Perceptual quality enhancement of speech corpora under optimal conditions of click detection

One factor which certainly affects the quality of speech recordings is the speakerpsilas voice characteristics. Ideally, the speakerpsilas voice should not contain any natural clicks (small cracks produced by the human vocal tract). In order to reduce the occurrence of such clicks, speech processing techniques have recently been proposed. These approaches usually consider two main stages: detection and treatment. The detection stage is advantageously carried out by using an automatic procedure (less tiring than a manual one). However, manual detection has a better performance. Our objective here is to determine the maximum improvement which could be achieved in speech quality if an automatic detection technique with a performance similar to that of the manual one were accomplished. We first present a manual detection and then a treatment by extrapolation of the clicks existing in a 45-minute long speech database. In order to compare the quality of speech recordings before and after click processing, formal listening tests are carried out. Results point out a significant improvement in speech quality. Up to our knowledge, this is the first paper which establishes an achievable bound in speech enhancement when a click treatment is considered.