Speaker verification using acoustic factor analysis with phonetic content compensation in limited and degraded test conditions
Akhil Babu Manam, Tummala Sai Revanth, Rohan Kumar Das, S. R. Mahadeva Prasanna · 2016
This work explores speaker verification (SV) under limited and degraded test conditions from the perspective of practical systems. An i-vector based SV system using maximum likelihood - acoustic factor analysis (ML-AFA) technique for modeling and short utterance variance normalization (SUVN) technique for enhancing the phonetic content of short utterance i-vectors is developed. As shorter utterances in text-independent scenarios may contain any phonetic content, the i-vectors for test utterances also vary significantly. SUVN helps in compensating the variation in i-vectors due to phonetic content and improves the performance in limited test data conditions. Whereas ML-AFA technique helps in reducing the effect of noise. Utilizing both these frameworks, a SV system under limited and degraded test conditions is proposed. The proposed system is found to outperform the baseline system by a significant margin. Finally, the proposed and the baseline systems are fused at scores level that further boosts the performance.