Statistical significance of a perceptually-derived noise suppression function for speech enhancement systems

Adrien DANIEL - Post-Doc, MM department
Multimedia Communications

Date: -
Location: Eurecom

This presentation is organized in two parts. First, basic statistical tools which are commonly used in psychology will be described. Sampling distributions and the logic behind hypothesis testing will be introduced, and a particular focus will be given to testing applied to means. In a second part, these tools will be applied to the following recent work. Most noise reduction systems for speech enhancement consist of a gain function applied in the frequency domain. In particular, a theoretically-optimal gain function can be derived based on mathematical criteria. We took a different way by designing a listening test which aimed to derive a perceptually-optimal gain function. After a description of this test and its results, we will use the previous statistical knowledge to answer the question: "does this perceptual function show a significant departure from the theoretical function?".