PSO-Optimized Feature Selection Results for Federated Speech Emotion Recognition

Results from a federated learning framework for speech emotion recognition, published by Mohammed Tawfik in May 2026. The dataset contains the outcomes of Particle Swarm Optimization (PSO) feature selection applied to multi-scale audio features from German (EmoDB) and English (RAVDESS) speech corpora.

Use Cases

• Benchmarking feature selection algorithms based on the described PSO-optimized results.
• Analyzing language-invariant emotion indicators based on the described prosodic features like fundamental frequency statistics.
• Studying cross-corpus generalization performance based on the reported accuracy drops for unseen datasets.
• Evaluating the impact of differential privacy on model utility based on the described membership inference attack resistance.

Strengths

• Results are associated with high reported accuracy on held-out test sets (e.g., 99.07% on EmoDB).
• Framework includes formal differential privacy guarantees, with membership inference attack resistance dropping to near-chance levels (AUC = 0.52).
• Feature selection and model components are validated through ablation experiments showing specific performance contributions (e.g., +2.80% from PSO).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is very small (5.5 KB), indicating it contains summary results rather than raw or extensive data.

Provenance

Source

figshare, author Mohammed Tawfik.

Collection Method

Generated from a federated learning experiment combining FedProx, a TCN-Transformer architecture, and PSO feature selection on speech emotion datasets.

Time Range

Experimental results from a study published in 2026.

Freshness

Last updated 2026-05-07 17:45:40; freshness should be verified.

Geography

Data derived from German (EmoDB) and English (RAVDESS, CREMA-D) speech corpora.