Genetic algorithm based feature selection for speaker trait classification

Dongrui WuInterspeech 2012

TL;DR

This paper introduces a novel approach for classifying speaker traits like personality and likability using acoustic features. By combining Fisher Information Metric for initial feature filtering with a Genetic Algorithm for refined feature selection, and then fusing multiple Support Vector Machines, the method achieves superior classification performance compared to baseline approaches on the INTERSPEECH 2012 Speaker Trait Challenge dataset.

What Problem Does This Paper Address?

Understanding a speaker's underlying traits—such as personality, likability, or even potential health issues (pathology)—from their voice is a fascinating area with applications in human-computer interaction, health monitoring, and security. While research has explored traits like emotion or age, personality, likability, and pathology have received less attention. A significant challenge in this domain is the sheer volume of acoustic features that can be extracted from speech (over 6000 in this case), many of which might be redundant or uninformative. Effectively selecting the most discriminative features is crucial for building accurate and efficient classification models. The INTERSPEECH 2012 Speaker Trait Challenge aimed to address this gap by providing a dataset and tasks for classifying these less-explored speaker traits.

Methods

The proposed method employs a multi-stage approach to feature selection and classification:

  1. Data Normalization: All acoustic features are normalized to the [0, 1] range to prevent features with larger numerical scales from dominating the learning process and to avoid numerical instability.
  2. Fisher Information Metric (FIM) Filtering: This step acts as a preliminary filter. The FIM quantifies the distinguishing power of each individual feature for a two-class problem. Features with low FIM scores are discarded, significantly reducing the feature pool. The paper uses the top 2000 features based on their FIM scores.
  3. Genetic Algorithm (GA) for Feature Selection: A Genetic Algorithm is used to further refine the feature selection from the filtered set. The GA evolves a population of "chromosomes," where each chromosome represents a subset of feature indices. The algorithm iteratively applies selection, crossover, and mutation operations to find feature subsets that maximize a fitness function.
  4. Fitness Function: The fitness of a chromosome is determined by the performance of a linear Support Vector Machine (SVM) trained on the selected features. It combines the unweighted average (UA) recall on a 5-fold cross-validation of the training data and the UA recall on a separate development dataset.
  5. Feature Subset Sizes: The GA is run for multiple target feature subset sizes (N = 100, 200, 300, 400, 500) to explore different levels of feature dimensionality.
  6. SVM Parameter Optimization: For each of the best feature subsets identified by the GA, the optimal complexity parameter (C) for the linear SVM is determined.
  7. SVM Fusion: Five individual SVM classifiers, each trained on a different feature subset (corresponding to the different N values) and optimized C, are combined using a majority vote. This fusion aims to improve the robustness and overall performance of the final classifier.

Datasets

The experiments were conducted on the INTERSPEECH 2012 Speaker Trait Challenge dataset. This dataset includes acoustic features extracted from speech recordings, with three sub-challenges: Personality, Likability, and Pathology. The paper specifically reports results for the Likability sub-challenge. The dataset was split into training, development, and test sets.

Key Insights and Ideas

  • Feature Redundancy is High: The Fisher Information Metric analysis revealed that a very small fraction of the initial 6125 acoustic features possess significant discriminative power. This strongly justifies the need for effective feature selection.
  • GA as a Powerful Feature Selector: The Genetic Algorithm proved effective in navigating the high-dimensional feature space to identify optimal feature subsets that lead to improved classification accuracy.
  • Fusion Enhances Robustness: Combining multiple SVMs trained on different feature subsets (obtained through the GA) via majority voting significantly boosts classification performance and robustness, mitigating the risk of relying on a single, potentially suboptimal, feature set.
  • Development Set is Crucial: Explicitly incorporating the development dataset into the GA's fitness function led to significantly better performance on that set, though it also introduced a risk of overfitting, as observed in the test set results.

Key Results

The proposed algorithm achieved superior performance compared to baseline SVM and Random Forest approaches on the Likability sub-challenge of the INTERSPEECH 2012 Speaker Trait Challenge.

  • Development Set Performance: The algorithm's UA and WA on the development dataset were significantly better than the baselines. For Likability, the algorithm achieved a UA of 68.6% and WA of 69.1%, compared to Baseline SVM (58.5% UA, 58.4% WA) and Baseline RF (57.6% UA, 57.5% WA).
  • Test Set Performance: The algorithm also outperformed the baselines on the test dataset, achieving a UA of 59.5% and WA of 59.6%, compared to Baseline SVM (55.9% UA, 56.1% WA) and Baseline RF (59.0% UA, 59.2% WA).
  • Overfitting Observation: While the development set performance was high, the test set performance was considerably lower. This indicates overfitting to the development set, a common challenge in machine learning.

Limitations and Future Work

  • Overfitting: The authors explicitly acknowledge that the algorithm suffers from overfitting to the development dataset, leading to a noticeable drop in performance on the unseen test set.
  • Future Work: The primary direction for future work is to reduce this overfitting. The authors suggest that further experiments are needed to explore methods for mitigating this issue. Additionally, while results for Likability are presented, the paper states that results for the Personality and Pathology sub-challenges will be reported in the final submission, implying these were also part of the experimental scope.

Summary updated

BibTeX
@inproceedings{wu2012genetic,
  author = {Dongrui Wu},
  title = {{Genetic algorithm based feature selection for speaker trait classification}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure