukturierten 23
Systemen. External diploma thesis at Fak. II, Fachgruppe Physikalische und Theoretische Chemie (April 2008) [6] Zakaria KASMI: Weiterentwicklung eines Indoor-Positionierungssystems auf der Basis von Ultra-Wide-Band-Verfahren. External diploma thesis at TU Darmstadt (April 2008) [7] Long JI: Optimale Versuchsplanung zur Modellbildung für nichtlineare dynamische Systeme (part 1). (Juni 2008) [8] Huaiyu ZHANG: Optimale Versuchsplanung zur Modellbildung für nichtlineare dynamische Systeme (part 2). (Juni 2008) [9] Sebastien BREILMANN: Entwicklung eines MC68332-Mikrocontrollerbasierten Gateways zur Erweiterung eines drahtgebundene Feldbusses um TCP/IP und Wireless LAN. (Juni 2008) [10] Andreas TANGL: Untersuchung zur Klassifikation von Inrush-Vorgängen mit neuronalen Netzen. External diploma thesis at Siemens AG (November 2008) [11] René KNOBLICH: Konzept und Implementierung eines Host-Client-Systems auf der Basis eines Java-GSM-Moduls und eines MSP430-Mikrokontrollers. External diploma thesis at Siemens AG (Juli 2008) [12] Peggy Laure NGUETSE JONGO: Berechnung visueller Features für multimodale Spracherkennung. (Juli 2008) [13] Ning MA: Entwicklung von HMM für deutsche, kontinuierlich gesprochene Sprache. (Juli 2008) [14] Piere KOLASA: Implementierung eines intelligenten Sensormoduls zur Phonokardiographie. (September 2008) [15] Michael SATTLER: Low-Power-Modul zur mobilen Erfassung der Sauerstoffsättigung und der Herzrate. (September 2008) [16] Andreas WILDGRUBER: Entwicklung und prototypischer Aufbau eines automatischen Prüfplatzes für eine Reizschwellenmessgerät zur Unterstützung der intraoperativen Elektrotherapie des Herzens. External diploma thesis at Biotronik GmbH & Co.KG. (September 2008) [17] Philipp GÜNTHER: Automatischer Messplatz zur Prüfung der elektromagnetischen Störfestigkeit von implantierbaren Cardioverter/Defibrillatoren (ICD). External diploma thesis at Biotronik GmbH & Co.KG. (September 2008) [18] Holger KIRCHHOFF: Synchronisation von Audiosignalen. External diploma thesis at zplane development (Oktober 2008) [19] Thomas WOTSCHKE: Entwicklung einer Test-Software zur Prüfung von aktiven Implantaten bei Ansteuerung eines vorhandenen Herzsimulators. External diploma thesis at Biotronik GmbH & Co.KG. (Oktober 2008) 24
[20] Michael SCHÜLER: Entwurf und Implementierung eines UDP-Layer-Stacks in VHDL für Gigabit-Ethernet-Streaming-Anwendungen. External diploma thesis at Heinrich-Hertz-Institut (Juli 2008) [21] Janne HAHNE: Entwicklung und Evaluierung von Steuerungskonzepten für eine neue Generation von Myoprothesen. External diploma thesis at Otto Bock Healthcare, Wien (Oktober 2008) [22] Sebastian FEESE: Implementierung und Erprobung von Algorithmen zu Fusion und Rekonstruktion gestörter Vitaldaten. (November 2008) Master Thesis [1] Philipp STÄUBER: Kontextbasierte Signalanalyse für implantierbare Herzschrittmacher. External master thesis at Biotronik GmbH & Co.KG. (September 2008) Student Research Projects [1] Fernando Rueda BALAREZO: Implementierung und Vergleich von Zeit- Frequenz-Maskierungsverfahren für Mikrofonsignale. (Januar 2008) [2] Marcus KLEINERT: Grundlagen der auditorischen Signalverarbeitung. External student research project at Telekom Labs (April 2008) [3] Sarah FELL: Methoden zur Permutationskorrektur. (Mai 2008) [4] Oliver BENJAK: Auslegung und Test von Komponenten zur Energieversorgung eines Picosatelliten. External student research project at Fak. V, LRT (Juli 2008) [5] Manuel BORCHERS: Erstellung eines Gateways zwischen einem Body-Area- Network und Matlab. (November 2008) 25
