Memory and complexity reduction for inventory-style speech enhancement systems

In this paper we are presenting a method that provides a dramatic reduction in memory requirement and computational complexity for an inventory-style speech enhancement scheme with only a small impact on the perceptual quality of the output of the system. Inventory-style or corpus-based speech enhancement generally attempts to generate a clean speech signal from a noisy speech signal by first estimating the characteristics of the underlying clean signal and then recreating it via corpus-based speech synthesis. As such, inventory-based enhancement is very different from most traditional methods which are typically relying on adaptive filtering or spectral subtraction. The advantage of inventory-based enhancement is its (principal) ability to deliver a very natural sounding output. A significant drawback is its large memory requirement and its large computational complexity (in comparison to traditional techniques)
¹
. The method proposed in this paper allows for a flexible reduction of the memory requirement as a function of the desired perceptual quality of the output. A data reduction by almost factor 10 is achievable with only minor losses in perceptual quality. Furthermore, a significant reduction of computational complexity is a possible choice in the implementation of the procedure.