ViVoLAB Speaker Diarization System for the DIHARD 2019 Challenge

Ignacio Viñals, Pablo Gimeno, Alfonso Ortega Giménez, Antonio Miguel, Eduardo LleidaInterspeech 2019 · DOI

Abstract

This paper presents the ViVoLAB speaker diarization system designed for the DIHARD 2019 Challenge, which focuses on robust performance in adverse acoustic environments. The system evaluates two primary embedding types—i-vectors and x-vectors—across two distinct clustering paradigms: a Variational Bayes (VB) PLDA approach and a novel tree-based sequential clustering method. Our findings highlight the continued relevance of traditional i-vectors in noisy scenarios and the trade-offs between the robustness of VB-based methods and the peak performance of sequential search strategies.

Introduction

Speaker diarization, the task of determining "who spoke when," remains a significant challenge in unconstrained acoustic environments characterized by noise, reverberation, and overlapping speech. The DIHARD 2019 Challenge provides a rigorous benchmark for these systems across diverse domains. This work specifically investigates the clustering stage of the diarization pipeline, comparing established statistical inference methods with a novel sequential path-finding approach.

Methodology

The ViVoLAB system follows a standard pipeline consisting of feature extraction, clustering, and resegmentation.

Feature Extraction

Two types of speaker embeddings are utilized:

  1. i-vectors: A traditional approach based on Gaussian Mixture Models (GMMs).
  2. x-vectors: A modern approach utilizing Deep Neural Networks (DNNs) to capture speaker characteristics.

Clustering Strategies

The core of our contribution lies in the comparison of two sophisticated clustering methods:

  • Variational Bayes (VB) PLDA: This method treats clustering as a statistical inference problem using a Fully Bayesian Probabilistic Linear Discriminant Analysis (PLDA) model. It provides robustness by integrating over the uncertainty of speaker assignments.
  • Tree-Based Sequential Clustering: This novel approach frames diarization as a path-finding problem within a decision tree. Each branch represents a potential speaker assignment for an audio segment. To manage the exponential growth of the tree, the M-algorithm is employed as a pruned search strategy to maintain only the most probable paths.
Graphical representation of the sequential clustering model used in the tree-based approach.

Resegmentation

To refine the temporal boundaries of speech segments, we apply a resegmentation step using Hidden Markov Models (HMMs) with eigenvoice priors. This stage consistently improves the final Diarization Error Rate (DER).

Experimental Setup and Datasets

The system was trained on a diverse data pool to ensure broad coverage of acoustic conditions. This pool included the MGB 2015 broadcast data (1,000+ hours), the AMI and ICSI meeting datasets, and the large-scale VoxCeleb I & II speaker identification datasets. Evaluation was performed on the DIHARD 2019 development and evaluation sets.

Results and Discussion

The performance of the system is measured using the Diarization Error Rate (DER).

Diarization Error Rate (DER) results for Track 1 (clean VAD) and Track 2 (system VAD) across different embedding and clustering configurations.

Our results indicate that while x-vectors are often superior in clean conditions, i-vectors demonstrate remarkable resilience in the adverse environments of the DIHARD challenge. In Track 1 (single-channel, clean Voice Activity Detection), the tree-based system using i-vectors achieved the best performance with a 25.67% DER. In Track 2 (single-channel, system VAD), the VB-PLDA system with x-vectors proved most effective, yielding a 38.29% DER.

A significant observation was the "domain mismatch" between development and evaluation data. While the tree-based approach achieved high peak performance, it was more sensitive to hyperparameter configuration compared to the more robust VB-PLDA approach.

Conclusion

The ViVoLAB system for DIHARD 2019 demonstrates that traditional i-vectors remain a competitive choice for diarization in noisy, real-world scenarios. While novel sequential clustering methods show promise for achieving state-of-the-art results, the robustness of Bayesian inference methods like VB-PLDA is crucial for handling domain shifts. Future work will focus on unsupervised domain adaptation to further bridge the performance gap in unseen acoustic environments.

Summary updated

BibTeX
@inproceedings{vinals2019vivolab,
  author = {Ignacio Viñals and Pablo Gimeno and Alfonso Ortega Giménez and Antonio Miguel and Eduardo Lleida},
  title = {{ViVoLAB Speaker Diarization System for the DIHARD 2019 Challenge}},
  booktitle = {Interspeech},
  year = {2019},
  doi = {10.21437/interspeech.2019-2462},
}
Figure