Output details
11 - Computer Science and Informatics
University of Edinburgh
Speech Recognition Using Augmented Conditional Random Fields
<22> Originality: New approach to discriminative acoustic modelling for speech recognition based on augmented conditional random fields (ACRFs), addressing conditional independence limitations of HMMs, and the ability to model acoustic context explicitly.
Significance: This was one of the first acoustic modelling approaches to use high-dimensional sparse spaces to improve discrimination. Phone error rate on TIMIT was significantly reduced compared with a state-of-the-art HMM/GMM system.
Rigour: A probabilistic acoustic model was carefully developed, in conjunction with an efficient learning algorithm. The approach was evaluated using the standard TIMIT corpus, and has been used as a benchmark for comparison (e.g. Hinton's group).