Ser Eval

ser-eval

Toward Efficient Speech Emotion Recognition via Spectral Learning and Attention — HyeYoung Lee, Muhammad Nadeem (arXiv:2507.03251, 2025)

What this evaluates

Evaluates the capability of speech emotion recognition models to classify emotional states from audio recordings using spectral features and attention mechanisms. The protocol measures classification performance across multiple standard SER benchmarks to assess robustness and generalization.

Datasets

• SAVEE — total ?; splits: train (-1), test (-1)
• RAVDESS — total ?; splits: train (-1), test (-1)
• CREMA-D — total ?; splits: train (-1), test (-1)
• TESS — total ?; splits: train (-1), test (-1)
• EMO-DB — total ?; splits: train (-1), test (-1)
• EMOVO — total ?; splits: train (-1), test (-1)

Metrics

• accuracy (primary) — range: percent
  • Ratio of correctly predicted labels to the total number of observations: (TP + TN) / (TP + TN + FP + FN).

Input / output format

Input: MFCC spectral features extracted from audio recordings, processed through a 1D-CNN architecture enhanced with channel and spatial attention mechanisms.

Output: Predicted emotional class label.

Scoring recipe

def calculate_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

Common pitfalls

• Datasets use different emotion category sets and label mappings, requiring careful alignment before cross-dataset comparison.
• The 80:20 train/test split is applied independently per dataset; cross-dataset generalization or leave-one-out evaluation is not reported.
• Data augmentation (noise injection, pitch modification) is applied only during training, not during inference, which may overstate real-world robustness.

Evidence (verbatim from paper)

In this experiment, we used accuracy as the primary evaluation metric to measure and enhance the model’s effectiveness. Accuracy is defined as the ratio of correctly predicted labels to the total number of observations. The classification accuracy is calculated using the following equation: Accuracy = (TP+TN)/(TP+TN+FP+FN)... Table[2] presents a comparative analysis of classification accuracy between the proposed model and existing state-of-the-art methods for speech emotion recognition (SER), evaluated on six benchmark datasets: SAVEE, RAVDESS, CREMA-D, TESS, EMO-DB, and EMOVO.

Citation

@misc{lee2025efficient,
  title={Toward Efficient Speech Emotion Recognition via Spectral Learning and Attention},
  author={HyeYoung Lee, Muhammad Nadeem},
  year={2025},
  note={arXiv:2507.03251}
}

• arXiv: 2507.03251