NASTAR: NOISE ADAPTIVE SPEECH ENHANCEMENT WITH TARGET-CONDITIONALRESAMPLING

Abstract: For speech-related applications, the acoustic mismatch between training and testing conditions can severely affect the performance. In this paper, we propose a novel noise adaptive speech enhancement with target-conditional resampling (NASTAR), which reduces the acoustic mismatch with only one noisy speech sample in a target environment.

Pseudo-Noise & Relevant-Cohort Demonstration

In the following sections, the query-noisy-speech contaminated by the selected noise condition is given in different SNR levels(from -8dB to 8dB in 4dB steps). The pseudo-noise estimated from the different query-noisy-speech in the assigned SNR level and the top-10 similar samples of relevant-cohort from the 0-dB query-noisy-speech are listed.
ACVacuum
Babble
CafeRestaurant
Car
MetroSubway