Speech processing and analysis

Abstract

In modern applications such as telecommunication, hand-free gadgets, hearing aid and human-machine interaction, the audio or speech signal is usually contaminated by noise and reverberations. Due to such challenging environment, the sound recording, speech information extraction usually requests sophisticated algorithms to increase the speech quality. The speech signal can be enhanced by noise reduction with single sensor or sensor array, channel de-reverberation and signal separation. In this project, a MATLAB toolkit was designed to achieve high fidelity speech acquisition. The toolkit is able to extract / separate audio signals from observations, under scenario of both instantaneous linear mixtures and convolutive mixtures. Users can choose different algorithms based on their judgments of their situations. This toolbox also includes a function for real-time signal separation using two microphones. A simulation package was included in this toolkit for performance assessment. Algorithms were evaluated in terms of convergence speed, robustness, and immunity to noise.