Acoustic Compensation Method for Accepting Different Recording Devices in Body-Conducted Voice Conversion
Daisuke Deguchi, Hironori Doi, Tomoki Toda, Hiroshi Saruwatari, Kiyohiro Shikano · Institutional Repositories DataBase (IRDB) · 2010
This paper presents an acoustic compensation method in body-conducted speech conversion that automatically compensates for acoustic differences caused by changes in recording conditions. An enhancement process for bodyconducted speech recorded with a Non-Audible Murmur (NAM) microphone has successfully applied a statistical voice conversion technique. Speech waveforms are generated from acoustic parameters of normal speech estimated from those of bodyconducted speech with a conversion model previously trained using stereo data of those two types of speech. This framework suffers from mismatched conditions between training and conversion processes. To alleviate this issue, an unsupervised acoustic compensation method based on constrained maximum likelihood linear regression (CMLLR) has been proposed and its effectiveness has been reported in the compensation of acoustic differences caused by attachment location of the NAM microphone. This paper further applies the CMLLR-based acoustic compensation method to the compensation of acoustic differences caused by different recording devices and evaluates its effectiveness. The experimental results demonstrate that the proposed method effectively reduces quality degradation and converted speech caused by differences in recording devices as well as attachment location of the NAM microphone.