Advanced signal processing techniques for speaker recognition and communications.

摘要

Advanced signal processing techniques can help us well analyze signals of interests and perform proper operations on signals of interests for many useful applications.;In this dissertation, we aim at developing signal processing techniques for speaker recognition (e.g. feature extraction, classifier design) and for communications (e.g. filtering, modulation, beamforming). In the first part, we focus on speaker recognition. For gender identification, we proposed a pitch-based system with a two-stage classifier to ensure accurate identification and low complexity. Our pitch extraction algorithm is able to produce robust pitch estimations. The proposed system is speech language/content independent, microphone independent, and robust against noisy recording conditions. For large population speaker identification under noisy conditions, we proposed a fuzzy-clustering-based decision tree approach. Our approach aims at partitioning the whole population into subgroups in a hierarchical way. We only apply mel-frequency cepstral coefficients (MFCC) + Gaussian mixture model (GMM) to the leaf node which has a very small population size and hence MFCC+GMM is effective. To achieve a low probability of classification error, we adopted fuzzy clustering in constructing the decision tree, i.e., a speaker may belong to multiple nodes at each level of the tree. We derived six features (including pitch and five vocal source characteristics) and constructed a six-level decision tree. Compared to MFCC+GMM, our proposed approach achieves much higher accuracy with much less complexity.;In the second part, we study signal processing for communications. To address limitations of orthogonal frequency-division multiplexing (OFDM), we proposed a multi-carrier transceiver based on frequency-shift filter. Compared with OFDM, the proposed transceiver is much less sensitive to carrier frequency offset and has a lower peak-to-average ratio; moreover, the proposed transceiver has the advantage of being able to mitigate strong co-channel interference and strong narrowband interference. To improve the anti-jamming capability of a space-time block coding system over fading channels, we proposed to use Capon's beamforming to extract the intended signal while suppressing jamming signals coming from directions different from the intended signal. We evaluate the anti-jamming performance and the system cost with different numbers of array elements under different types of jamming signals.