Symposium on Machine Learning in Speech and Language Processing (MLSLP)

September 14, 2012
Portland, Oregon, USA

Speaker: Georg Heigold (Google)

Title: Exemplar-Based Speech Recognition in a Rescoring Approach

Abstract:
The acoustic models in state-of-the-art speech recognition systems are based on phones in context that are represented by hidden Markov models. This modeling approach may be insufficient to model long-span acoustic context. Exemplar-based approaches may be an attractive alternative to this conventional approach. In this talk, we will discuss a general rescoring framework based on conditional random fields and show how to do exemplar-based speech recognition within this framework, with the focus on massive, noisy data. Experimental results for the Voice Search and the YouTube tasks are presented and analyzed.