220 research outputs found
Acoustic Event Detection and Classification in Smart-Room Environment: Evaluation of CHIL Project Systems
No full textThe identification of acoustic events that are produced in ameeting room environment may help to detect and describethe human and social activity that takes place in the room. Inthe framework of the CHIL project, three different sites havedeveloped and tested several preliminary systems foracoustic event classification (AEC) and acoustic eventdetection (AED). A primary AED evaluation task with thetesting portions of the isolated sound databases and theseminar recordings produced in CHIL was designed, and itwas carried out in February 2006. Additionally, a secondaryAEC evaluation task was carried out using only the isolatedsound databases. In this paper, a short description of thesystems is presented, and the evaluation setup and results,for both AED and AEC, are reported and discussed
UPM-UC3M system for music and speech segmentation
No full textThis paper describes the UPM-UC3M system for the Albayzín evaluation 2010 on Audio Segmentation. This evaluation task consists of segmenting a broadcast news audio document into clean speech, music, speech with noise in background and speech with music in background. The UPM-UC3M system is based on Hidden Markov Models (HMMs), including a 3-state HMM for every acoustic class. The number of states and the number of Gaussian per state have been tuned for this evaluation. The main analysis during system development has been focused on feature selection. Also, two different architectures have been tested: the first one corresponds to an one-step system whereas the second one is a hierarchical system in which different features have been used for segmenting the different audio classes. For both systems, we have considered long term statistics of MFCC (Mel Frequency Ceptral Coefficients), spectral entropy and CHROMA coefficients. For the best configuration of the one-step system, we have obtained a 25.3% average error rate and 18.7% diarization error (using the NIST tool) and a 23.9% average error rate and 17.9% diarization error for the hierarchical one
CLEAR Evaluation of Acoustic Event Detection and Classification Systems
No full textAbstract. In this paper, we present the results of the Acoustic Event Detection (AED) and Classification (AEC) evaluations carried out in February 2006 by the three participant partners from the CHIL project. The primary evaluation task was AED of the testing portions of the isolated sound databases and seminar recordings produced in CHIL. Additionally, a secondary AEC evaluation task was designed using only the isolated sound databases. The set of meetingroom acoustic event classes and the metrics were agreed by the three partners and ELDA was in charge of the scoring task. In this paper, the various systems for the tasks of AED and AEC and their results are presented.
Form und Wachstum der InAs Quantenpunkte auf hoch-indizierten GaAs(113)A, B und GaAs(2 5 11)A, B Substraten
No full textDie vorliegende Dissertation ist der Untersuchung der selbst-organisierten InAs Quantenpunkte (QD) auf vier Substraten, GaAs(113)A, B und GaAs(2 5 11)A, B, gewidmet. Die Proben wurden mit Molekularstrahlepitaxie präpariert, und mit in situ Rastertunnelmikroskopie (STM), Elektronenbeugung und Photolumineszenz (PL) untersucht. Der Stranski-Krastanow Wachstummodus tritt auf allen untersuchten Oberflächen auf. Die Symmetrie der QDs ergibt sich durch die Symmetrie des Substrates und beweist damit das epitaktische Wachstum. Die QD Ensembles auf B Flächen zeigen schmalere Grössenverteilungen (GV) und grössere Inseldichten als die QDs auf B Flächen. InAs QDs auf GaAs(113)A werden durch {110}, (111)A und {2 5 11}A Facetten und einen gerundeten (001) Bereich begrenzt. Diese Form wird später durch eine Verlängerung unter Reduzierung der (111)A Facette, verändert. Dabei tritt eine Abflachung des gerundeten Bereichs durch {113}B-Facetten auf. Die wellige Morphologie von GaAs(113)A und von der InAs-Benetzungsschicht (WL) erklärt die breite GV und niedrige QD Inseldichte. InAs QDs auf GaAs(113)B wachsen mit einem steileren zentralen Teil, der auf einem flachen Sockel sitzt. Die Form des steilen Teils wird durch {110} und (111)B Facetten und einen (001)-vizinalen Bereich gegeben. Die {135}B- und (112)B-Facetten rahmen den flachen Sockel ein. Die flache Morphologie von GaAs(113)B bleibt in der InAs WL erhalten. Deshalb wachsen die QDs überall gleichzeitig und erbringen eine schmalle GV und hohe Inseldichte. GaAs(2 5 11)A wächst unter Bildung von (011) Stufenbündeln. Nach Aufbringen von InAs zeigen STM Bilder von vizinalen GaAs(2 5 11)A Oberflächen 3D Inseln, die genau auf den (011)-Stufenbündeln sitzen. Ihre sehr breite GV ist typisch für die inkoherente Inseln und wird durch eine niedrige PL Intensität bestätigt. Auf dem nominalen Substrat zeigt die GV einen schmalen und einen breiten Anteil. Der letztere stammt von den Inseln auf den Stufenbündeln. Der schmale Anteil wird den koherenten QDs zugeschrieben, die nahezu die gleiche Form wie die QDs auf GaAs(113)A aufweisen. Jedoch zeigt die PL ähnliche Spektren für vizinale und nominale InAs/GaAs(2 5 11)A Systeme. Die GaAs(2 5 11)B-Oberfläche, deren atomare Struktur zum ersten Mal in dieser Arbeit bestimmt wurde, ergibt ein homogenes InAs QD Ensemble mit einer hohen Inseldichte. Die Form der QDs ist dieselbe wie für die QDs auf der GaAs(113)B Oberfläche bis auf die fehlende Symmetrie. Die PL Spektren von InAs QDs auf GaAs(2 5 11)B und