Mixture of PLDA for Noise Robust I-Vector Speaker Verification

Abstract

In real-world environments, noisy utterances with variable noise levels are recorded and then converted to i-vectors for cosine distance or PLDA scoring. This paper investigates the effect of noise-level variability on i-vectors. It demonstrates that noise-level variability causes the i-vectors to shift, causing the noise contaminated i-vectors to form clusters in the ivector space. It also demonstrates that optimal subspaces for discriminating speakers are noise-level dependent. Based on these observations, this paper proposes using signal-to-noise ratio (SNR) of utterances as guidance for training mixture of PLDA models. To maximize the coordination among the PLDA models, mixtures of PLDA models are trained simultaneously via an EM algorithm using the utterances contaminated with noise at various levels. For scoring, given a test i-vector, the marginal likelihoods from individual PLDA models are linearly combined by the posterior probabilities of the test utterance's SNR. Verification scores are the ratio of the marginal likelihoods. Results based on NIST 2012 SRE suggest that the SNR-dependent mixture of PLDA is not only suitable for the situations where the test utterances exhibit a wide range of SNR, but also beneficial for the test utterances with unknown SNR distribution. Supplementary materials containing full derivations of the EM algorithms and scoring functions can be found in http://bioinfo.eie.polyu.edu.hk/mPLDA/SuppMaterials.pdf.

Channel Variability Synthesis in i-vector Speaker Recognition

“…In [8], a mixture of channel-dependent PLDA models are trained to account for the channel conditions of each test utterance presented at the detection phase. Another mixture of PLDA Models is presented in [9], where each one is trained with different levels of noise and used according to the signal-to-noise ratio of the test utterance. The work in [10] assessed one channel features-domain noise compensation combined with multi-condition training.…”

Twin Model G-PLDA for Duration Mismatch Compensation in Text-Independent Speaker Verification

“…In addition to comparing these histograms, a second measure of differences between the distribution of i-vectors from long and short utterances based on the Partition Coefficient [15,16] is employed in this paper. The partition coefficient is an index that indicates the clustering tendency in a dataset and lies in the range [1/݇, 1] , where ݇ is the number of clusters.…”

A Generalization of PLDA for Joint Modeling of Speaker Identity and Multiple Nuisance Conditions

“…The approach was shown to outperform standard PLDA with pooled training data when each class in the training data is seen under both considered conditions, frontal and profile, in a face recognition task. A similar approach is proposed by [12]; but in this case, the mixture component is not given during training but rather is dependent on a continuous metadata value. The approach is tested by adding noise to the training data at different signal-to-noise (SNR) levels, resulting in gains compared to pooling all the data to train a single PLDA model.…”