Scalable I-vector concatenation for PLDA based language identification system

Language identification systems combining i-vectors estimated from different acoustic feature spaces have recently been shown to be superior to i-vector systems based on a single acoustic feature space. Specifically, i-vectors estimated using MFCC and PLP front-ends were concatenated prior to using LDA to obtain a combined i-vector. In this work, we investigate the scalability of this i-vector concatenation based framework to incorporate a larger number of front-ends, in particular, phonotactic front-ends. A modification to the framework is also proposed in order to improve this scalability. The proposed framework is evaluated on the 30, 10 and 3 seconds test set of NIST 2007 LRE database.