Speaker-Adaptive Postfiltering
Few-shot neural postfilters that adapt embedded text-to-speech to a new speaker from very little data. This work is the core of my M.S. thesis.
Text-to-speech running on an embedded device cannot afford a large model, and adapting it to a new speaker usually means collecting far more recordings than a user is willing to provide. This work attacks both constraints at once with speaker-adaptive neural postfilters, a light model that sits after synthesis and repairs the characteristic degradation, adapted from only a few utterances.
The study compares Transformer- and GAN-based postfilters under limited adaptation data, and examines how to stay data- and parameter-efficient without giving up quality. It grew out of my M.S. thesis at Ozyegin University with Prof. Cenk Demiroglu.
Citation
@article{eren2023postfiltering,
title = {Deep learning-based speaker-adaptive postfiltering with limited adaptation data for embedded text-to-speech synthesis systems},
author = {Eren, Eray and Demiroglu, Cenk},
journal = {Computer Speech \& Language},
volume = {81},
pages = {101520},
year = {2023},
doi = {10.1016/j.csl.2023.101520}
}