Modeling code-Switching speech on under-resourced languages for language identification

This paper presents an integration of phonotactic information to perform language identification (LID) in a mixed-language speech. A single-pass front-end recognition system is employed to convert the spoken utterances into a statistical occurrence of phone sequences. To process such phone sequences, a hidden Markov model (HMM) is utilized to build robust acoustic models that can handle multiple languages within an utterance. A supervised Support Vector Machine (SVM) learns the language transition of the phonotactic information given the recognized phone sequences. The back-end SVM-based decision classifies language identity given the likelihood scores phone occurrences. The experiments are conducted on commonly mixed-language Northern Sotho and English speech utterances. We evaluate the system measuring the performance of the phone recognition and LID portions separately. We obtained a phone error rate of 15.7% when a data-driven phoneme mapping approach is modeled with 16 Gaussian mixtures per state. However, the proposed integrated LID system has achieved a considerable performance with an acceptable LID accuracy of 85.0% and average of 81% on code-switched speech and monolingual speech segments respectively. Index Terms Code-switching speech, under-resourced languages, phonotactic information, acoustic models, language model