Djupinlärningsbaserat Tydligare Tal i Digitalt Underhållningsmaterial

Speech intelligibility in stereo audio remains a persistent challenge for digital entertainment content, particularly for users with hearing impairments. This thesis investigates the potential of a deep learning-based approach to improve speech clarity in stereo formats, using content from Sveriges Television AB (SVT). A fullband speech enhancement model, DeepFilterNet, was applied through two processing strategies: discrete-channel enhancement and a novel Center Enhanced Stereo (CES) method. These were compared to SVT’s existing FFmpeg-based Dialogue Enhance (DE) solution through both objective and subjective evaluations. Objective analysis was performed using non-intrusive metrics Non-Intrusive Speech Quality Assessment (NISQA) and Deep Noise Supression Mean Opinion Score (DNSMOS), while subjective assessments involved over 1,000 SVT Play users rating paired clips across diverse genres. Results indicate that CES consistently improves speech clarity without compromising background audio quality. These findings demonstrate the feasibility of integrating deep learning models into existing streaming pipelines, offering meaningful improvements in accessibility and user experience.