Data-driven Speech Intelligibility Enhancement and Prediction for Hearing Aids

Hearing impairment is a widespread problem around the world. It is estimated that one in six people are living with some degree of hearing loss. Moderate and severe hearing impairment has been recognised as one of the major causes of disability, which is associated with declines in the quality of life, mental illness and dementia. However, investigation shows that only 10-20\% of older people with significant hearing impairment wear hearing aids. One of the main factors causing the low uptake is that current devices struggle to help hearing aid users understand speech in noisy environments. For the purpose of compensating for the elevated hearing thresholds and dysfunction of source separation processing caused by the impaired auditory system, amplification and denoising have been the major focuses of current hearing aid studies to improve the intelligibility of speech in noise. Also, it is important to derive a metric that can fairly predict speech intelligibility for the better development of hearing aid techniques. This thesis aims to enhance the speech intelligibility of hearing impaired listeners. Motivated by the success of data-driven approaches in many speech processing applications, this work proposes the differentiable hearing aid speech processing (DHASP) framework to optimise both the amplification and denoising modules within a hearing aid processor. This is accomplished by setting an intelligibility-based optimisation objective and taking advantage of large-scale speech databases to train the hearing aid processor to maximise the intelligibility for the listeners. The first set of experiments is conducted on both clean and noisy speech databases, and the results from objective evaluation suggest that the amplification fittings optimised within the DHASP framework can outperform a widely used and well-recognised fitting. The second set of experiments is conducted on a large-scale database with simulated domestic noisy scenes. The results from both objective and subjective evaluations show that the DHASP-optimised hearing aid processor incorporating a deep neural network-based denoising module can achieve competitive performance in terms of intelligibility enhancement. A precise intelligibility predictor can provide reliable evaluation results to save the cost of expensive and time-consuming subjective evaluation. Inspired by the findings that automatic speech recognition (ASR) models show similar recognition results as humans in some experiments, this work exploits ASR models for intelligibility prediction. An intrusive approach using ASR hidden representations and a non-intrusive approach using ASR uncertainty are proposed and explained in the third and fourth experimental chapters. Experiments are conducted on two databases, one with monaural speech in speech-spectrum-shaped noise with normal hearing listeners, and the other one with processed binaural speech in domestic noise with hearing impaired listeners. Results suggest that both the intrusive and non-intrusive approaches can achieve top performances and outperform a number of widely used intelligibility prediction approaches. In conclusion, this thesis covers both the enhancement and prediction of speech intelligibility for hearing aids. The proposed hearing aid processor optimised within the proposed DHASP framework can significantly improve the intelligibility of speech in noise for hearing impaired listeners. Also, it is shown that the proposed ASR-based intelligibility prediction approaches can achieve state-of-the-art performances against a number of widely used intelligibility predictors.