What it does

This repo trains classifiers to detect emotion from audio—think angry, happy, sad, surprised—using four approaches under one roof: LSTM, CNN, SVM, and MLP. It handles the full pipeline from feature extraction through training to prediction, with pre-trained checkpoints included so you can run inference without waiting for a GPU to warm up.

The interesting bit

The author claims a bump to ~80% accuracy from an earlier version, mostly by improving feature extraction. The real utility is the side-by-side comparison: you can pit a lightweight SVM against a CNN or LSTM using the same preprocessed features and see which architecture actually suits your dataset. It also supports OpenSMILE’s standard INTERSPEECH feature sets (up to 6,373 features in ComParE_2016), which is more thorough than the usual librosa-only tutorials.

Key highlights

• Supports four model families: LSTM, CNN, SVM, and MLP, all sharing a common base class
• Feature extraction via librosa or OpenSMILE (IS09 through ComParE_2016)
• Pre-trained checkpoints available; YAML configs control the pipeline
• Built-in plotting: radar charts for prediction probabilities, waveform/spectrogram visualization, training curves
• Supports four datasets: RAVDESS, SAVEE, EMO-DB, and CASIA (English, German, Chinese)

Caveats

• The ~80% accuracy figure lacks specifics on which dataset or model achieved it
• Python 3.8 and TensorFlow 2 are pinned; no mention of newer versions
• OpenSMILE integration is optional and requires manual installation

Verdict

Worth bookmarking if you need a clean, comparable baseline for speech emotion recognition or you’re teaching the topic. Skip it if you need production-grade inference or real-time streaming—the repo is research scaffolding, not a deployed system.