Ponedeljak, 28.3.2022. od 12h

UŽIVO: svečana sala FTN-a
ONLINE: Join Zoom Meeting
Meeting ID: 829 7035 8300
Passcode: 870188


'FAR-FIELD AUTOMATIC SPEECH RECOGNITION'

prof. Reinhold Haeb-Umbach, Padeborn University
https://www.uni-paderborn.de/en/person/242


ABSTRACT: Recognizing speech uttered at a distance from the microphones, so-called far-field or distant speech recognition, is a major success story of recent years and at present still a very active field of research. It is a key component for the success of digital home assistants. In this talk we open the lid of such devices and have a look at the speech technology inside. In particular, we will discuss the various techniques to achieve reliable speech recognition in the presence of an acoustically very challenging environment. We will discuss modern microphone array processing, dereverberation, source separation, beamforming and multi-condition training techniques, all of them contributing to high-performance far-field ASR. It will become evident that in all these fields, deep learning occupies a critical role, but it will also be seen that a clever combination of signal processing and deep learning leads to highly effective solutions with far less computational demands than pure deep learning solutions.
The lecture is an introduction to the field, touching upon different techniques, but not going into too much detail.