Abstract
This paper presents a novel approach to prosodic speaker verification by adapting the "i-vector" dimensionality reduction technique to handle discrete prosodic feature counts. By employing a Multinomial Subspace Model, the researchers achieved a significant 30% relative improvement in accuracy over traditional GMM-SVM systems, demonstrating that high-level behavioral features can be effectively modeled in low-dimensional subspaces.
Introduction
In speaker verification, systems typically rely on low-level acoustic features like Mel-Frequency Cepstral Coefficients (MFCCs) to capture vocal tract characteristics. However, high-level prosodic features—such as speech rhythm, pitch contours, and syllable durations—carry unique behavioral signatures. The primary challenge in utilizing prosody is its irregular nature; features like pitch are undefined during unvoiced segments, making them incompatible with standard continuous modeling techniques like Gaussian Mixture Models (GMMs).
Methodology
To address these challenges, the authors propose a Multinomial Subspace Model. The process begins with discretization, where Syllable-based prosodic features (SNERFs) are passed through an ensemble of small GMMs. The system counts the "occupation" of each Gaussian component, transforming irregular data into high-dimensional count vectors (approximately 30,000 dimensions).
Dimensionality Reduction and Optimization
The core of the approach is the Multinomial Shift. Unlike standard i-vector systems that assume Gaussian means live in a low-dimensional subspace, this model assumes the log-probabilities of a multinomial distribution reside in that subspace. This allows for massive dimensionality reduction, compressing the 30k-dimension vectors into a 250-dimension "i-vector."
As shown in the figure above, the Equal Error Rate (EER) improves significantly as the number of factors $r$ increases, eventually plateauing. This compression acts as a filter, retaining speaker-discriminative information while discarding noise.
Intersession Compensation and Scoring
To mitigate session noise (e.g., differences between landline and cellular recordings), the researchers applied Within-Class Covariance Normalization (WCCN) and Linear Discriminant Analysis (LDA) directly to the prosodic i-vectors. For classification, a simple dot product (cosine similarity) was found to be highly effective, drastically reducing computational costs compared to traditional Support Vector Machine (SVM) training.
Experimental Results
The system was evaluated using the NIST 2006 Speaker Recognition Evaluation (SRE) benchmark, with background training data sourced from NIST 2004 and 2005 SRE datasets.
The proposed system achieved a best Equal Error Rate (EER) of 8.83% using LDA and fast scoring. This represents a 30% relative improvement over the baseline GMM-SVM system (12.7% EER). Furthermore, the use of "fast scoring" reduced the computational cost of the classification phase by a factor of more than 100.
Conclusion
This study demonstrates that multinomial distributions are naturally suited for the event-based nature of prosody. The ability to apply intersession compensation to low-dimensional prosodic vectors proved to be the "secret sauce" for performance gains. Future work may involve further isolating session noise by splitting the total variability subspace into distinct speaker and channel subspaces.
Summary updated