I built the speech‑emotion‑onnx project as a Jupyter Notebook repository on GitHub to perform speech emotion recognition with a custom PyTorch model, exported to ONNX and further optimized with INT8 dynamic quantization.
The project achieves a 3.56x inference speedup and a 73% model size reduction.
I use a 4-layer MLP with BatchNorm + Dropout:
Input → 512 → 256 → 128 → num_classes
Training uses Adam, lr=0.001 | StepLR scheduler | 60 epochs | Batch 32.
Data is split with an 80/20 stratified split.
Test set accuracy is 78.8% accuracy on test set.
INT8 quantization yields 73% size reduction via INT8 quantization.