FedEmoNet: Statistical Validation Results for Federated Speech Emotion Recognition

99.07% accuracy was achieved on the EmoDB corpus and 98.96% on RAVDESS in a federated learning framework for speech emotion recognition. This 5.5 KB Excel dataset contains statistical validation results, including 10-fold cross-validation and zero-shot cross-corpus evaluation on CREMA-D, authored by Mohammed Tawfik and last updated in May 2026.

Use Cases

• Benchmark federated learning algorithms for audio tasks based on the described FedProx protocol and non-IID client setup.
• Analyze the impact of differential privacy on model utility and privacy based on the reported membership inference attack resistance.
• Study feature importance for cross-corpus emotion recognition based on the described SHAP-LIME analysis and prosodic features.
• Evaluate the generalization gap for speech emotion models based on the reported 20-40% typical accuracy drop on unseen datasets.

Strengths

• Results are validated with 10-fold cross-validation and reported with standard deviations (e.g., 99.07% ± 0.35%).
• Includes zero-shot cross-corpus evaluation on CREMA-D with 1,488 samples, showing a clear arousal-dependent transfer pattern.
• Ablation experiments quantify the contribution of key components like PSO selection (+2.80%) and the Transformer blocks (+2.10%).
• The dataset is openly licensed under CC-BY-4.0.

Limitations

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

Provenance

Source

figshare

Collection Method

Results from the FedEmoNet framework, combining federated optimization, a TCN-Transformer architecture, PSO feature selection, and differential privacy.

Time Range

null

Freshness

Last updated 2026-05-07 17:45:47; freshness should be verified.

Geography

null