This paper presents a real-time method for Spoken Language Identification based on the entropy of the posterior probabilities of language specific phoneme recognisers. Entropy based discriminant functions computed on short speech segments are used to compare the model fit to a specific set of observations and language identification is performed as a model selection task. The experiments, performed on a closed set of four Germanic languages on the SpeechDat telephone speech recordings, give 95% accuracy of the method for 10 seconds long speech utterances and 99% accuracy for 20 seconds long utterances.
tmh_import_11_12_14. QC 20111216