The I2R’s Submission to VOiCES Distance Speaker Recognition Challenge 2019

Hanwu Sun, Kah Kuan Teh, Ivan Kukanov, Huy Dat TranInterspeech 2019

TL;DR

This paper details I2R's submission to the VOiCES distance speaker recognition challenge, focusing on fusing multiple x-vector and i-vector systems. Their key innovation lies in a robust de-reverberation frontend and a novel "Blend" fusion strategy, achieving a competitive 2% EER on the development dataset.

What Problem Does This Paper Address?

The paper tackles the challenging problem of speaker recognition in noisy, reverberant environments, specifically for distant speech captured by single microphones. This is a significant hurdle for real-world applications like voice assistants or security systems, where clean, close-talking audio is not always available. The VOiCES challenge aims to push the boundaries of speaker recognition under these difficult conditions, requiring systems that can effectively handle audio distortions like reverberation and background noise.

Methods

The I2R team employed a multi-pronged approach, combining established techniques with novel strategies:

  • Frontend De-reverberation: A Weighted Prediction Error (WPE) method was used to mitigate the effects of reverberation. This technique leverages linear prediction to estimate and subtract the reverberant components of the audio signal.
  • Feature Extraction: They utilized both x-vectors (DNN-based embeddings) and i-vectors (probabilistic linear discriminative models) as speaker representations. Specifically, they trained four separate systems: two x-vector systems (from SITW and Voxceleb datasets) and two i-vector systems (also from SITW and Voxceleb).
  • Backend Modeling:
    • LDA: Linear Discriminant Analysis (LDA) was used to reduce the dimensionality of the x-vector and i-vector embeddings.
  • PLDA: Probabilistic Linear Discriminant Analysis (PLDA) was employed for speaker modeling. They explored both "out-of-domain" PLDA (trained on general datasets like Voxceleb) and "in-domain" PLDA (trained on a subset of the VOiCES development data, specifically speaker utterances not present in the training models).
  • Score Normalization: Symmetric score normalization (s-norm) with adaptive cohort selection was applied to calibrate the scores from individual systems.
  • Fusion Strategies:
    • Linear Fusion: A standard linear fusion approach using the Bosaris toolkit was implemented.
  • Blend-based Non-linear Fusion: A novel "Blend" model, inspired by deep super learning, was developed. This method uses a multi-layer ensemble of classifiers (XGBoost, random forest, gradient boosting) trained on the outputs of the individual subsystems.

Datasets

  • Training Data: Voxceleb 1 & 2, MUSAN (noise and music), and RIRS noise datasets were used for training the x-vector and i-vector models. Voxceleb 1 & 2 were also used for training the out-of-domain PLDA backend.
  • Development Data: The VOiCES Dev dataset was used for tuning parameters and training the "in-domain" PLDA model. A subset of this data, consisting of 93 speaker utterances not present in the training models, was specifically extracted for in-domain PLDA design.
  • Evaluation Data: The VOiCES evaluation dataset was used for final system performance assessment.

Key Insights and Ideas

  • De-reverberation is Crucial: The WPE de-reverberation frontend consistently improved performance across all subsystems, demonstrating its significant impact on distant speech recognition. The improvement was more pronounced on the evaluation dataset, suggesting it had stronger reverberation.
  • DNN Embeddings Outperform i-vectors: X-vector systems, leveraging DNN embeddings, showed a substantial performance advantage (around 50% relative improvement) over i-vector systems, aligning with recent trends in speaker recognition.
  • In-Domain PLDA Boosts Performance: Using an "in-domain" PLDA model, trained on speaker data from the target challenge dataset, significantly improved EER and DCF compared to out-of-domain PLDA.
  • Fusion is Key: Combining multiple subsystems through fusion strategies (both linear and non-linear) was essential for achieving the best results. The novel "Blend" fusion model showed promising results, outperforming linear fusion in some aspects.
  • Complementary Subsystems: The four different x/i-vector subsystems exhibited complementary effects, meaning their combined performance was better than any individual system, highlighting the benefit of diversity in feature extraction.

Key Results

  • The de-reverberation approach achieved relative improvements of 5-10% on EER and DCF for individual subsystems on the Dev set, and over 10% on the final fusion system. On the evaluation dataset, improvements exceeded 15%.
  • The final fused system achieved an EER of approximately 2% and a minDCF of 0.240 on the Development Dataset.
  • The "Blend" based non-linear fusion model (System02) achieved an EER of 1.76% and minDCF of 0.2264 on the development dataset.
  • The best performing submission (System03, a linear fusion of System01 and System02) achieved an EER of 1.53% on the development dataset.
  • On the evaluation dataset, the best system (System01, linear fusion) achieved an EER of 6.28% and minDCF of 0.4975.

Limitations and Future Work

  • Dataset Differences: The paper notes that the evaluation dataset was more challenging than the development dataset due to higher noise levels, more serious reverberation, and shorter utterances. This led to significantly worse performance on the evaluation set compared to the development set.
  • Overfitting Concerns: The authors observed that further linear fusion of already combined systems (System03) led to worse performance on the evaluation set, suggesting potential overfitting.
  • Future Work: While not explicitly detailed, the consistent performance improvements from de-reverberation and in-domain modeling suggest these areas are crucial for future research. Exploring more sophisticated non-linear fusion techniques and robust feature extraction methods for highly reverberant and noisy conditions would be logical next steps.

Summary updated

BibTeX
@inproceedings{sun2019i,
  author = {Hanwu Sun and Kah Kuan Teh and Ivan Kukanov and Huy Dat Tran},
  title = {{The I2R’s Submission to VOiCES Distance Speaker Recognition Challenge 2019}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure