Introduction
Speaker recognition systems aim to identify individuals based on their vocal characteristics. However, speech is inherently variable, consisting of diverse phonetic events like vowels, consonants, and silences. Traditional Gaussian Mixture Model (GMM) based systems often treat an entire utterance as a single homogeneous unit, potentially diluting the identity-bearing information present in specific acoustic segments. While supervised methods using Automatic Speech Recognition (ASR) to segment speech into phonetic units have been explored, they are limited by ASR accuracy and language dependence. This paper proposes an unsupervised ensemble learning framework that segments speech based on acoustic properties without requiring linguistic knowledge.
Methodology
The proposed approach utilizes unsupervised clustering to partition speech frames into distinct acoustic subsets. By training an ensemble of "weak" GMM classifiers on these subsets, the system can specialize in different acoustic regions.
Unsupervised Data Selection
Instead of relying on transcripts, the system employs clustering algorithms—specifically K-means and Divisive Hierarchical Clustering—to group speech frames based on raw acoustic features. The researchers found that using the Mahalanobis distance metric, which accounts for feature correlations, significantly improved clustering performance compared to standard Euclidean distance.
Ensemble Architecture
The original speech data is divided into $K$ non-overlapping subsets. For each subset, a separate GMM is trained. During the testing phase, the final decision is a weighted combination of the scores from these ensemble classifiers. The weights are determined by the frame count of the test utterance assigned to each cluster, ensuring that classifiers with more representative data for a given sample have a greater influence on the final decision.
Experimental Setup
The methodology was evaluated on two industry-standard benchmarks:
- NIST SRE 2008: Core task involving English trials.
- NIST SRE 2010: Telephone-to-telephone condition (det5). Universal Background Models (UBMs) were trained using data from previous NIST SRE cycles (2004-2006). The system also incorporated standard techniques such as Eigenchannel compensation and ZT-norm score normalization.
Results and Discussion
The ensemble approach demonstrated consistent improvements over baseline single-classifier systems across various configurations.
Key findings include:
- Error Reduction: On the NIST SRE 2010 dataset, an ensemble with 4 subsets and 256-mixture GMMs achieved a 20.85% reduction in the Detection Cost Function (DCF), improving from 0.5603 to 0.4435.
- Optimal Complexity: The performance generally improved as the number of subsets increased, reaching an optimal "sweet spot" between 8 and 16 subsets. Beyond this point, data sparsity began to degrade the performance of individual weak learners.
- Robustness: The system achieved a 14.97% reduction in DCF on the NIST SRE 2008 benchmark compared to a standard eigenchannel-compensated baseline.
Conclusion
This research demonstrates that unsupervised ensemble learning is a powerful tool for enhancing speaker recognition. By leveraging acoustic clustering rather than linguistic segmentation, the system remains language-independent and avoids the errors associated with ASR pre-processing. Future work could explore adaptive clustering techniques to automatically determine the optimal number of subsets based on audio characteristics and integrate these ensemble methods with modern deep learning architectures.
Summary updated