@InProceedings{Ciba2023_1194,
author = {Stefan Ciba and Mohammed Krini and Amir Rajabi},
booktitle = {Studientexte zur Sprachkommunikation: Elektronische Sprachsignalverarbeitung 2023},
title = {Approach to Speaker-Generalized Spectral Envelope Estimation by Deep Recurrent Neural Network for Speech Reconstruction in a Speech Enhancement System},
year = {2023},
editor = {Christoph Draxler},
month = mar,
pages = {202--208},
publisher = {TUDpress, Dresden},
abstract = {Classical algorithms for Speech Enhancement (SE) often show unsatisfactory
results in loud or exquisite noise scenarios which can be stationary or
transient. Data-driven estimation techniques can outperform classical algorithms
by means of machine learning. In this work machine learning and the classical
approach are brought together in a Speech Enhancement System (SE-System). Furthermore,
the source-filter model of speech production is used in such a way, that
spectral speech features, namely excitation and envelope, can be estimated separately
and subsequently combined. The estimation of the envelope was done by
a Deep Recurrent Neural Network (Deep-RNN) as regressive model. The spectral
envelope is extracted by an Infinite Impulse Response-Filter (IIR-Filter). The
Deep-RNN is trained with many speakers and tested with several unseen speakers,
to approach speaker generalization. Finally, the estimations as well as the potential
of signal improvements, by applying the ideal excitation by the SE-System, are
measured and discussed.},
isbn = {978-3-95908-303-4},
issn = {0940-6832},
keywords = {Poster},
url = {https://www.essv.de/pdf/pdf/2023_202_208.pdf},
}