UltraSpeech
Han Ding, Yizhan Wang, Hao Li, Cui Zhao, Ge Wang, Xi Wei, Jizhong Zhao
Abstract
Speech enhancement can benefit lots of practical voice-based interaction applications, where the goal is to generate clean speech from noisy ambient conditions. This paper presents a practical design, namely UltraSpeech, to enhance speech by exploring the correlation between the ultrasound (profiled articulatory gestures) and speech. UltraSpeech uses a commodity smartphone to emit the ultrasound and collect the composed acoustic signal for analysis. We design a complex masking framework to deal with complex-valued spectrograms, incorporating the magnitude and phase rectification of speech simultaneously. We further introduce an interaction module to share information between ultrasound and speech two branches and thus enhance their discrimination capabilities. Extensive experiments demonstrate that UltraSpeech increases the Scale Invariant SDR by 12dB, improves the speech intelligibility and quality effectively, and is capable to generalize to unknown speakers.