Speech intelligibility comparison of standalone and two-stage deep learning architectures for behind-the-ear-to-binaural enhancement

Abstract The relative perceptual performance of alternative deep learning based speech processing architectures remains uncertain under complex acoustic conditions. This study compared speech intelligibility, in listeners with normal hearing and with mild-to-moderate hearing loss, obtained with two processing models: a speech-target-oriented two-stage model (SED + U-Net), designed to estimate speech-target activity before denoising, and a standalone U-Net model. These two architectures were selected as representative instances of two contrasting design philosophies in deep-learning-based speech enhancement: unconstrained direct denoising versus denoising explicitly guided by a prior estimate of speech-target activity. Intelligibility was assessed using a keyword-recognition task with sentences from the Chilean SHARVARD corpus, presented under controlled virtual acoustic conditions. Twenty-four adults participated in the study, consisting of 12 normal-hearing (NH) listeners and 12 listeners with mild-to-moderate hearing loss (HL). Experimental conditions varied by processing model, masker configuration (Masker Around, Masker Front, and Masker Side), reverberation time (RT = 0.2 s and 0.8 s), and signal-to-noise ratio (− 3, 0, 3, 6, and 9 dB). Keyword recognition was analyzed using a binomial generalized linear mixed-effects model. Results showed significant effects of SNR, processing model, masker configuration, and reverberation time. Unlike the original hypothesis, the standalone U-Net performed better in keyword recognition than the speech-target-oriented two-stage model, emphasizing the need for perceptual validation in evaluating speech-processing systems.