Language Recognition on Albayzin 2010 LRE using PLLR features

Resumen

Phone Log-Likelihood Ratios (PLLR) have been recently proposed as alternative features to MFCC-SDC for iVector Spoken Language Recognition (SLR). In this paper, PLLR features are first described, and then further evidence of their usefulness for SLR tasks is provided, with a new set of experiments on the Albayzin 2010 LRE dataset, which features wide-band multi speaker TV broadcast speech on six languages: Basque, Catalan, Galician, Spanish, Portuguese and English. iVector systems built using PLLR features, computed by means of three open-source phone decoders, achieved significant relative improvements with regard to the phonotactic and MFCC-SDC iVector systems in both clean and noisy speech conditions. Fusions of PLLR systems with the phonotactic and/or the MFCC-SDC iVector systems led to improved performance, revealing that PLLR features provide complementary information in both cases.