Sparse representations for audio

点击翻译按钮获取中文摘要

Over the past decades, advances in digital audio processing have allowed numerous applications to emerge, mature and in many cases become commodities in everyday life. Cellphones, smartphones, MP3 players, Voice over IP (VoIP) software and even Youtube all have in common advanced audio processing in their core. Specifically, speech and/or audio compression algorithms allow using transmission bandwidth and storage space sparingly while providing users the audio quality they expect for a given application. The need for performant audio processing algorithms, and in particular compression algorithms, can be expected to grow with demanding applications such as digital multichannel radio broadcast for mobile users (in cars, for example) and wireless multimedia of all sorts to be used on mobile devices.The purpose of this research project is to study new approaches for analysing, decomposing and representing audio signals in order to obtain sparse and very sparse representations suitable for compression, but also for signal enhancement, sound source separation and signal classification. Instead of attempting to improve on the classical approaches based on block-by-block analysis using transforms and prediction filters, the focus will be on adaptive signal segmentation, perceptually derived basis functions and audio object representation. The resulting algorithms will thus produce audio signal approximations which will range from synthetic sound objects (as in MIDI, for example) to high-fidelity source reproduction, rather than from high distortion to low distortion. To achieve the next significant gains in sparse audio representations, waveform tracking constraints have to be relaxed, and ultimately completely replaced by excitation pattern similarities in the auditory system. This research project aims to contribute knowledge and technological advances towards that goal.