Language identification of individual words with joint sequence models

Abstract

Abstract Within a multilingual automatic speech recognition (ASR) sys-tem, knowledge of the language of origin of unknown wordscan improve pronunciation modelling accuracy. This is of par-ticular importance for ASR systems required to deal with code-switched speech or proper names of foreign origin. For wordsthat occur in the language model, but do not occur in the pro-nunciation lexicon, text-based language identification (T-LID)of a single word in isolation may be required. This is a chal-lenging task, especially for short words. We motivate for theimportance of accurate T-LID in speech processing systems andintroduce a novel way of applying Joint Sequence Models to theT-LID task. We obtain competitive results on a real-world 4-language task: for our best JSM system, an F-measure of 97:2%is obtained, compared to a F-measure of 95:2% obtained with astate-of-the-art Support Vector Machine (SVM).Index Terms: text-based language identification, joint se-quence models, multilingual speech recognition