Project Overview
An EPSRC funded collaborative research project (Pro. No GR/S30238/01) between Brunel University, Southampton University, University of East Anglia, Norwich and Queen's University Belfast.
The project is divided into three parts:
- Prediction and Modelling of Noise and its Diversity in Mobile Environments (Southampton University)
- Speaker Modelling, Adaptation and Identification in Noise in Mobile Environments (UEA, Norwich)
- Enhancement of Noisy Speech in Mobile Environments (Brunel University)
Mobile phones are often used in noisy outdoor environments such as a noisy street, airports, cafe, or in a moving car/train. The quality and intelligibility of speech can be severely degraded by the ambient noise. Therefore noise reduction is an increasingly important aspect of improving the quality of service (QoS) and reliability of speech communication. With the increasingly deployment of speech recognition and voice-based systems across a wide range of multimedia mobile services, it is important to the users and providers of mobile phones that speech communication and access to voice recognition systems is not impaired by noise. Noise degrades the accuracy of automatic speech recognition even for such modest tasks as name dialling, automatic directory enquiry, or voice control of the accessories in a moving car. Furthermore, the future generation of very low bit rate coders will increasingly depend on correct speech classification and accurate estimation of speech parameters for improved performance. The aim of this proposal is to develop an integrated system for both speech enhancement and speech recognition. This will be achieved through the development of a unified speech model for both speech recognition and synthesis/reconstruction, together with decision-tree predictive models of non-stationary noise sources typically encountered in mobile environments.
Objective
- Re-configurable multi-resolution decision-tree modelling and prediction of the time-varying spectrum of non-stationary noise sources in mobile phone environments such as streets, cars, trains, airports, and conferences.
- Development of a unified model of speech integrating features for both recognition and synthesis and including speaker adaptation.
- Dynamic multi-resolution models of temporal and spectral masking effects to mitigate the perceptual impact of the distortion of low-amplitude short-duration speech segments with lower than the average SNR.
People
Academic Staff
Research Staff and Students
Database
NoiseX_0 NoiseX_1
Aurora
TIDIGITS ADULTS
Cellular Telephone Acoustic-Phonetic Continuous Speech Corpus (CTIMIT)
Progress Report
Project meeting 23/01/2004
- Speech Communication for Mobile and Hands-Free Devices in Noisy Environments by Prof Saeed Vaseghi
- Analysis of Effects of Train/car Noise in Formant Track Estimation by Qin Yan
- Analysis of the Effects of Train Noise on Recognition Rate Using Formants and MFCC by Esfandiar Zavarehei
- Some Related Research Topics by Dr Ben Milner
- Progress Report by Jonathan Darch
- Modelling of Non-stationary Noise by Ioannis Andrianakis
Project meeting 25/05/2004
- Formant track restoration in Train noisy speech by Qin Yan
- Speech Recognition in Noise by Esfandiar Zavarehei
- Progress Report by Jonathan Darch
- Modelling of Mel Frequency Features forNon-stationary Noise by Ioannis Andrianakis
Project meeting 28/07/2004
- Formant based Synthesizer by Qin Yan
- Frequency Kalman Noise Reduction by Esfandiar Zavarehei
- Progress Report by Jonathan Darch
- Modelling of Non-stationary Noise by Ioannis Andrianakis
Project meeting 14/02/2005
- Report of Formant Tracking LP models and Plans on Integration of Harmonic plus Noise model by Qin Yan
- Analysis of the Effects of Train Noise on Recognition Rate Using Formants and MFCC by Esfandiar Zavarehei
- Progress Report by Jonathan Darch
- Modelling of Non-stationary Noise by Ioannis Andrianakis
Project meeting 06/07/2005
- FLTP with HNM in Speech Enhancement by Qin Yan
- Meet 6 by Esfandiar Zavarehei
- Progress Report by Jonathan Darch
- Bayesian Methods for Speech Enhancement by Ioannis Andrianakis
Project meeting 23/11/2005
- Refinement of FLTP- HNM Speech Enhancement System by Qin Yan
- Meet 7 by Esfandiar Zavarehei
- Progress Report by Jonathan Darch
- Bayesian Methods for Speech Enhancement by Ioannis Andrianakis
Links
News
List of MPhil/PhD Projects in MultiMedia Mobile Digtial Signal Processing
Advanced Signal processing and Noise Reduction 3rd Ed, S. Vaseghi, John Wiley 2006
DeNoise Toolkit
This is a collection of established and new speech denoising methods developed at Brunel in collaboration with Southampton and UEA (Project sponsored by EPSRC).
Voice Morph
VoiceMorph is a software tool developed in our lab for analysis, modelling modification and of voice profile parameters including speaker correlates and Accent correlates of voice.
MSc Digital Signal Processing
Multimedia, communication and inteligent systems.