Article details

Research area
Speech enhancement

A System for Enhanced Speech Recognition Series: Signals and Communication Technology


Tobias Herbig, Franz Gerl, Wolfgang Minker

Self-learning speaker identification


Current speech recognition systems suffer from variation of voice 
characteristics between speakers as they are usually based on speaker 
independent speech models. In order to resolve this issue, adaptation 
methods have been developed in many state-of-the-art systems. However, 
information acquired over time is still lost whenever another speaker intermittently uses the recognition system. This work therefore develops an integrated approach for speech and speaker recognition in order to improve the self-learning opportunities of the system. A speaker adaptation scheme
is introduced. It is suited for fast short-term and detailed long-term adaptation. These adaptation profiles are then used for an efficient
speaker recognition system. The speaker identification enables the speaker adaptation to track different speakers which results in an optimal long-term adaptation.

Read/download now