Use this URL to cite or link to this record in EThOS:
Title: Non-intrusive estimation of acoustic parameters from degraded speech
Author: Eaton, Derek James
ISNI:       0000 0004 6422 6906
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Access from Institution:
Estimation of the acoustic parameters Signal-to-Noise Ratio (SNR), Reverberation Time (T60), Direct-to-Reverberant Ratio (DRR), and clipping level from degraded speech are open research questions. These parameters are important for determining speech quality and intelligibility, and they are widely applicable to speech enhancement and speech recognition systems. Whilst SNR, T60, and DRR are useful priors for dereverberation schemes, indications of clipping and the clipping level are useful for signal restoration. This thesis investigates how accurately and robustly to noise and, in the case of clipping detection, robustness to the coding and decoding process it is possible to estimate these parameters non-intrusively from degraded speech in real-time or near real-time, and introduces a range of novel algorithms. Alongside the algorithms, an international research challenge was staged for which a novel noisy reverberant speech corpus was developed to determine the state-of-the-art in T60 and DRR estimation. In tests, the algorithms presented in this thesis were highly competitive, being able to estimate T60 with Pearson correlation coefficient, ρ = 0.608 and DRR with ρ = 0.314. Both algorithms achieved very low computational complexity with Real-Time Factors (RTFs) of 0.0164 and 0.0589 respectively. The clipping detection algorithms achieved F1 approximately 0.6 for Global System For Mobile Communications (GSM) 06.10 decoded speech, babble noise at 20 dB SNR clipping levels in the range -5 to -20dBFS, and also produce an estimate of the original unclipped signal level.
Supervisor: Naylor, Patrick Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral