Natural click processing through wavelet analysis and extrapolation for speech enhancement

Nowadays, several applications based on the use of the speech technology are expecting a higher quality of speech. Applications like movie dubbing, advertisements, and spoken books are just some examples. One factor that has considerably influenced speech quality is the speaker´s voice itself. Ideally, the speaker´s voice should not contain any natural clicks as well as other kinds of audible degradation. However, in practice, such a condition is not always fulfilled. This paper discusses an automatic procedure to treat the click-type degradations present in speech databases. The approach in question includes two main stages: click detection and treatment. The first stage is based on a continuous wavelet transform (CWT), which is often used to analyze transient signals (such as clicks). The second stage performs a treatment of the previously detected clicks by using an extrapolation technique. By considering an ad hoc database, Mean Opinion Score (MOS) results obtained from absolute category rating (ACR) tests confirm the applicability of the proposed approach for improving speech quality.