Graph-based Semi-Supervised Learning in Acoustic Modeling for Automatic Speech Recognition

Thesis (Ph.D.)--University of Washington, 2016-06

Rotation invariant iris feature extraction using Gaussian Markov random fields with non-separable wavelet

Rotation invariance is important for an iris recognition system since changes of head orientation and binocular vergence may cause eye rotation. The conventional methods of iris recognition cannot achieve true rotation invariance. They only achieve approximate rotation invariance by rotating the feature vector before matching or unwrapping the iris ring at different initial angles. In these methods, the complexity of the method is increased, and when the rotation scale is beyond the certain scope, the error rates of these methods may substantially increase. In order to solve this problem, a new rotation invariant approach for iris feature extraction based on the non-separable wavelet is proposed in this paper. Firstly, a bank of non-separable orthogonal wavelet filters is used to capture characteristics of the iris. Secondly, a method of Markov random fields is used to capture rotation invariant iris feature. Finally, two-class kernel Fisher classifiers are adopted for classification. Experimental results on public iris databases show that the proposed approach has a low error rate and achieves true rotation invariance. © 2010.

Machine learning classification of surgical pathology reports and chunk recognition for information extraction noise reduction

Background and aims: Machine learning techniques for the text mining of cancer-related clinical documents have not been sufficiently explored. Here some techniques are presented for the pre-processing of free-text breast cancer pathology reports, with the aim of facilitating the extraction of information relevant to cancer staging.

Materials and methods: The first technique was implemented using the freely available software RapidMiner to classify the reports according to their general layout: ‘semi-structured’ and ‘unstructured’. The second technique was developed using the open source language engineering framework GATE and aimed at the prediction of chunks of the report text containing information pertaining to the cancer morphology, the tumour size, its hormone receptor status and the number of positive nodes. The classifiers were trained and tested respectively on sets of 635 and 163 manually classified or annotated reports, from the Northern Ireland Cancer Registry.

Results: The best result of 99.4% accuracy – which included only one semi-structured report predicted as unstructured – was produced by the layout classifier with the k nearest algorithm, using the binary term occurrence word vector type with stopword filter and pruning. For chunk recognition, the best results were found using the PAUM algorithm with the same parameters for all cases, except for the prediction of chunks containing cancer morphology. For semi-structured reports the performance ranged from 0.97 to 0.94 and from 0.92 to 0.83 in precision and recall, while for unstructured reports performance ranged from 0.91 to 0.64 and from 0.68 to 0.41 in precision and recall. Poor results were found when the classifier was trained on semi-structured reports but tested on unstructured.

Conclusions: These results show that it is possible and beneficial to predict the layout of reports and that the accuracy of prediction of which segments of a report may contain certain information is sensitive to the report layout and the type of information sought.

Spectral peak resolution and speech recognition in quiet: Normal hearing, hearing impaired and cochlear implant listeners

Spectral peak resolution was investigated in normal hearing (NH), hearing impaired (HI), and cochlear implant (CI) listeners. The task involved discriminating between two rippled noise stimuli in which the frequency positions of the log-spaced peaks and valleys were interchanged. The ripple spacing was varied adaptively from 0.13 to 11.31 ripples/octave, and the minimum ripple spacing at which a reversal in peak and trough positions could be detected was determined as the spectral peak resolution threshold for each listener. Spectral peak resolution was best, on average, in NH listeners, poorest in CI listeners, and intermediate for HI listeners. There was a significant relationship between spectral peak resolution and both vowel and consonant recognition in quiet across the three listener groups. The results indicate that the degree of spectral peak resolution required for accurate vowel and consonant recognition in quiet backgrounds is around 4 ripples/octave, and that spectral peak resolution poorer than around 1–2 ripples/octave may result in highly degraded speech recognition. These results suggest that efforts to improve spectral peak resolution for HI and CI users may lead to improved speech recognition

Speech recognition in noise for cochlear implant listeners: Benefits of residual hearing

The purpose of this study was to explore the potential advantages, both theoretical and applied, of preserving low-frequency acoustic hearing in cochlear implant patients. Several hypotheses are presented that predict that residual low-frequency acoustic hearing along with electric stimulation for high frequencies will provide an advantage over traditional long-electrode cochlear implants for the recognition of speech in competing backgrounds. A simulation experiment in normal-hearing subjects demonstrated a clear advantage for preserving low-frequency residual acoustic hearing for speech recognition in a background of other talkers, but not in steady noise. Three subjects with an implanted "short-electrode" cochlear implant and preserved low-frequency acoustic hearing were also tested on speech recognition in the same competing backgrounds and compared to a larger group of traditional cochlear implant users. Each of the three short-electrode subjects performed better than any of the traditional long-electrode implant subjects for speech recognition in a background of other talkers, but not in steady noise, in general agreement with the simulation studies. When compared to a subgroup of traditional implant users matched according to speech recognition ability in quiet, the short-electrode patients showed a 9-dB advantage in the multitalker background. These experiments provide strong preliminary support for retaining residual low-frequency acoustic hearing in cochlear implant patients. The results are consistent with the idea that better perception of voice pitch, which can aid in separating voices in a background of other talkers, was responsible for this advantage.

