Learning Filterbanks from Raw Speech for Phone Recognition

We train a bank of complex filters that operates on the raw waveform and is fed into a convolutional neural network for end-to-end phone recognition. These time-domain filterbanks (TD-filterbanks) are initialized as an approximation of mel-filterbanks, and then fine-tuned jointly with the remaining...

Full description

Saved in:
Bibliographic Details
Published in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 5509 - 5513
Main Authors Zeghidour, Neil, Usunier, Nicolas, Kokkinos, Iasonas, Schaiz, Thomas, Synnaeve, Gabriel, Dupoux, Emmanuel
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.04.2018
Subjects
Computer architecture
Convolution
Scattering
Speech recognition
Time-domain analysis
Training
Online AccessGet full text

Cover

More Information
Summary:We train a bank of complex filters that operates on the raw waveform and is fed into a convolutional neural network for end-to-end phone recognition. These time-domain filterbanks (TD-filterbanks) are initialized as an approximation of mel-filterbanks, and then fine-tuned jointly with the remaining convolutional architecture. We perform phone recognition experiments on TIMIT and show that for several architectures, models trained on TD- filterbanks consistently outperform their counterparts trained on comparable mel-filterbanks. We get our best performance by learning all front-end steps, from pre-emphasis up to averaging. Finally, we observe that the filters at convergence have an asymmetric impulse response, and that some of them remain almost analytic.
ISSN:2379-190X
DOI:10.1109/ICASSP.2018.8462015