New Approaches to Measuring and Improving Speech Intelligibility

Abstract

Hearing impaired and elderly listeners have trouble understanding speech in the presence of noise, which may be caused by the reduced speech processing ability in the brain. Slowing down the speech that is presented, may provide the listener’s brain more opportunities to process stimuli or collect crucial speech cues. Speech rate modulation algorithms that generate a slowed down version of speech have been suggested as an approach to improve comprehension. However, existing non-uniform speech rate modulation algorithms are difficult to experimentally replicate, often requiring specialized linguistics knowledge of the speech, such as the locations of phonetic elements. Additionally, no previous study has objectively measured intelligibility of rate modulated speech. This thesis proposes new methods for objectively measuring rate modulated speech characteristics through Speech Rate Based Intelligibility Metrics (SRBIM), and for improving speech through SOLely Acoustic Non-Uniform (SOLANU) speech rate modulation. SRBIM was compared to existing metrics in situations where the existing metrics were known to work, involving speech contaminated by various noise types and noise levels, and speech processed through three types of noise reduction algorithms. New and existing metric behaviours were compared by analyzing the correlation and scatter plots between all possible metric pairs. The metrics were also used to rank the performances of the three noise reduction algorithms. Simulations were extended to speech processed through two existing methods and two variants of the proposed SOLANU speech rate modulation algorithm. Analysis through correlation and scatter plots, as well as subjective observations were included. When compared against two of the three existing metrics, SRBIM showed a linear relationship in measuring speech intelligibility of noisy speech and speech processed through noise reduction. Subjective observations showed that clean rate modulated speech was more intelligible than noisy rate modulated speech. SRBIM were the only metrics that reflected the impact of rate modulation on clean speech compared to noisy speech. Thus, SRBIM has potential in measuring speech intelligibility of noisy speech, and speech that had been enhanced through noise reduction or rate modulation. Future research involving human observers with rate modulated speech is recommended to further validate SRBIM to subjective intelligibility scores.