Application of Subspace Gaussian Mixture Models in Contrastive Acoustic Scenarios

This paper describes experimental results of applying Subspace Gaussian Mixture Models (SGMMs) in two completely diverse acoustic scenarios: (a) for Large Vocabulary Continuous Speech Recognition (LVCSR) task over (well-resourced) English meeting data and, (b) for acoustic modeling of underresourced Afrikaans telephone data. In both cases, the performance of SGMM models is compared with a conventional context-dependent HMM/GMM approach exploiting the same kind of information available during the training. LVCSR systems are evaluated on standard NIST Rich Transcription dataset. For under-resourced Afrikaans, SGMM and HMM/GMM acoustic systems are additionally compared to KL-HMM and multilingual Tandem techniques boosted using supplemental out-of-domain data. Experimental results clearly show that the SGMMapproach (having considerably less model parameters) outperforms conventional HMM/GMM system in both scenarios and for all examined training conditions. In case of under-resourced scenario, the SGMM trained only using indomain data is superior to other tested approaches boosted by data from other domain.