GaAs(001) weisen ähnliche Intensitäten aber kleinere Werte für Emissionsenergie und Linienbreite für GaAs(2 5 11)B auf. Eine Rotverschiebung der Emission wurde durch geänderte Wachstumsparameter erzielt.The present thesis has been devoted to the investigation of self-organised InAs quantum dots (QD) on four high-index substrates: GaAs(113)A, B and GaAs(2 5 11)A, B. The samples were prepared by molecular beam epitaxy, and characterised by in situ scanning tunneling microscopy (STM), electron diffraction and photoluminescence (PL). The Stranski-Krastanow (SK) growth mode occurs on all investigated surfaces. The symmetry of the QDs derives from the bulk-truncated substrate that proves epitaxial growth during and after the SK transition. The QD ensembles on the B faces exhibit a narrower size distribution (SD) and larger density than those on respective A faces. InAs QDs on GaAs(113)A are given by {110}, (111)A and {2 5 11}A bounding facets and a rounded region due to a stacking of vicinal (001) surfaces. Later in the growth this shape alters by an elongation with a size reduction of the (111)A facet, induced by the flattening of the rounded region with {113}B facets. An undulating morphology of the bare GaAs(113)A and of the wetting layer (WL) accounts for a broad QD SD and low density. InAs QDs on GaAs(113)B evolve with a central steep part sitting on a flat base. The shape of the central part is given by {110} and (111)B bounding facets, and a (001) rounded region. High-index {135}B and (112)B facets are derived for the flat base. The flat morphology of bare GaAs(113)B is retained in the InAs WL. The QDs grow simultaneously with an equal rate everywhere that results in a narrow SD and high density. A remarkable feature of GaAs(2 5 11)A is the formation of GaAs(011) step bunches found on nominal and vicinal substrates. STM images from vicinal GaAs(2 5 11)A reveal 3D InAs islands, sitting on the step bunches with a very broad SD, that is believed to be characteristic for incoherent islands. It is confirmed by a low PL emission intensity. On nominal GaAs(2 5 11)A, the islands grow with narrow and broad SD. The latter stem from the islands appearing on the (011) step bunches. The narrow SD is ascribed to the coherent QDs which exhibit the same shape as those on GaAs(113)A except for a missing mirror symmetry. However, the PL shows similar spectra for vicinal and nominal GaAs(2 5 11)A. The GaAs(2 5 11)B surface, which structure has been determined in this thesis for the first time, yields a uniform InAs QD ensemble with a high density. The shape of the QDs is mainly the same as that on GaAs(113)B, except for the missing mirror symmetry. The QD PL peak exhibits similar intensity as reference InAs QDs on GaAs(001), but with smaller energy and linewidth, indicating a smaller sized and more uniform QD ensemble. A red shift of the emission energy has been achieved by using a modified preparation
Deep learning for EEG seizure detection in preterm infants.
No full textEEG is the gold standard for seizure detection in the newborn infant, but EEG interpretation in the preterm group is particularly challenging; trained experts are scarce and the task of interpreting EEG in real-time is arduous. Preterm infants are reported to have a higher incidence of seizures compared to term infants. Preterm EEG morphology differs from that of term infants, which implies that seizure detection algorithms trained on term EEG may not be appropriate. The task of developing preterm specific algorithms becomes extra-challenging given the limited amount of annotated preterm EEG data available. This paper explores novel deep learning (DL) architectures for the task of neonatal seizure detection in preterm infants. The study tests and compares several approaches to address the problem: training on data from full-term infants; training on data from preterm infants; training on age-specific preterm data and transfer learning. The system performance is assessed on a large database of continuous EEG recordings of 575h in duration. It is shown that the accuracy of a validated term-trained EEG seizure detection algorithm, based on a support vector machine classifier, when tested on preterm infants falls well short of the performance achieved for full-term infants. An AUC of 88.3% was obtained when tested on preterm EEG as compared to 96.6% obtained when tested on term EEG. When re-trained on preterm EEG, the performance marginally increases to 89.7%. An alternative DL approach shows a more stable trend when tested on the preterm cohort, starting with an AUC of 93.3% for the term-trained algorithm and reaching 95.0% by transfer learning from the term model using available preterm data. The proposed DL approach avoids time-consuming explicit feature engineering and leverages the existence of the term seizure detection model, resulting in accurate predictions with a minimum amount of annotated preterm data
Neonatal seizure detection: a deep learning approach
No full textThe detection of neonatal seizures is an important step in identifying neurological dysfunction in newborn infants. Research indicates that seizures in infants are underreported; specialist EEG monitoring equipment and neonatal neurophysiological expertise are required to reliably detect neonatal seizures due to the predominance of sub-clinical seizure events. Often, the required level of clinical expertise is not available around-the-clock in the NICU; medical experts have cited the need for decision support algorithms to assist staff during these times.
