Robust speech recognition in noisy environments

The topic of this research project is the development of adaptivenoise reduction filters for automatic speech recognition (ASR). In contrast to systems using a close-talk microphone the capture of speech shall be accomplished by means of a hands-free interface. This in turn requires an ASR which is immune to environmental noise.

The increased distance between the speaker and the microphones in hands-free systems results in an increased level of environmental noise that is captured together with the speech signal. This reduction of the signal-to-noise-ratio (SNR) has a detrimental effect on the recognition rate. A way of making the recognizer more robust is a noise reducing preprocessing of the captured signal (see figure below). This approach leads to statistical signal processing algorithms which are used, e.g., in the area of telecommunication. As those algorithms are not optimized for speech recognition, the fusion of noise reduction and speech recognition is an up-to-date challenging field of research.

The combination of ASR and hands-free speech capture affords the development of highly comfortable user-interfaces. Complex applications like navigation systems in cars or technical installations in rooms can be controlled in a more natural way.

signal graph