<p>With timeliness and efficiency being critical in the aviation maintenance industry, the need has been growing for smart technological solutions that help in optimizing and streamlining the different underlying tasks. One such task is the technical documentation of the performed maintenance operations. Instead of paper-based documentation, voice tools that transcribe spoken logbook entries allow technicians to document their work right away in a hands-free and time efficient manner. However, an accurate automatic speech recognition (ASR) model requires large training corpora, which are lacking in the domain of aviation maintenance. In addition, ASR models which are trained on huge corpora in standard English perform poorly in such a technical domain with non-standard terminology. Hence, this thesis investigates the extent to which fine-tuning an ASR model, pre-trained on standard English corpora, on limited in-domain data improves its recognition performance in the technical domain of aviation maintenance. The thesis presents a case study on one such pre-trained ASR model, wav2vec 2.0. Results show that fine-tuning the model on a limited anonymized dataset of maintenance logbook entries brings about a significant reduction in its error rates when tested on not only an anonymized in-domain dataset, but also a non-anonymized one. This suggests that any available aviation maintenance logbooks, even if anonymized for privacy, can be used to fine-tune general-purpose ASR models and enhance their in-domain performance. Lastly, an analysis on the influence of voice characteristics on model performance stresses the need for balanced datasets representative of the population of aviation maintenance technicians.</p>

No description available.

Despite the documented need for reliable speech audiometry materials for measures such as speech recognition threshold and word recognition score, such recorded materials are not available in the Vietnamese language. The purpose of this study was to develop, digitally record, evaluate, and psychometrically equate a set of Vietnamese bisyllabic and monosyllabic word lists for use in the measurement of speech recognition and word recognition ability, respectively. To create the speech recognition threshold materials, common Vietnamese bisyllabic words were digitally recorded by male and female talkers of Vietnamese and presented for evaluation to 20 native speakers of Vietnamese with normal hearing. Based on listener response, a set of 48 bisyllabic words with relatively steep psychometric function slopes were selected and digitally adjusted to ensure equivalency for psychometric function slope and to equate threshold to the mean pure-tone average for the test participants. To create the word recognition materials, 250 words were digitally recorded by one male and one female talker of Vietnamese and presented to the listeners for evaluation. Based on listener response, 200 words were selected and divided into 4 lists of 50 monosyllabic words and 8 half-lists of 25 monosyllabic words. The lists were digitally adjusted to ensure intensity threshold equivalency. The resulting mean psychometric function slopes at 50% for the speech recognition threshold materials is 11.3%/dB for the male talker and 10.2%/dB for the female talker. Analysis of the word recognition materials indicates no significant difference between the lists or half-lists. The mean psychometric function slope at 50% for the monosyllabic lists and half-lists is 5.1%/dB for the male recordings and 5.2%/dB for the female recordings. The results of the current study are comparable to those found in other languages. Digital recordings of the bisyllabic and monosyllabic word lists are available on compact disc.

Proteins and nanoparticles (NPs) provide a promising platform for supramolecular interaction. We are currently exploring both fundamental and applied aspects of this interaction. On the fundamental side, we have fabricated a series of water-soluble anionic and cationic NPs to interact with cationic and anionic proteins respectively. A Varity of studies such as, activity assay, fluorescence titration, isothermal titration calorimetry etc. were carried out to quantify the binding properties of these functional NPs with those proteins. Those studies reveal the prospect of tuning the affinity between the nanoparticles and proteins by the surface modification. On the application side, we have used this protein-nanoparticle interaction in protein refolding where we successfully refolded the thermally denatured proteins toward its native structure. We have also applied this particle-protein recognition to create a biocompatible protein sensor using a protein-NP conjugate. Green fluorescent protein and a series of cationic NPs were used for a protein sensor for the identification of protein analytes through displacement process. We have extended this application even in sensing the proteins in human serum.

No description available.

Yes / Accurate identification of siblings through face recognition is a challenging task. This is predominantly because of the high degree of similarities among the faces of siblings. In this study, we investigate the use of state-of-the-art deep learning face recognition models to evaluate their capacity for discrimination between sibling faces using various similarity indices. The specific models examined for this purpose are FaceNet, VGGFace, VGG16, and VGG19. For each pair of images provided, the embeddings have been calculated using the chosen deep learning model. Five standard similarity measures, namely, cosine similarity, Euclidean distance, structured similarity, Manhattan distance, and Minkowski distance, are used to classify images looking for their identity on the threshold defined for each of the similarity measures. The accuracy, precision, and misclassification rate of each model are calculated using standard confusion matrices. Four different experimental datasets for full-frontal-face, eyes, nose, and forehead of sibling pairs are constructed using publicly available HQf subset of the SiblingDB database. The experimental results show that the accuracy of the chosen deep learning models to distinguish siblings based on the full-frontal-face and cropped face areas vary based on the face area compared. It is observed that VGGFace is best while comparing the full-frontal-face and eyes—the accuracy of classification being with more than 95% in this case. However, its accuracy degrades significantly when the noses are compared, while FaceNet provides the best result for classification based on the nose. Similarly, VGG16 and VGG19 are not the best models for classification using the eyes, but these models provide favorable results when foreheads are compared.

