Encrypted domain cloud-based speech noise reduction with comb filter
Abukari Mohammed Yakubu, Namunu Chinthaka Maddage, Pradeep K. Atrey · 2016
During the acquisition process of speech by recording, noise often contaminates the signal which degrades its quality and (1) makes it unpleasant for human perception and (2) causes inaccuracies in speech processing application such as speech recognition and speech transcription. Due to the storage and computational requirements of speech records, clients constrained in resources often outsource them to Cloud Data Centers (CDCs). However, the use of third party servers such as CDCs raises security concerns. In this work we enhance the quality of speech records contaminated with humming noise in a privacy-preserving manner in Encrypted Domain (ED). Our proposed scheme is based on (K;N) Shamir's Secret Sharing (SSS) method and Finite Impulse Response (FIR) comb filter. Experimental results for our proposed scheme in ED produces similar results as compared to its Plaintext Domain (PD) implementation version with minimal overheads while maintaining security and privacy.