Benefits of amplification for speech recognition in background noise

The purpose of the present study was to examine the benefits of providing audible speech to listeners with sensorineural hearing loss when the speech is presented in a background noise. Previous studies have shown that when listeners have a severe hearing loss in the higher frequencies, providing audible speech (in a quiet background) to these higher frequencies usually results in no improvement in speech recognition. In the present experiments, speech was presented in a background of multitalker babble to listeners with various severities of hearing loss. The signal was low-pass filtered at numerous cutoff frequencies and speech recognition was measured as additional high-frequency speech information was provided to the hearing-impaired listeners. It was found in all cases, regardless of hearing loss or frequency range, that providing audible speech resulted in an increase in recognition score. The change in recognition as the cutoff frequency was increased, along with the amount of audible speech information in each condition (articulation index), was used to calculate the "efficiency" of providing audible speech. Efficiencies were positive for all degrees of hearing loss. However, the gains in recognition were small, and the maximum score obtained by an listener was low, due to the noise background. An analysis of error patterns showed that due to the limited speech audibility in a noise background, even severely impaired listeners used additional speech audibility in the high frequencies to improve their perception of the "easier" features of speech including voicing

Automatic segmentation and 3D feature extraction of protein aggregates in Caenorhabditis Elegans

In the last years, it has become increasingly clear that neurodegenerative diseases involve protein aggregation, a process often used as disease progression readout and to develop therapeutic strategies. This work presents an image processing tool to automatic segment, classify and quantify these aggregates and the whole 3D body of the nematode Caenorhabditis Elegans. A total of 150 data set images, containing different slices, were captured with a confocal microscope from animals of distinct genetic conditions. Because of the animals’ transparency, most of the slices pixels appeared dark, hampering their body volume direct reconstruction. Therefore, for each data set, all slices were stacked in one single 2D image in order to determine a volume approximation. The gradient of this image was input to an anisotropic diffusion algorithm that uses the Tukey’s biweight as edge-stopping function. The image histogram median of this outcome was used to dynamically determine a thresholding level, which allows the determination of a smoothed exterior contour of the worm and the medial axis of the worm body from thinning its skeleton. Based on this exterior contour diameter and the medial animal axis, random 3D points were then calculated to produce a volume mesh approximation. The protein aggregations were subsequently segmented based on an iso-value and blended with the resulting volume mesh. The results obtained were consistent with qualitative observations in literature, allowing non-biased, reliable and high throughput protein aggregates quantification. This may lead to a significant improvement on neurodegenerative diseases treatment planning and interventions prevention

Data mining framework for fatty liver disease classification in ultrasound: a hybrid feature extraction paradigm

PURPOSE: Fatty liver disease (FLD) is an increasing prevalent disease that can be reversed if detected early. Ultrasound is the safest and ubiquitous method for identifying FLD. Since expert sonographers are required to accurately interpret the liver ultrasound images, lack of the same will result in interobserver variability. For more objective interpretation, high accuracy, and quick second opinions, computer aided diagnostic (CAD) techniques may be exploited. The purpose of this work is to develop one such CAD technique for accurate classification of normal livers and abnormal livers affected by FLD. METHODS: In this paper, the authors present a CAD technique (called Symtosis) that uses a novel combination of significant features based on the texture, wavelet transform, and higher order spectra of the liver ultrasound images in various supervised learning-based classifiers in order to determine parameters that classify normal and FLD-affected abnormal livers. RESULTS: On evaluating the proposed technique on a database of 58 abnormal and 42 normal liver ultrasound images, the authors were able to achieve a high classification accuracy of 93.3% using the decision tree classifier. CONCLUSIONS: This high accuracy added to the completely automated classification procedure makes the authors' proposed technique highly suitable for clinical deployment and usage.

XDense: A Dense Grid Sensor Network for Distributed Feature Extraction

XXXIII Simpósio Brasileiro de Redes de Computadores e Sistemas Distribuídos (SBRC 2015). 15 to 19, May, 2015, III Workshop de Comunicação em Sistemas Embarcados Críticos. Vitória, Brasil.

XDense: A Dense Grid Sensor Network for Distributed Feature Extraction

XXXIII Simpósio Brasileiro de Redes de Computadores e Sistemas Distribuídos (SBRC 2015), III Workshop de Comunicação em Sistemas Embarcados Críticos. Vitória, Brasil.

Feature Extraction in Densely Sensed Environments: Extensions to Multiple Broadcast Domains

The vision of the Internet of Things (IoT) includes large and dense deployment of interconnected smart sensing and monitoring devices. This vast deployment necessitates collection and processing of large volume of measurement data. However, collecting all the measured data from individual devices on such a scale may be impractical and time consuming. Moreover, processing these measurements requires complex algorithms to extract useful information. Thus, it becomes imperative to devise distributed information processing mechanisms that identify application-specific features in a timely manner and with a low overhead. In this article, we present a feature extraction mechanism for dense networks that takes advantage of dominance-based medium access control (MAC) protocols to (i) efficiently obtain global extrema of the sensed quantities, (ii) extract local extrema, and (iii) detect the boundaries of events, by using simple transforms that nodes employ on their local data. We extend our results for a large dense network with multiple broadcast domains (MBD). We discuss and compare two approaches for addressing the challenges with MBD and we show through extensive evaluations that our proposed distributed MBD approach is fast and efficient at retrieving the most valuable measurements, independent of the number sensor nodes in the network.

Estimation of nonparametric probability density functions with applications to automatic speech recognition

Magdeburg, Univ., Fak. für Elektrotechnik und Informationstechnik, Diss., 2007

Investigation of ICA algorithms for feature extraction of EEG signals in discrimination of Alzheimer disease

In this paper we present a quantitative comparisons of different independent component analysis (ICA) algorithms in order to investigate their potential use in preprocessing (such as noise reduction and feature extraction) the electroencephalogram (EEG) data for early detection of Alzhemier disease (AD) or discrimination between AD (or mild cognitive impairment, MCI) and age-match control subjects.