A discriminative method for speaker verification using the difference information

Zhenchun Lei, Yingchun Yang, Zhaohui WuInterspeech 2006

TL;DR

This paper introduces a novel discriminative approach for speaker verification by mapping utterances into vectors based on their "difference information" to a codebook. This method, applicable to both VQ and GMM frameworks, offers a more direct measure of utterance-codebook fitness than existing derivative-based methods and achieves superior performance on the YOHO database.

What Problem Does This Paper Address?

Speaker verification, the task of confirming a person's identity based on their voice, often relies on comparing an unknown utterance to a stored voice model. Traditional methods like Gaussian Mixture Models (GMMs) are generative, meaning they model how speech is produced. While effective, they can be less discriminative than methods that directly learn to distinguish between speakers. Support Vector Machines (SVMs) are a powerful discriminative tool, but applying them to variable-length utterances requires clever "kernel" functions that map entire utterances into a fixed-size vector space. Existing utterance-based kernels, like the Fisher kernel, often rely on derivative operations of generative models, which can be complex or even infeasible to combine with certain models like Vector Quantization (VQ). This paper aims to bridge this gap by proposing a new, simpler discriminative mapping for utterances that can be readily integrated with both VQ and GMMs, leading to improved speaker verification performance.

Methods

The core innovation lies in how utterances are transformed into vectors for SVM input. Instead of relying on derivative-based features, the authors propose using "difference information."

  • VQ-based Method: An utterance is treated as a sequence of acoustic feature vectors. For each frame (feature vector), the method finds the closest vector in a pre-trained speaker-specific codebook. A matrix is then constructed where each element represents the difference between the frame's feature vector and its nearest codebook vector. This matrix is summed across all frames to create a single utterance-level vector. This vector implicitly captures how well the utterance "fits" the codebook.
  • GMM-based Method: This method leverages the components of a GMM. For each frame, it identifies the GMM component that yields the maximum likelihood. Similar to the VQ approach, a matrix is formed based on the difference between the frame's features and the mean of this most likely component. Covariance matrices of the GMM components are used for normalization, providing a more robust representation.

Both methods then use these generated utterance vectors as input to an SVM classifier. The paper also explores linear, polynomial, and Radial Basis Function (RBF) kernels for the SVM.

Datasets

  • YOHO Database: This is the primary dataset used for experiments. It consists of 138 speakers, each with enrollment and verification sessions. The features extracted are 12th order Mel-Frequency Cepstral Coefficients (MFCCs) with deltas, resulting in 32-dimensional feature vectors per frame.

Key Insights and Ideas

  • "Difference Information" as a Discriminative Feature: The paper demonstrates that directly measuring the difference between utterance features and codebook/GMM components is a powerful discriminative signal for speaker verification, bypassing the need for complex derivative calculations.
  • Unified Framework for VQ and GMM: The proposed mapping elegantly allows for the integration of discriminative SVMs with both VQ and GMMs, two common paradigms in speaker recognition. This is particularly significant as it enables combining SVMs with VQ, which was previously challenging with derivative-based methods.
  • Codebook/GMM Component Fitness: The resulting vectors directly reflect how well an utterance aligns with the learned speaker models (codebook or GMM components), providing an intuitive measure of speaker identity.
  • Improved Normalization: The GMM-based method's use of covariance matrix normalization offers a more sophisticated way to handle feature variations compared to simpler approaches.

Key Results

The experiments on the YOHO database show significant improvements:

  • GMM/SVM vs. GMM: The GMM/SVM approach consistently outperforms the standalone GMM across various GMM orders (8 to 64 components), achieving substantially lower Equal Error Rates (EERs). For instance, with 32-order GMMs, GMM/SVM reduced EER from 6.0% to 1.5% (without cohort normalization).
  • VQ/SVM vs. GMM/SVM: The GMM/SVM method also outperforms the VQ/SVM method, especially after cohort normalization. For 32-order models, GMM/SVM+cohort achieved an EER of 0.17%, while VQ/SVM+cohort was 0.5%.
  • Cohort Normalization: Applying cohort normalization further boosts performance for both VQ/SVM and GMM/SVM, highlighting its importance in robust speaker verification.
  • Kernel Function Performance: While linear kernels performed well, polynomial and RBF kernels did not show a clear advantage over linear kernels in this specific setup.

Limitations and Future Work

  • Codebook/GMM Dependence: The performance of the proposed methods is inherently dependent on the quality and representativeness of the codebook (for VQ) or the trained GMM.
  • Future Research Directions: The authors suggest exploring higher-order deviations and further optimizing the combination of GMMs with their proposed method to fully leverage the nature of GMMs. They also hint at investigating more complex kernel functions.

Summary updated

BibTeX
@inproceedings{lei2006discriminative,
  author = {Zhenchun Lei and Yingchun Yang and Zhaohui Wu},
  title = {{A discriminative method for speaker verification using the difference information}},
  booktitle = {Interspeech},
  year = {2006},
}

Loading related content...

Figure