This thesis describes the implementation of a speaker dependent connected-digit recognizer using continuous Hidden Markov Modeling (HMM). The speech recognition system was implemented using MATLAB and on the ADSP-2181, a digital signal processor manufactured by Analog Devices. Linear predictive coding (LPC) analysis was first performed on a speech signal to model the characteristics of the vocal tract filter. A 7 state continuous HMM with 4 mixture density components was used to model each digit. The Viterbi reestimation method was primarily used in the training phase to obtain the parameters of the HMM. Viterbi decoding was used for the recognition phase. The system was first implemented as an isolated word recognizer. Recognition rates exceeding 99% were obtained on both the MATLAB and the ADSP-2181 implementations. For continuous word recognition, several algorithms were implemented and compared. Using MATLAB, recognition rates exceeding 90% were obtained. In addition, the algorithms were implemented on the ADSP-2181 yielding recognition rates comparable to the MATLAB implementation. / Master of Science

Many real world datasets are characterized by having a long tailed distribution, with several samples for some classes and only a few samples for other classes. While many Deep Learning based solutions exist for object recognition when hundreds of samples are available, there are not many solutions for the case when there are only a few samples available per class. Recognition in the regime where the number of training samples available for each class are low, ranging from 1 to couple of tens of examples is called Lowshot Recognition. In this work, we attempt to solve this problem. Our framework is similar to [1]. We use a related dataset with sufficient number (a couple of hundred) of samples per class to learn representations using a Convolutional Neural Network (CNN). This CNN is used to extract features of the lowshot samples and learn a classifier . During representation learning, we enforce the learnt representations to obey certain property by using a custom loss function. We believe that when the lowshot sample obey this property the classification step becomes easier. We show that the proposed solution performs better than the softmax classifier by a good margin. / Master of Science / Deep learning, a branch of Artificial Intelligence(AI) is revolutionizing the way computers can learn and perform artificial intelligence tasks. The power of Deep Learning comes from being able to model very complex functions using huge amounts of data. For this reason, deep learning is criticized as being data hungry. Although AI systems are able to beat humans in many tasks, unlike humans, they still lack the ability to learn from less data. In this work, we address the problem of teaching AI systems with only a few examples, formally called the “low-shot learning”. We focus on low-shot visual recognition where the AI systems are taught to recognize different objects from images using very few examples. Solving the low-shot recognition problem will enable us to apply AI based methods to many real world tasks. Particularly in the cases where we cannot afford to collect huge number of images because it is either costly or it is impossible. We propose a novel technique to solve this problem. We show that our solution performs better at low-shot recognition than the regular image classification solution, the softmax classifier.

A key enabler of pervasive computing is the ability to drive service delivery through the analysis of situations: Semantically meaningful classifications of system state, identified through analysing the readings from sensors attached to the everyday objects that people interact with. Situation recognition is a mature area of research, with techniques primarily falling into two categories. Knowledge-based techniques use inference rules crafted by experts; however often they compensate poorly for sensing peculiarities. Learning-based approaches excel at extracting patterns from noisy training data, however their lack of transparency can make it difficult to diagnose errors. In this thesis we propose a novel hybrid approach to situation recognition that combines both techniques. This offers improvements over each used individually, through not sacrificing the intelligibility of the decision processes that the use of machine learning alone often implies, and through providing better recognition accuracy through robustness to noise typically unattainable when developers use knowledge-based techniques in isolation. We present an ontology model and reasoning framework that supports the uniform modelling of pervasive environments, and infers additional knowledge from that which is specified, in a principled way. We use this as a basis from which to learn situation recognition models that exhibit comparable performance with more complex machine learning techniques, while retaining intelligibility. Finally, we extend the approach to construct ensemble classifiers with either improved recognition accuracy, intelligibility or both. To validate our approach, we apply the techniques to real-world data sets collected in smart-office and smart-home environments. We analyse the situation recognition performance and intelligibility of the decision processes, and compare the results to standard machine learning techniques and results published in the literature.

Condition based monitoring is widely used for the determination of the health of machines. The Nottingham Trent University Computing Department has developed a new system, the Incremental Motion Encoder (!ME), which is based on the time interpolation of the digital signals produced by an optical encoder. Experiments have shown that the !ME can be used as a condition based maintenance sensor as it is possible to detect rolling element defects, an unbalanced shaft and oil contamination of a bearing. The system uses a geometrically configured optical device to scan a precision encoder disc and Digital Signal Processing technology is used to interpret the signals. Previous work has demonstrated the qualitative usefulness of the 1ME. However, further work was needed to assess the accuracy of the measurements, to analyse the principles of the 1ME, to validate the performance of the existing device and to develop methods for error definition and error compensation. Testing and experimentation on the existing experimental system have been carried out by the Candidate and an understanding gained of the device. The sources of error of the 1ME have been identified, which had not been quantified previously. Measuring and compensating for the three main sources of error, read head position, eccentricity of the encoder disc and encoder abnormalities are the three major tasks of the project. Modifications to the experimental rig have been developed in order to allow these tasks to be addressed. The Candidate has developed three different types of techniques to measure the position error of the read heads. A pattern recognition method was developed and is successful for 1ME systems that use an encoder disc with significant grating line errors. A second method using Fast Fourier Transform (FFT) has been developed to exploit the fact that the difference in the phase angles, obtained using a FFT, gives the angle between the read head positions. The new experimental system is now able to obtain the angular position of the read heads by using the index grating line. The third method relies on the presence of the index grating line on the encoder disc which may not be present in all systems. Eccentricity of disc centre relative to the centre of rotation affects the correct calculation of the angular position of the encoder disc. Algorithms have been developed by the Candidate in order to compensate for this type of error. Experimental results have shown that angular position error can be corrected successfully. The Candidate has developed methods for detection of small abnormalities of the encoder disc by using a multiple averaging technique. Computational algorithms have been developed to correct the encoder disc abnormalities by using individual information from each read head, promising results have been obtained from the experimental 1ME. An 1ME device can be tailored to fulfil the desired requirements of resolution, bandwidth and accuracy. A self calibration instrument can be developed by using the previously mentioned techniques in order to self calibrate and increase the accuracy and reliability of an IME's results.