@misc{10481/96807,
year = {2024},
month = {11},
url = {https://hdl.handle.net/10481/96807},
abstract = {This paper presents a speech restoration system that generates audible speech from articulatory movement data captured using Permanent Magnet Articulography (PMA). Several algorithms were explored for speech synthesis, including classical unit-selection and deep neural network (DNN) methods. A database containing simultaneous PMA and speech recordings from healthy subjects was used for training and validation. The system generates either direct waveforms or acoustic parameters, which are converted to audio via a vocoder. Results show intelligible speech synthesis is feasible, with Mel-Cepstral Distortion (MCD) values between 9.41 and 12.4 dB, and Short- Time Objective Intelligibility (STOI) scores ranging from 0.32 to 0.606, with a maximum near 0.9. Unit selection and recurrent neural network (RNN) methods performed best. Informal listening tests further confirmed the effectiveness of these},
organization = {MICIU/AEI/10.13039/501100011033  PID2022-141378OBC22},
organization = {ERDF/EU},
publisher = {Internation Speech Communication Association (ISCA)},
title = {Direct Speech Synthesis from Non-audible Speech Biosignals: A Comparative Study},
doi = {10.21437/IberSPEECH.2024-18},
author = {Lobato Martín, Javier and Pérez Córdoba, José Luis and González López, José Andrés},
}