Unsupervised Language Model Adaptation using Utterance-based Web Search for Clinical Speech Recognition

Robert Herms, Daniel Richter, Maximilian Eibl, Marc Ritter
2015 Conference and Labs of the Evaluation Forum  
In this working notes paper we present our methodology in clinical speech recognition for the Task 1.a.1 of the CLEF eHealth Evaluation Lab 2015. The goal of this task is to minimize the worddetection errors. Our approach is based on the assumption that each spoken clinical document has its own context. Hence, the recognition system is adapted for each document separately. The proposed method performs two-pass decoding whereas the first transcript is processed to queries which are used for
more » ... eving web resources as adaptation data to build a document-specific dictionary and language model. The second pass decodes the same document using the adapted dictionary and language model. The experimental results show a reduction of the insertion errors in comparison to the baseline system, but no improvement of the overall incorrectness percentage across all spoken documents.
dblp:conf/clef/HermsRER15 fatcat:g2qjl2n76jarney3s7jgfgnkvu