Adaptive individual background model for speaker verification

Yossi Bar-Yosef, Yuval BistritzInterspeech 2009

TL;DR

This paper introduces a novel speaker verification system that replaces the traditional universal background model (UBM) with an "individual background model" (IBM) tailored to each speaker. This IBM, built using a new adaptation algorithm, significantly improves verification accuracy compared to standard GMM-UBM systems, even when combined with existing score normalization techniques.

What Problem Does This Paper Address?

Speaker verification systems aim to identify individuals based on their unique voice characteristics. A common approach uses Gaussian Mixture Models (GMMs) to model a speaker's voice and compares it against a "universal background model" (UBM) that represents a general population of speakers. While effective, this UBM can be too general, leading to less robust decisions, especially when dealing with variations in recording conditions or speaker-specific vocal traits. Existing methods to improve robustness, like score normalization, add computational overhead. This paper addresses the need for a more speaker-specific background model that enhances accuracy without a significant increase in computational cost.

Methods

The core innovation is the Individual Background Model (IBM). Instead of a single UBM for all speakers, an IBM is created for each target speaker. This IBM is generated by adapting the parameters of the UBM to a set of K "cohort" speaker models that are most similar to the target speaker. The adaptation algorithm is parametric, meaning it operates directly on the GMM parameters (means, covariances, and weights) without requiring the original audio data. This new algorithm leverages the Kullback-Leibler (KL) divergence between Gaussian distributions.

The paper also explores combining the GMM-IBM system with adaptive test normalization (ATnorm), a technique that further refines scores by considering imposter model scores. The authors propose a fast scoring mechanism for the GMM-IBM system that minimizes the computational increase when used with ATnorm.

Datasets

  • NIST-2004-SRE database: Used for evaluating the performance of the proposed GMM-IBM system. This dataset includes multi-lingual speakers and various channel/handset types.
  • SPIDRE corpus: Used in conjunction with a subset of the NIST-2004 data for generating the UBM and cohort models.

Key Insights and Ideas

  • Speaker-specific background models matter: Replacing a generic UBM with an IBM tailored to each speaker significantly boosts verification performance. This suggests that accounting for speaker-specific "imposter" characteristics leads to better discrimination.
  • Parametric adaptation is efficient: The proposed method for creating the IBM relies solely on model parameters, avoiding the need for raw audio or extensive simulations, which keeps the adaptation process computationally feasible.
  • Synergy with score normalization: The GMM-IBM system not only performs well on its own but also benefits from existing score normalization techniques like ATnorm, leading to further improvements.
  • Reduced cohort size: The IBM approach requires a smaller number of cohort models (K=15 in experiments) compared to traditional ATnorm (50-75), suggesting a more focused and potentially more effective background representation.

Key Results

  • The GMM-IBM system consistently outperformed a standard GMM-UBM system across various metrics.
  • In plain settings (without score normalization), the GMM-IBM achieved a 10% improvement in Equal Error Rate (EER) and a 6.4% improvement in minimum Decision Cost Function (minDCF) compared to the baseline GMM-UBM.
  • When combined with ATnorm, the GMM-IBM system showed even greater gains, achieving a 16.5% improvement in EER and a 16.4% improvement in minDCF over the GMM-UBM with ATnorm.
  • The proposed GMM-IBM system with ATnorm achieved comparable performance to a standard GMM-UBM system with ATnorm, but with a slightly lower DCF.
  • The computational overhead for incorporating ATnorm with GMM-IBM was estimated to be around 20% in a typical setting.

Limitations and Future Work

  • Adaptation coefficient (α): The optimal value for the adaptation coefficient 'α' (used in the IBM generation) depends on other parameters like the number of cohorts (K) and the number of adapted parameters. Further investigation into this dependency is suggested.
  • Parameter dependency: The paper notes that the trend suggests increasing the number of adapted parameters (means, variances, weights) requires lower values of the adaptation factor 'α'.
  • Future work: The authors suggest further investigation into the optimal choice of 'α' and how it relates to other system parameters. The paper also implies that exploring different combinations of adapted parameters within the IBM generation could yield further improvements.

Summary updated

BibTeX
@inproceedings{baryosef2009adaptive,
  author = {Yossi Bar-Yosef and Yuval Bistritz},
  title = {{Adaptive individual background model for speaker verification}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure