Real-time multichannel source separation of acoustic signals

UoM administered thesis: Phd

  • Authors:
  • Shahab Faiz

Abstract

The project is concerned with the design and the real-time implementation of a multichannel source separation algorithm for acoustic source signals. The acoustic signals considered are non-stationary and do not have any information available that helps in the separation of sources in a supervised way. The source signals considered in this research project are speech signals and percussion instrument signals. This is a non-trivial and highly challenging problem in the field of digital signal processing. So far the algorithms presented (in the literature) for unsupervised speech separation have shown partial success in a constrained environment for only two sources. For percussion instruments there is no system presently available that can separate and preserve the integrity of the sounds using fewer microphones than sources (typically seven or more sources).A novel hybrid algorithm will be presented in this thesis that separates speech signals for at least three sources in a real room environment. The experimental results show that the algorithm gives improved performance and has a low computational load. Due to its adaptive methodology and its efficient use of computational resources, it is well suited for real-time implementation. For further enhancement of the signal-to-interference ratio in a real room environment following the hybrid algorithm, a harmonic alignment algorithm has been proposed. It exploits the harmonic property of speech by working on pitch, for which a novel frequency estimator algorithm has been proposed. This frequency estimator algorithm has the least computational load compared to all competitive algorithms whilst giving optimum uniform variance performance. This research also highlights the limitations of performing the separation in a completely unsupervised way. For that it proposes a theory on how to avoid those limitations and presents a novel source separation system that is partially supervised. The application of this system to percussion instruments in a real room recording environment achieves separation with fewer microphones while the integrity of the sounds is preserved. This system can lead to the development of a commercial recording software system for offline and real-time separation and processing of percussion instrument sounds.

Details

Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date1 Aug 2014