Noise-Adaptive Statistical Speech Synthesis in Varying Acoustic Environments

Period: August 2013 – March 2015
Grant type: Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research (Research Activity Start-up Support)
Issue number: 25880027
URL: https://kaken.nii.ac.jp/ja/grant/KAKENHI-PROJECT-25880027/

In recent years, speech synthesis based on statistical models, which has been actively researched and developed, has become capable of achieving the same level of clarity as natural human speech in clean environments. In this study, we investigated the effects of environmental noise on human speech production and hearing, and based on the findings, aimed to realize a method to automatically and adaptively improve the clarity of speech synthesis systems under various noise conditions.
For this purpose, we recorded speech with the Lombard effect, which is seen in noisy environments. We also introduced an auditory model called Glimpse that automatically predicts intelligibility, taking into account the masking effect of noise in noisy environments, and proposed a method to automatically and adaptively change the spectral parameters of a speech synthesis system based on this auditory model, and confirmed that the intelligibility of the speech synthesis system was improved.