A discriminative performance metric for GMM-UBM speaker identification

Omid Dehzangi, Bin Ma, Engsiong Chng, Haizhou Li

Introduction

In the field of speaker identification (SI), the objective is to accurately assign an unknown voice sample to one of $N$ enrolled speakers. The prevailing industry standard for this task is the Gaussian Mixture Model-Universal Background Model (GMM-UBM) framework. Typically, individual speaker models are derived from a UBM using Maximum A Posteriori (MAP) estimation.

While effective, MAP is fundamentally a generative approach. It optimizes how well a model represents a specific speaker's acoustic data in isolation, often failing to explicitly account for interspeaker variability. This can result in overlapping models in the acoustic space, leading to misidentifications. This paper proposes a Discriminative Performance Metric (DPM) to address this gap by directly optimizing model parameters for classification performance.

Methodology: The DPM Approach

The core innovation lies in a learning algorithm that treats each Gaussian component within a speaker's GMM as a discriminative unit. Unlike standard MAP, which focuses on the statistical fit of the data, the DPM approach re-estimates Gaussian mixture weights ($w_i$) to minimize frame-level classification errors.

The DPM optimization framework, illustrating the iterative adjustment of mixture weights based on loss calculated from a comparator.

The optimization process follows these steps:

  1. One-vs-Rest Optimization: For each speaker model, the target speaker is treated as the "positive" class, while all other speakers in the cohort are treated as the "negative" class.
  2. Operating Point Selection: Drawing from Receiver Operating Characteristic (ROC) analysis, the algorithm calculates a measure ($\Phi$) for each training frame to determine the weight required for correct classification.
  3. Weight Tuning: The algorithm identifies an optimal threshold that maximizes frame-level accuracy on training data. This threshold is then assigned as the new weight for the corresponding Gaussian component.
  4. Iterative Refinement: This process is repeated across multiple iterations to sharpen the decision boundaries between speakers.

Experimental Setup

The proposed method was evaluated using the 2001 NIST Speaker Recognition Evaluation (SRE) corpus. The development set included 60 speakers from the Switchboard-Cellular corpora, while the evaluation set comprised 174 target speakers. Performance was assessed across three test segment lengths: 10 seconds, 20 seconds, and "whole" utterances.

Key Results

The DPM method demonstrated consistent and significant improvements over the baseline MAP system. On average, the "MAP+DPM" approach achieved a 13.2% relative improvement in identification accuracy.

Identification Error Rate (IER) comparison between MAP and MAP+DPM across different Top-N ranks and segment lengths.

As shown in the results table, the error reduction is evident across all Top-N categories. For instance, in Top-1 identification for whole segments, the error rate dropped from 23.57% with MAP to 20.46% with MAP+DPM. Notably, these gains were robust regardless of the number of Gaussian mixtures used, with consistent performance boosts observed for models ranging from 32 to 512 mixtures.

Conclusion and Future Work

By shifting the focus from generative modeling to discriminative weight tuning, the DPM approach effectively captures the nuances that distinguish speakers. While the current work focuses on mixture weights, future research could extend this discriminative logic to Gaussian means and covariances. Additionally, applying this metric to more modern, noisy datasets and exploring its utility in speaker verification tasks remain promising avenues for further development.

Summary updated

BibTeX
@inproceedings{dehzangi2010discriminative,
  author = {Omid Dehzangi and Bin Ma and Engsiong Chng and Haizhou Li},
  title = {{A discriminative performance metric for GMM-UBM speaker identification}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-158},
}
Figure