Thesis (Ph.D.)--McMaster University, 2004. / Advisor: Zhi-Quan (Tom) Luo. Includes bibliographical references (leaves 90-97). Also available via World Wide Web.

由於在盲語音分離的應用，線性準平穩源訊號混合的盲識別獲得了巨大的研究興趣。在這個問題上，我們利用準穩態源訊號的時變特性來識別未知的混合系統系數。傳統的方法有二：i)基於張量分解的平行因子分析(PARAFAC);ii)基於對多個矩陣的聯合對角化的聯合對角化算法(JD)。一般來說，PARAFAC和JD 都採用了源聯合的提取方法；即是說，對應所有訊號源的系統係數在升法上是用時進行識別的。 / 在這篇論文中，我利用Khati-Rao(KR)子空間來設計一種新的盲識別算法。在我設計的算法中提出一種與傳統的方法不同的提法。在我設計的算法中，盲識別問題被分解成數個結構上相對簡單的子問題，分別對應不同的源。在超定混合模型，我們提出了一個專門的交替投影算法(AP)。由此產生的算法，不但能從經驗發現是非常有競爭力的，而且更有理論上的利落收斂保證。另外，作為一個有趣的延伸，該算法可循一個簡單的方式應用於欠混合模型。對於欠定混合模型，我們提出啟發式的秩最小化算法從而提高算法的速度。 / Blind identification of linear instantaneous mixtures of quasi-stationary sources (BI-QSS) has received great research interest over the past few decades, motivated by its application in blind speech separation. In this problem, we identify the unknown mixing system coefcients by exploiting the time-varying characteristics of quasi-stationary sources. Traditional BI-QSS methods fall into two main categories: i) Parallel Factor Analysis (PARAFAC), which is based on tensor decomposition; ii) Joint Diagonalization (JD), which is based on approximate joint diagonalization of multiple matrices. In both PARAFAC and JD, the joint-source formulation is used in general; i.e., the algorithms are designed to identify the whole mixing system simultaneously. / In this thesis, I devise a novel blind identification framework using a Khatri-Rao (KR) subspace formulation. The proposed formulation is different from the traditional formulations in that it decomposes the blind identication problem into a number of per-source, structurally less complex subproblems. For the over determined mixing models, a specialized alternating projections algorithm is proposed for the KR subspace for¬mulation. The resulting algorithm is not only empirically found to be very competitive, but also has a theoretically neat convergence guarantee. Even better, the proposed algorithm can be applied to the underdetermined mixing models in a straightforward manner. Rank minimization heuristics are proposed to speed up the algorithm for the underdetermined mixing model. The advantages on employing the rank minimization heuristics are demonstrated by simulations. / Detailed summary in vernacular field only. / Detailed summary in vernacular field only. / Lee, Ka Kit. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2012. / Includes bibliographical references (leaves 72-76). / Abstracts also in Chinese. / Abstract --- p.i / Acknowledgement --- p.ii / Chapter 1 --- Introduction --- p.1 / Chapter 2 --- Settings of Quasi-Stationary Signals based Blind Identification --- p.4 / Chapter 2.1 --- Signal Model --- p.4 / Chapter 2.2 --- Assumptions --- p.5 / Chapter 2.3 --- Local Covariance Model --- p.7 / Chapter 2.4 --- Noise Covariance Removal --- p.8 / Chapter 2.5 --- Prewhitening --- p.9 / Chapter 2.6 --- Summary --- p.10 / Chapter 3 --- Review on Some Existing BI-QSS Algorithms --- p.11 / Chapter 3.1 --- Joint Diagonalization --- p.11 / Chapter 3.1.1 --- Fast Frobenius Diagonalization [4] --- p.12 / Chapter 3.1.2 --- Pham’s JD [5, 6] --- p.14 / Chapter 3.2 --- Parallel Factor Analysis --- p.16 / Chapter 3.2.1 --- Tensor Decomposition [37] --- p.17 / Chapter 3.2.2 --- Alternating-Columns Diagonal-Centers [12] --- p.21 / Chapter 3.2.3 --- Trilinear Alternating Least-Squares [10, 11] --- p.23 / Chapter 3.3 --- Summary --- p.25 / Chapter 4 --- Proposed Algorithms --- p.26 / Chapter 4.1 --- KR Subspace Criterion --- p.27 / Chapter 4.2 --- Blind Identification using Alternating Projections --- p.29 / Chapter 4.2.1 --- All-Columns Identification --- p.31 / Chapter 4.3 --- Overdetermined Mixing Models (N > K): Prewhitened Alternating Projection Algorithm (PAPA) --- p.32 / Chapter 4.4 --- Underdetermined Mixing Models (N <K) --- p.34 / Chapter 4.4.1 --- Rank Minimization Heuristic --- p.34 / Chapter 4.4.2 --- Alternating Projections Algorithm with Huber Function Regularization --- p.37 / Chapter 4.5 --- Robust KR Subspace Extraction --- p.40 / Chapter 4.6 --- Summary --- p.44 / Chapter 5 --- Simulation Results --- p.47 / Chapter 5.1 --- General Settings --- p.47 / Chapter 5.2 --- Overdetermined Mixing Models --- p.49 / Chapter 5.2.1 --- Simulation 1 - Performance w.r.t. SNR --- p.49 / Chapter 5.2.2 --- Simulation 2 - Performance w.r.t. the Number of Available Frames M --- p.49 / Chapter 5.2.3 --- Simulation 3 - Performance w.r.t. the Number of Sources K --- p.50 / Chapter 5.3 --- Underdetermined Mixing Models --- p.52 / Chapter 5.3.1 --- Simulation 1 - Success Rate of KR Huber --- p.53 / Chapter 5.3.2 --- Simulation 2 - Performance w.r.t. SNR --- p.54 / Chapter 5.3.3 --- Simulation 3 - Performance w.r.t. M --- p.54 / Chapter 5.3.4 --- Simulation 4 - Performance w.r.t. N --- p.56 / Chapter 5.4 --- Summary --- p.56 / Chapter 6 --- Conclusion and Future Works --- p.58 / Chapter A --- Convolutive Mixing Model --- p.60 / Chapter B --- Proofs --- p.63 / Chapter B.1 --- Proof of Theorem 4.1 --- p.63 / Chapter B.2 --- Proof of Theorem 4.2 --- p.65 / Chapter B.3 --- Proof of Observation 4.1 --- p.65 / Chapter B.4 --- Proof of Proposition 4.1 --- p.66 / Chapter C --- Singular Value Thresholding --- p.67 / Chapter D --- Categories of Speech Sounds and Their Impact on SOSs-based BI-QSS Algorithms --- p.69 / Chapter D.1 --- Vowels --- p.69 / Chapter D.2 --- Consonants --- p.69 / Chapter D.1 --- Silent Pauses --- p.70 / Bibliography --- p.72

