Evolution of an Adaptive Unsupervised Speech Controlled System

Abstract

In the preceding chapter a first solution and a reference implementation for combined speaker identification and speaker specific speech recognition were presented and evaluated as complete systems. However, some deficiencies discussed in the preceding section motivate to extend this first solution.

Speaker identification as discussed so far has been treated on an utterance level based on a likelihood criterion. The discussion in this chapter goes beyond that. Long-term speaker identification is investigated to track different speakers across several utterances. The evolution of the self-adaptive system is taken into consideration. Together this results in a more confident guess of the current speaker’s identity. Long-term stability and the detection of unknown speakers can be significantly improved.