The Team Head of Department Prof. Dr.-Ing. Reinhold Orglmeister Office EN 3 Mrs. Edeltraud Esser Scientific Assistant Dr.-Ing. Dorothea Kolossa Research Assistants Dipl.-Ing. Eugen Hoffmann Dipl.-Ing. Achim Volmer Dipl.-Ing. Alexander Vorwerk Dipl.-Ing. Florian Kohl Ph.D. Students Dipl.-Ing. Ovidiu Codreanu Dipl.-Ing. Ramon Fernandez Astudillo Master Eng. Heriberto Zavala-Fernandez Master Eng. Georgios Tsontzos Guest Researchers Prof. Dr.-Ing. Hans-Heinrich Bothe (DTU, Technical University of Denmark) Dr.-Ing. Bert-Uwe Köhler (IAV GmbH) Dr.-Ing. Guntram Liebsch (Siemens AG) Dr.-Ing. René Straßnick (TU Berlin) Dipl.-Ing. Henry Westphal (Tigris Elektronik GmbH) Dipl.-Ing. Steffen Zeiler Student Teaching Assistants (Tutors) Manuel Borchers (Praktikum Digitale Systeme - netx, sponsored by Hilscher GmbH, Hattersheim) Bennet Fischer (Projekt Elektronik) Janis Döbler (Projekt Elektronik) Lilo Timm (Projket Elektronik), until 09/2008 Maik Pflugradt (Mikroprozessortechnik) 26
Staff common to the Chairs of "Electronics and Medical Signal Processing" (Prof. Orglmeister) and "Electronic Measurement and Diagnostic Technology" (Prof. Gühmann): Office Mrs. Edeltraud Esser (EN 3 EMSP) and Ms. Brigitte Auerbach (EN 13 MDT) Institute Engineers Dipl.-Ing. Rüdiger Seidel and Dipl.-Ing. Frank Baeumer Electronic Technicians Michael Hackbarth and Hans-Ulrich Timm Mechanic Technicians Peter Jaeck and Uwe Kurlbaum Teaching Summer Term 2008 VL IV VL UE VL VL SE SE PJ Medizinelektronik (Orglmeister) - Medical Electronics - Medizinelektronik (Orglmeister, Vorwerk) - Medical Electronics - Einführung in die automatische Spracherkennung (Kolossa) - Introductory Automatic Speech Recognition - Rechenübungen zur automatischen Spracherkennung (Kolossa) - Introductory Automatic Speech Recognition - Neuro-Fuzzy-Methoden (Bothe) - Neuro-Fuzzy-Methods - Ausgewählte Kapitel zur Signalverabeitung (Köhler) - Advanced Topics in Signal Processing - Neuronale Netze (Orglmeister, Hoffmann, Vorwerk) - Neural Networks - Ausgewählte Themen zu Elektronik und Signalverarbeitung (Orglmeister) - Selected Topics - Projekt Elektronik (Fischer, Döbler, Timm, Volmer) - Electronics Project Course - 27
PJ PJ IV CO Praktikum Digitale Systeme (Straßnick) - Microcontroller Project Course - Praktikum Signalverarbeitung (Liebsch, Hoffmann) - Signal Processor Project Course - Mixed-Signal-Baugruppen (Westphal) - Mixed Signal Board Design - Forschungscolloquium zur medizinischen Signalverarbeitung (Orglmeister) - Medical Signal Processing Research Seminar - Winter Term 2008/2009 VL UE VL UE VL IV VL PJ PJ PJ SE CO Analog- und Digitalelektronik (Tschirley) - Analog and Digital Electronics - Rechenübungen zu Analog- und Digitalelektronik (Volmer) - Analog and Digital Electronics - Mikroprozessortechnik (Kolossa) - Microprocessor Technology - Rechenübungen zu Mikroprozessortechnik (Pflugradt) - Microprocessor Technology - Signalverarbeitung (Köhler) - Signal Processing - Signalverarbeitung (Vorwerk) - Signal Processing - Neuro-Fuzzy-Methoden (Bothe) - Neuro-Fuzzy-Methods - Projekt Elektronik (Fischer, Döbler, Volmer) - Electronics Project Course - Praktikum Digitale Systeme (Straßnick, Borchers) - Microcontroller Project Course - Praktikum Signalverarbeitung (Liebsch, Hoffmann) - Signal Processor Project Course - Ausgewählte Themen zu Elektronik und Signalverarbeitung (Orglmeister) - Selected Topics - Forschungscolloquium zur medizinischen Signalverarbeitung (Orglmeister) - Medical Signal Processing Research Seminar - 28