Restoration of degraded speech based on example-based speech enhancement

Abstract

It is becoming popular to record meeting discussions or daily conversations as life log speech data. The recoded speech is highly degraded by quantization noises if the amplitude of the input speech level is very low. Recently a large database can be available as a cloud service and it is possible to collect a large amount of speech data. In this paper, we propose a novel speech enhancement method which utilizes large clean speech data as an example database. In our method, the most suitable speech sequences are retrieved in the clean database and modified to replace the degraded speech with the clean segments. In the preliminary experiment, we obtained effectiveness of the example database at low bit quantization conditions. Accordingly, the speech quality that deteriorated by quantization can be improved by the example database.