Although distinguishing different sounds in noisy environment is a relative easy task for human, source separation has long been extremely difficult in audio signal processing. The problem is challenging for three reasons: the large variety of sound type, the abundant mixing conditions and the unclear mechanism to distinguish sources, especially for similar sounds. In recent years, the neural network based methods achieved impressive successes in various problems, including the speech enhancement, where the task is to separate the clean speech out of the noise mixture. However, the current deep learning based source separator does not perform well on real recorded noisy speech, and more importantly, is not applicable in a more general source separation scenario such as overlapped speech. In this thesis, we firstly propose extensions for the current mask learning network, for the problem of speech enhancement, to fix the scale mismatch problem which is usually occurred in real recording audio. We solve this problem by combining two additional restoration layers in the existing mask learning network. We also proposed a residual learning architecture for the speech enhancement, further improving the network generalization under different recording conditions. We evaluate the proposed speech enhancement models on CHiME 3 data. Without retraining the acoustic model, the best bi-direction LSTM with residue connections yields 25.13% relative WER reduction on real data and 34.03% WER on simulated data. Then we propose a novel neural network based model called “deep clustering” for more general source separation tasks. We train a deep network to assign contrastive embedding vectors to each time-frequency region of the spectrogram in order to implicitly predict the segmentation labels of the target spectrogram from the input mixtures. This yields a deep network-based analogue to spectral clustering, in that the embeddings form a low-rank pairwise affinity matrix that approximates the ideal affinity matrix, while enabling much faster performance. At test time, the clustering step “decodes” the segmentation implicit in the embeddings by optimizing K-means with respect to the unknown assignments. Experiments on single channel mixtures from multiple speakers show that a speaker-independent model trained on two-speaker and three speakers mixtures can improve signal quality for mixtures of held-out speakers by an average over 10dB. We then propose an extension for deep clustering named “deep attractor” network that allows the system to perform efficient end-to-end training. In the proposed model, attractor points for each source are firstly created the acoustic signals which pull together the time-frequency bins corresponding to each source by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. We showed that this frame work can achieve even better results. Lastly, we introduce two applications of the proposed models, in singing voice separation and the smart hearing aid device. For the former, a multi-task architecture is proposed, which combines the deep clustering and the classification based network. And a new state of the art separation result was achieved, where the signal to noise ratio was improved by 11.1dB on music and 7.9dB on singing voice. In the application of smart hearing aid device, we combine the neural decoding with the separation network. The system firstly decodes the user’s attention, which is further used to guide the separator for the targeting source. Both objective study and subjective study show the proposed system can accurately decode the attention and significantly improve the user experience.

