What it does

The project trains machine learning models to classify emotions from speech audio. It bundles four datasets (RAVDESS, TESS, EMO-DB, plus a custom noisy set), extracts standard audio features via librosa (MFCC, chromagram, mel spectrogram, contrast, tonnetz), and wraps both sklearn classifiers and Keras RNNs behind a unified Python interface. A test.py script lets you speak into a microphone for live prediction.

The interesting bit

The author did the tedious data-wrangling so you don’t have to: datasets are pre-formatted, features are pre-extracted on first run, and grid-search results are pickled for reuse. The “glue code” here is the value — it lowers the barrier from “I want to try speech emotion recognition” to actually running rec.train() in a few lines.

Key highlights

• Supports 9 emotion labels across 4 datasets, with optional class balancing
• 8 sklearn classifiers + RNNs (LSTM/GRU via Keras), plus regressor variants for 3- and 5-emotion subsets
• Automatic audio conversion to 16000Hz mono via ffmpeg if your files don’t match
• Pre-computed grid search results in grid/ folder; histogram plotting for model comparison
• Live microphone testing with python test.py

Caveats

• The README shows train scores of 1.0 against test scores of ~0.81–0.89, suggesting overfitting is present and unaddressed
• The “custom” dataset is explicitly described as “unbalanced noisy” — quality unclear
• librosa and scikit-learn versions are pinned to releases from 2018–2021; compatibility with current Python/package versions is untested

Verdict

Good for students or researchers who need a working baseline fast and don’t mind dated dependencies. Skip it if you need production-grade robustness or modern deep-learning architectures; the RNN implementation is basic (128-unit LSTM stacks) and the evaluation metrics are thin.