SNR-dependent mixture of PLDA for noise robust speaker verification

Man‐Wai Mak2014 · DOI

TL;DR

This paper tackles the challenge of speaker verification in noisy environments by proposing a novel "SNR-dependent mixture of PLDA" model. This approach allows the system to adapt its verification strategy based on the estimated Signal-to-Noise Ratio (SNR) of the audio, leading to more robust performance compared to traditional methods that use a single model for all noise levels.

What Problem Does This Paper Address?

Speaker verification systems aim to confirm a person's identity based on their voice. A major hurdle in deploying these systems in real-world scenarios is the presence of background noise and varying acoustic conditions, which can significantly degrade performance. Existing state-of-the-art methods, like the i-vector/PLDA framework, often rely on a single Probabilistic Linear Discriminant Analysis (PLDA) model to handle all test utterances, regardless of their noise levels. This paper argues that a single model might not be optimal for the wide range of SNRs encountered in real-world audio, leading to a performance bottleneck. The authors propose a solution that dynamically adjusts the verification process based on the estimated SNR of the audio.

Methods

The core innovation is the SNR-dependent mixture of PLDA (SD-mPLDA) model. Instead of training a single PLDA model, they train multiple PLDA models, each optimized for a specific range of SNRs.

Here's a breakdown of the key components:

  • i-vectors: The system first extracts i-vectors, which are low-dimensional representations of entire speech utterances, capturing both speaker and channel variability.
  • PLDA (Probabilistic Linear Discriminant Analysis): This is a standard backend technique for speaker verification that models the distribution of i-vectors for speakers and non-speakers.
  • Mixture of PLDA: The paper extends the standard PLDA by creating a mixture of several PLDA models.
  • SNR-Dependent Training: Crucially, these PLDA models are trained simultaneously using an Expectation-Maximization (EM) algorithm. The contribution of each training i-vector to a specific PLDA model is determined by the posterior probability of the utterance's SNR. This means the models learn to specialize in different noise conditions.
  • Soft-Decision Scoring: During verification, the system doesn't hard-assign an i-vector to a single PLDA model. Instead, it computes marginal likelihoods from all the trained PLDA models and combines them probabilistically. This combination is weighted by the posterior probabilities of the test utterance's and the target speaker's utterance SNRs. This "soft" approach allows for smoother transitions between models and better handling of ambiguous SNR cases.

Datasets

  • Training Data: Speech files from NIST 2005-2010 SREs were used for training gender-dependent Universal Background Models (UBMs), total variability matrices, LDA-WCCN, and PLDA models. For generating noisy data, clean training files were augmented with noise from the PRISM dataset at SNRs of 6dB and 15dB.
  • Evaluation Data: The core set of NIST 2012 Speaker Recognition Evaluation (SRE) was used. Specifically, the paper focused on "common condition 4" (CC4) and "common condition 5" (CC5), which include noisy test segments.

Key Insights and Ideas

  • SNR is a critical factor: The paper strongly suggests that treating all noise levels with a single model is suboptimal. Different SNRs require different modeling approaches for optimal performance.
  • Soft-decision is better than hard-decision: By probabilistically combining the outputs of multiple SNR-specific PLDA models, the system achieves better performance than simply selecting one model based on a hard threshold of the SNR. This "soft" approach is more robust to uncertainty in SNR estimation.
  • Simultaneous training is key: Training the mixture of PLDA models simultaneously via EM, with SNR probabilities influencing the contribution of training data, allows for better coordination between the specialized models.
  • Eliminates score normalization: A significant advantage of the proposed soft-decision mixture approach is that it naturally produces scores within a comparable range, eliminating the need for complex score normalization techniques often required by hard-decision mixture models.

Key Results

The proposed Soft-Decision Mixture of PLDA (SD-mPLDA) consistently outperformed both a baseline multi-condition PLDA (trained on pooled clean and noisy data) and a Hard-Decision Mixture of PLDA (HD-mPLDA).

  • In CC4 (female speakers), SD-mPLDA achieved an Equal Error Rate (EER) of 2.88% and a minimum Detection Cost Function (minDCF) of 0.329, compared to 3.49% EER and 0.308 minDCF for the baseline PLDA.
  • In CC5 (female speakers), SD-mPLDA achieved 2.80% EER and 0.287 minDCF, outperforming the baseline's 2.79% EER and 0.290 minDCF.
  • Similar improvements were observed for male speakers.
  • The paper also highlights that SD-mPLDA is less sensitive to the choice of decision thresholds compared to HD-mPLDA, making it more practical.

Limitations and Future Work

  • Limited noise types: The authors suggest extending the work by including a wider variety of noise types (e.g., traffic, office, restaurant) in the training data for the mixture models.
  • Reverberation: They propose applying the mixture model concept to handle reverberant speech, where each reverberant environment could be modeled by a separate PLDA model.
  • SNR estimation: While the paper uses SNR for model selection, the accuracy of the SNR estimation itself can impact performance. Future work could explore more robust SNR estimation techniques.

Summary updated

Abstract

15th Annual Conference of the International Speech Communication Association, INTERSPEECH 2014, Singapore, September 14-18, 2014
BibTeX
@misc{mak2014snr,
  author = {Man‐Wai Mak},
  title = {{SNR-dependent mixture of PLDA for noise robust speaker verification}},
  year = {2014},
  doi = {10.21437/interspeech.2014-421},
}
Figure