M. Tech. Electrical Engineering. / Aims to develop directional light emitters, wave guiding of light along CMOS structures and coupling of light into secondary elements in CMOS integrated circuitry that can be modulated at a high speed up to (1Gb/sec PLUS). This could enable high density in connection in integrated circuit packages, helping to implement on-chip optical processing and to generate optical highway communication channels from and to the chip tp produce photonic signal processing using standard CMOS and bipolar-base integrated circuitry.

Thesis (Ph. D.)--Rochester Institute of Technology, 2007. / Typescript. Includes bibliographical references (leaves 123-129).

We propose a Bayesian monaural source separation system to extract each individual tone from mixture signals of piano music performance. Specifically, tone extractions can be facilitated by model-based inference. Two signal models based on summation of sinusoidal waves were employed to represent piano tones. The first model is the traditional General Model, which is a variant of sinusoidal modeling, for representing a tone for high modeling quality; but this model often fails for mixtures of tones. The second model is an instrument-specific model tailored for the piano sound; its modeling quality is not as high as the traditional General Model, but its structure makes source separation easier. To exploit the benefits offered by both the traditional General Model and our proposed Piano Model, we used the hierarchical Bayesian framework to combine both models in the source separation process. These procedures allowed us to recover suitable parameters (frequencies, amplitudes, phases, intensities and fine-tuned onsets) for thorough analyses and characterizations of musical nuances. Isolated tones from a target recording were used to train the Piano Model, and the timing and pitch of individual music notes in the target recording were supplied to our proposed system for different experiments. Our results show that our proposed system gives robust and accurate separation of signal mixtures, and yields a separation quality significantly better than those reported in previous works. / What makes a good piano performance? An expressive piano performance owes its emotive power to the performer's skills in shaping the music with nuances. For the purpose of performance analysis, nuance can be defined as any subtle manipulation of sound parameters including attack, timing, pitch, loudness and timbre. A major obstacle to a systematic computational analysis of musical nuances is that it is often difficult to uncover relevant sound parameters from the complex audio signal of a piano music performance. A piano piece invariably involves simultaneous striking of multiple keys, and it is not obvious how one may extract the parameters of individual keys from the combined mixed signal. This problem of parameter extraction can be formulated as a source separation problem. Our research goal is to extract individual tones (frequencies, amplitudes and phases) from a mixture of piano tones. / Szeto, Wai Man. / Adviser: Wong Kim Hong. / Source: Dissertation Abstracts International, Volume: 73-03, Section: B, page: . / Thesis (Ph.D.)--Chinese University of Hong Kong, 2010. / Includes bibliographical references (leaves 120-128). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Electronic reproduction. [Ann Arbor, MI] : ProQuest Information and Learning, [201-] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstract also in Chinese.

Music recordings contain the mixed contribution of multiple overlapping instruments. In order to better understand the music, it would be beneficial to understand each instrument independently. This thesis focuses on separating the individual instrument recordings within a song. In particular, we propose novel algorithms for separating instrument recordings given only their mixture. When the number of source signals does not exceed the number of mixture signals, we focus on a subclass of source separation algorithms based on joint diagonalization. Each approach leverages a different form of source structure. We introduce repetitive structure as an alternative that leverages unique repetition patterns in music and compare its performance against the other techniques. When the number of source signals exceeds the number of mixtures (i.e. the underdetermined problem), we focus on spectrogram factorization techniques for source separation. We extend single-channel techniques to utilize the additional spatial information in multichannel recordings, and use phase information to improve the estimation of the underlying components.

M. Tech. Electrical Engineering. / Discusses node location as a major problem when considering several areas of application based on wireless sensor networks. Many localisation algorithms have been proposed in the literature to solve the problem of locating sensor nodes in WSN. However, most of these algorithms have poor localisation accuracy and high computational cost. Due to these limitations, this research study considers the modelling of an efficient and robust localisation scheme to determine the location of individual sensor nodes in WSN. To successfully solve this task, this research study focuses on the aspect of improving the position accuracy of wireless sensor nodes in WSN. The study considers a distance based cooperative localisation algorithm called Curvilinear Component Analysis Mapping (CCA-MAP) to accurately localise the sensor nodes in WSN. CCA-MAP is used because it delivers improved position accuracy and computational efficiency.