Text-informed speech inpainting via voice conversion

摘要

The problem of speech inpainting consists in recovering some parts in a speech signal that are missing for some reasons. To our best knowledge none of the existing methods allows satisfactory inpainting of missing parts of large size such as one second and longer. In this work we address this challenging scenario. Since in the case of such long missing parts entire words can be lost, we assume that the full text uttered in the speech signal is known. This leads to a new concept of text-informed speech inpainting. To solve this problem we propose a method that is based on synthesizing the missing speech by a speech synthesizer, on modifying its vocal characteristics via a voice conversion method, and on filling in the missing part with the resulting converted speech sample. We carried subjective listening tests to compare the proposed approach with two baseline methods.