Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Electrical and Computer Engineering

Supervisor

Dr. Vijay Parsa

Abstract

Traditional landline and cellular communications use a bandwidth of 300 - 3400 Hz for transmitting speech. This narrow bandwidth impacts quality, intelligibility and naturalness of transmitted speech. There is an impending change within the telecommunication industry towards using wider bandwidth speech, but the enlarged bandwidth also introduces a few challenges in speech processing. Echo and noise are two challenging issues in wideband telephony, due to increased perceptual sensitivity by users.

Subjective and/or objective measurements of speech quality are important in benchmarking speech processing algorithms and evaluating the effect of parameters like noise, echo, and delay in wideband telephony. Subjective measures include ratings of speech quality by listeners, whereas objective measures compute a metric based on the reference and degraded speech samples. While subjective quality ratings are the "gold - standard'', they are also time- and resource- consuming. An objective metric that correlates highly with subjective data is attractive, as it can act as a substitute for subjective quality scores in gauging the performance of different algorithms and devices.

This thesis reports results from a series of experiments on subjective and objective speech quality evaluation for wideband telephony applications. First, a custom wideband noise reduction database was created that contained speech samples corrupted by different background noises at different signal to noise ratios (SNRs) and processed by six different noise reduction algorithms. Comprehensive subjective evaluation of this database revealed an interaction between the algorithm performance, noise type and SNR. Several auditory-based objective metrics such as the Loudness Pattern Distortion (LPD) measure based on the Moore - Glasberg auditory model were evaluated in predicting the subjective scores. In addition, the performance of Bayesian Multivariate Regression Splines(BMLS) was also evaluated in terms of mapping the scores calculated by the objective metrics to the true quality scores. The combination of LPD and BMLS resulted in high correlation with the subjective scores and was used as a substitution for fine - tuning the noise reduction algorithms.

Second, the effect of echo and delay on the wideband speech was evaluated in both listening and conversational context, through both subjective and objective measures. A database containing speech samples corrupted by echo with different delay and frequency response characteristics was created, and was later used to collect subjective quality ratings. The LPD - BMLS objective metric was then validated using the subjective scores.

Third, to evaluate the effect of echo and delay in conversational context, a realtime simulator was developed. Pairs of subjects conversed over the simulated system and rated the quality of their conversations which were degraded by different amount of echo and delay. The quality scores were analysed and LPD+BMLS combination was found to be effective in predicting subjective impressions of quality for condition-averaged data.


Share

COinS