ABSTRACT
Accent conversion is a field closely related to voice conversion that converts audio from one accent to another. Traditional methods for accent conversion rely on MFCC features and the use ofparallel accent speech corpora, which can be ineffective. Recent researches look into models that use new features and without the need of a parallel corpus. This paper proposes a Spanish accent conversion system, which can be easily extended to more accents or other languages. First of all, we build a Spanish ASR system that allows us to do forced alignment. Then, we extract CELP features and trains the accented audios using a phone-based encoder-decoder model. Although it has not been entirely finished due to time limitation, we plan to finish constructing this the accent conversion system in the future.
Index Terms — Accent conversion, CELP, Codeexcited linear prediction, Spanish, Spanish accent, Speex