Supervised domain adaptation for I-vector based speaker recognition

Autoencoder Based Domain Adaptation for Speaker Recognition Under Insufficient Channel Information

“…The dataset 'SRE-1phn' contains audio from only a single telephone number per speaker and use of such a poor phone number diversity hinders the effective estimation of within-speaker variability of in-domain. In this case, the conventional approaches [13], [14] that estimate within-speaker variability from in-domain unlabeled dataset would fail, in spite of perfect speaker label estimation, due to insufficient channel information. Singer [25] also tackled same issue and suggested dataset selection criteria to prevent this situation in advance.…”

“…State-of-the-art techniques from other studies are examined for comparison. For Garcia-Romero's Interpolated approach [13] referred to as system 5 in Table III, the true speaker label is used for ideal case rather than clustering with AHC algorithm. Then, WC and AC from SWB and SRE-1phn are interpolated, as indicated in Table III by "SWB + SRE-1phn".…”

“…In this case, if a dataset domain is matched with operation domain, it is called in-domain dataset; if it is mismatched, it is called out-of-domain dataset. GarciaRomero [7] found that training the UBM and total variability subspace on some types of domains have limited effects on performance improvement. Instead, the performance depends heavily on PLDA parameter estimation.…”

“…Villalba [12] introduced a variational Bayesian approach for adapting PLDA models from out-of-domain datasets to in-domain. Garcia-Romero [13] introduced a clustering approach for unlabeled in-domain datasets that uses well-known Agglomerative Hierarchical Clustering (AHC) to estimate Within-speaker Covariance (WC) and Across-speaker Covariance (AC) with the PLDA model. In this approach, the estimated in-domain WC and AC are interpolated from out-ofdomain WC and AC [7], [13], [14].…”

“…Garcia-Romero [13] introduced a clustering approach for unlabeled in-domain datasets that uses well-known Agglomerative Hierarchical Clustering (AHC) to estimate Within-speaker Covariance (WC) and Across-speaker Covariance (AC) with the PLDA model. In this approach, the estimated in-domain WC and AC are interpolated from out-ofdomain WC and AC [7], [13], [14]. Kanagasundaram [15] introduced another IDVC technique, called Inter-Dataset Variability (IDV), to capture the variability between out-ofdomain and in-domain.…”