Spherical Discriminant Analysis in Semi-supervised Speaker Clustering

Hao Tang, Hao Tang 0001, Stephen M. Chu, Thomas S. HuangNAACL 2009

Abstract

This paper presents Spherical Discriminant Analysis (SDA), a semi-supervised dimensionality reduction technique tailored for speaker clustering. Recognizing that speaker-specific data in high-dimensional Gaussian Mixture Model (GMM) supervector space is primarily directional, SDA optimizes for cosine similarity rather than Euclidean distance. Evaluated on the GALE broadcast news database, SDA significantly outperforms traditional methods like PCA and LDA, particularly as the number of speakers increases, demonstrating its robustness and generalizability to unseen speakers.

Introduction

Speaker clustering, the process of grouping audio segments by speaker identity, is a cornerstone of speaker diarization systems. Traditional approaches often rely on unsupervised methods, ignoring the wealth of available labeled speech data from unrelated tasks. Furthermore, standard clustering techniques typically operate in Euclidean space, which may not be optimal for high-dimensional speech representations where directional information is paramount.

This paper introduces Spherical Discriminant Analysis (SDA), a novel semi-supervised dimensionality reduction technique. By recognizing that speaker-specific patterns in GMM supervector space are more "directional" than "positional," SDA optimizes for cosine similarity, leading to significant improvements in clustering accuracy.

Methodology

The proposed framework leverages a semi-supervised pipeline where an independent, labeled training set is used to learn a discriminative subspace for the target clustering task.

Utterance Representation

Each speech segment is represented as a GMM mean supervector. By adapting a Universal Background Model (UBM) to a specific utterance, the resulting mean shifts are concatenated into a high-dimensional vector (e.g., 2560 dimensions). These vectors capture the unique acoustic characteristics of a speaker relative to a general population model.

Spherical Discriminant Analysis (SDA)

The core innovation, SDA, moves beyond the Euclidean assumptions of Linear Discriminant Analysis (LDA). While LDA maximizes the ratio of between-class to within-class variance in Euclidean terms, SDA seeks a linear transformation that maximizes cosine similarity for segments from the same speaker while minimizing it for different speakers.

Given the non-linear nature of spherical projections, the authors utilize a graph embedding framework. The optimization problem is solved using a steepest descent algorithm, effectively finding a lower-dimensional subspace where the most discriminative directional features are preserved and the hypersphere is effectively "flattened" for easier clustering.

Experimental Setup

The researchers evaluated SDA using the GALE database, which consists of broadcast news speech.

  • Training Set: 498 speakers (18,327 utterances) used to train the SDA transformation and the UBM.
  • Test Set: 630 speakers (19,024 utterances) with no speaker overlap with the training set, ensuring the generalizability of the learned features to new individuals.

Results and Analysis

The performance of SDA was compared against raw features (Orig) and standard dimensionality reduction techniques like PCA and LDA across varying numbers of speakers (5 to 50). The metric used was average clustering accuracy across 100 trials.

Clustering accuracy comparison across different subspaces and metrics.

As shown in the results above, SDA consistently outperforms all baselines. Notably, as the number of speakers increases—making the clustering task more difficult—the performance gap between SDA and standard LDA widens. For a 50-speaker task, SDA achieves 85.9% accuracy compared to 79.6% for LDA and 69.4% for the original Euclidean space. This suggests that SDA is particularly robust in complex, multi-speaker scenarios.

Discussion

The success of SDA highlights the importance of directional scattering in high-dimensional speech modeling. In GMM supervector space, the "angle" of voice characteristics is often more informative than the absolute magnitude. By linearizing the spherical projection through graph embedding, the authors provide a practical way to leverage this insight without the computational complexity of non-linear spherical models.

Furthermore, the results demonstrate the power of semi-supervision as a bridge. Labeled data from general speaker recognition tasks can significantly enhance unsupervised clustering on new, unseen speakers, provided the underlying discriminative features are well-modeled.

Conclusion and Future Work

SDA represents a significant step forward in speaker clustering by aligning dimensionality reduction with the natural geometry of speech features. Future work could explore the application of spherical approaches to other speech processing domains, such as language or emotion recognition, where directional patterns are likely to be equally critical. Additionally, investigating the sensitivity of SDA to the representativeness of the training data remains an important area for further study.

Summary updated

BibTeX
@inproceedings{tang2009spherical,
  author = {Hao Tang and Hao Tang and Stephen M. Chu and Thomas S. Huang},
  title = {{Spherical Discriminant Analysis in Semi-supervised Speaker Clustering}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
}
Figure