Normal-rate to fast-rate speech conversion using non-linear compression maps

Abstract

This paper presents a new technique to convert normal-rate speech into intelligible fast-rate, speeded speech. Speeded speech has long been recognized for its potential to improve spoken media comprehension; however, current tools to significantly speed playback of non-text media are insufficient due to their reliance on inaccurate phoneme analysis. With the ever increasing amount of non-text media online, a method to speed playback that is agnostic of phonemes is needed. Our technique uses spectral and source components of the acoustics to generate a non-linear compression map that characterizes how conversational-rate speech signals are compressed to achieve analogue fast-rate speech signals. A data set containing conversational- and fast-rate speech pairs was processed to determine compression maps corresponding to each pair. A Recursive Neural Network (RNN) was trained on the set of normal-rate speech and the corresponding compression maps. The RNN was then used to generate compression maps for novel normal-rate speech and ultimately output a fast-rate speech signal. Elicited fast-rate speech and speeded speech conversions technique are now being compared perceptually for intelligibility and naturalness.