This paper explores the significance of an ensemble of boosted Support Vector Machine (SVM) classifiers in the i-vector framework for speaker verification (SV) in noisy environments. Prior work in this field have established the significance of supervector-based approaches and more specifically the i-vector extraction paradigm for robust SV. However, in highly degraded environments, SVMs trained using i-vectors are susceptible to misclassifications. For enhanced classification accuracy, we explore the impact of multiple SVM classifiers trained by adaptive boosting. To mitigate the effect of statistical mismatches due to difference in utterance lengths and data imbalance caused by a disproportionate ratio of target speaker and impostor utterances, we propose a novel combination scheme of the adaptive boosting algorithm with a data generation technique using partitioned utterances. All experiments are conducted on the NIST-SRE-2003 database under mismatched conditions with training utterances degraded by 4 types of additive noises (car, factory, pink and white) collected from the NOISEX-92 database, at 0 dB and 5 dB SNRs. Results indicate that the proposed method significantly outperforms the baseline i-vector SVM based SV systems across all noisy environments.
Bibliographic reference. Sarkar, Sourjya / Rao, K. Sreenivasa (2014): "A novel boosting algorithm for improved i-vector based speaker verification in noisy environments", In INTERSPEECH-2014, 671-675.