Loading...

State-Space Model for Speech Enhancement in Presence of Additive Noise and Packet Loss

Sarafnia, Ali | 2013

486 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: English
  3. Document No: 44239 (55)
  4. University: Sharif University of Technology
  5. Department: Science and Engineering
  6. Advisor(s): Ghorshi, Mohammad Ali
  7. Abstract:
  8. This thesis aims to develop speech enhancement methods in the presence of additive noise, packet loss and band-limitation of speech signal. Noise reduction of speech improves the perceived quality and intelligibility. High performance noise reduction methods are based on Bayesian methods requiring estimates of the parameters of the functions that describe the likelihood and the prior distributions of the signal and noise processes. The Bayesian noise reduction method which is used in this thesis is Kalman filter.In packet-based speech processing applications such as Internet and Voice over Internet Protocol (VoIP), it is possibly expected that some packets during signal transmission are being lost. Therefore, a packet loss concealment phase is required to estimate and interpolate lost packets. Two methods of packet loss concealment are implemented at this thesis. At the first method adjacent packets of lost packet are employed to interpolate the lost frame while second method is based on prediction of lost packet using preceding packet. The last phase of speech enhancement is defined as bandwidth extension of narrowband speech signal. The bandwidth extension problem can be defined as estimation of missing signal components in a particular frequency range. Bandwidth extension problem could be solved efficiently by implementing two sets of stages separately based on source-filter model of speech production. The first stage is extension of excitation signal and the second stage is spectral envelope extension. Codebook method is utilized for purpose of envelope extension in this thesis
  9. Keywords:
  10. Speech Enhancement ; Bayesian Method ; Kalman Filters ; Source-Filter Model ; Excitation Signal

 Digital Object List

 Bookmark

No TOC