Reproduce “Emotion Recognition from Speech” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kannan Venkataramanan, Haresh Rengaraj Rajamohan
Researchers at New York University conducted a comprehensive comparative study on speech emotion recognition, demonstrating that Log-Mel Spectrogram features combined with 2D Convolutional Neural Networks achieved 66% accuracy for 14-class emotion classification, outperforming MFCCs and other models. The work also showed that feature choice impacts results more than model complexity and confirmed the model's robustness on an external dataset.
Key insights
- The choice of audio features has a greater impact on emotion recognition accuracy than the complexity of the deep learning model architecture.
- Log-Mel Spectrogram features were found to be consistently superior to Mel-Frequency Cepstral Coefficients (MFCCs) for speech emotion recognition.
- Two-dimensional Convolutional Neural Networks (2D CNNs), particularly with Global Average Pooling, emerged as the most effective model architecture among those tested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.