Date of Thesis


Thesis Type

Masters Thesis

Degree Type

Master of Science in Electrical Engineerging

First Advisor

Robert Nickel


A persistent challenge in speech processing is the presence of noise that reduces the quality of speech signals. Whether natural speech is used as input or speech is the desirable output to be synthesized, noise degrades the performance of these systems and causes output speech to be unnatural. Speech enhancement deals with such a problem, typically seeking to improve the input speech or post-processes the (re)synthesized speech. An intriguing complement to post-processing speech signals is voice conversion, in which speech by one person (source speaker) is made to sound as if spoken by a different person (target speaker). Traditionally, the majority of speech enhancement and voice conversion methods rely on parametric modeling of speech. A promising complement to parametric models is an inventory-based approach, which is the focus of this work. In inventory-based speech systems, one records an inventory of clean speech signals as a reference. Noisy speech (in the case of enhancement) or target speech (in the case of conversion) can then be replaced by the best-matching clean speech in the inventory, which is found via a correlation search method. Such an approach has the potential to alleviate intelligibility and unnaturalness issues often encountered by parametric modeling speech processing systems. This work investigates and compares inventory-based speech enhancement methods with conventional ones. In addition, the inventory search method is applied to estimate source speaker characteristics for voice conversion in noisy environments. Two noisy-environment voice conversion systems were constructed for a comparative study: a direct voice conversion system and an inventory-based voice conversion system, both with limited noise filtering at the front end. Results from this work suggest that the inventory method offers encouraging improvements over the direct conversion method.