How renovamen/speech-emotion-recognition works
Project type and purpose
• This is a machine-learning pipeline/tool for Speech Emotion Recognition (SER).
• It provides end-to-end functionality: raw audio → feature extraction → model training → checkpointing → inference/prediction → visualization.
File structure
• Top-level “scripts”:
• preprocess.py – orchestrates feature extraction & data preprocessing
• train.py – runs model training based on a YAML config
• predict.py – loads a checkpointed model to infer emotion on new audio
• Core modules:
• extract_feats/ – two interchangeable feature extractors (librosa.py and opensmile.py)
• models/
– base.py & dnn/dnn.py – base classes
– dnn/cnn.py and dnn/lstm.py – neural architectures
– ml.py – classical ML (SVM & MLP via scikit-learn)
• utils/ – shared helpers (file operations, CLI opts, plotting)
• Data and metadata:
• configs/*.yaml – all run-time parameters (paths, model hyperparams, feature set)
• features/ – serialized feature outputs (.p or .csv)
• checkpoints/ – serialized trained model weights & configs
• Support files: requirements.txt, LICENSE, README.md
README for additional insights
• Lists external libraries: TensorFlow/Keras (for CNN/LSTM), scikit-learn (SVM/MLP), joblib, librosa, SciPy, pandas, matplotlib, NumPy, optional Opensmile.
• Describes multi-dataset support, configurable feature sets (IS09, IS10, etc.).
• Emphasizes two feature-extraction backends and four model families.