In this thesis novel automated neonatal seizure detection algorithms are proposed. Here deep learning end-to-end optimised algorithms, which detect seizures from raw multi-channel EEG, are presented. The deep learning approach differs from previous rule-based and machine learning systems, which rely on hand-engineered features. An appropriate input EEG representation is selected through empirical analysis; 2-dimensional time-frequency representations of EEG are compared to 1-dimensional temporal representations in a series of experiments. Deep learning algorithms prove capable of extracting discriminative intermediate hierarchical representations from raw EEG.
A deep learning algorithm for detecting seizures in term neonates is proposed. The designed algorithm utilises only convolutional layers to process multi-channel temporal EEG and is designed to exploit the large quantity of weakly labelled data available in the training stage. The effect of varying architectural parameters is thoroughly studied, and the designed architecture compares favourably in terms of performance, inference run time and number of parameters when compared with baseline systems. The developed system outperforms state-of-the-art machine learning algorithms when tested on a large database of continuous EEG recordings (duration 834h) and when further validated on a held-out publicly available dataset (duration 112h), achieving AUC results of 98.5% and 95.6% respectively.
The challenges associated with detecting seizures in term EEG are exacerbated in preterm EEG due to the large variations in EEG morphology depending on gestational age and the limited amount of annotated preterm EEG available. A deep learning algorithm development framework is proposed where classifiers are trained to detect seizures for specific gestational age ranges. The developed algorithms leverage the existence of robustly trained term EEG models through transfer learning and classifier ensembling; this results in accurate seizure detection despite the scarcity of labelled training data. This preterm seizure detection framework represents the first time an algorithm of this kind has been developed; it achieves an AUC of 95.4% on a held-out test dataset (duration 575h), and detects approximately 50% of all seizure events at a false detection rate of one false alarm every four hours.
In this work, algorithms are investigated through a series of visualisation techniques. This analysis gives an understanding of the EEG patterns which contribute to algorithm decisions and highlights the differences between term and preterm EEG. The developed algorithm framework is also utilised as part of a mobile brain monitoring device where the light-weight nature of the designed network and the simplicity of the inference computations are exploited through AI-on-the-edge decision support. The ability to decipher deep learning algorithms and to integrate algorithms into existing brain monitoring systems are necessary steps for the translation of the work in this thesis into clinical domain
Accurate wearable heart rate monitoring during physical exercises using PPG
No full textObjective: The challenging task of heart rate (HR) estimation from the photoplethysmographic (PPG) signal, during intensive physical exercises is tackled in this paper. Methods: The study presents a detailed analysis of a novel algorithm (WFPV) that exploits a Wiener filter to attenuate the motion artifacts, a phase vocoder to refine the HR estimate and user-adaptive postprocessing to track the subject physiology. Additionally, an offline version of the HR estimation algorithm that uses Viterbi decoding is designed for scenarios that do not require online HR monitoring (WFPV+VD). The performance of the HR estimation systems is rigorously compared with existing algorithms on the publically available database of 23 PPG recordings. Results: On the whole dataset of 23 PPG recordings, the algorithms result in average absolute errors of 1.97 and 1.37 BPM in the online and offline modes, respectively. On the test dataset of 10 PPG recordings which were most corrupted with motion artifacts, WFPV has an error of 2.95 BPM on its own and 2.32 BPM in an ensemble with 2 existing algorithms. Conclusion: The error rate is significantly reduced when compared with the state-of-the art PPG-based HR estimation methods. Significance: The proposed system is shown to be accurate in the presence of strong motion artifacts and in contrast to existing alternatives has very few free parameters to tune. The algorithm has a low computational cost and can be used for fitness tracking and health monitoring in wearable devices. The Matlab implementation of the algorithm is provided online
PPG-based heart rate estimation using Wiener filter, phase vocoder and Viterbi decoding
No full textAccurate heart rate (HR) estimation from the photoplethysmographic (PPG) signal during intensive physical exercises is tackled in this paper. Wiener filters are designed to attenuate the influence of motion artifacts. The phase vocoder is used to improve the initial Discrete Fourier transform (DFT) based frequency estimation. Additionally, Viterbi decoding is used as a novel post-processing step to find the path through time-frequency state-space plane. The system performance is assessed on a publically available dataset of 23 PPG recordings. The resulting algorithm is designed for scenarios that do not require online HR monitoring (swimming, offline fitness statistics). The resultant system with an error rate of 1.31 beats per minute outperforms all other systems reported to-date in literature and in contrast to existing alternatives requires no parameter to tune at the post-processing stage and operates at a much lower computational cost. The Matlab implementation is provided online
- …
