How rayanwang/speech_emotion_recognition_blstm works
Main Components
a. Data Sources
• Raw audio datasets (e.g., Berlin speech dataset)
b. Data Ingestion & Preprocessing
• dataset.py: loads raw audio and labels
• wav_extraction.py + pydub (utility/audio, utility/audiosegment): extracts WAV segments
• sentence_extraction.py + webrtcvad: splits audio into utterances
c. Feature Engineering
• External libs: librosa, pyAudioAnalysis
• utility/metrics_util.py, globalvars.py: compute & normalize features
• stFeatureSpeed patch in pyAudioAnalysis for mfcc, spectral features
• .p file storage of feature streams
d. Model Training & Hyperparameter Tuning
• DataAugmentationTraining.py: optional data-augmentation routines
• find_best_model.py: Hyperas + Keras (TensorFlow backend) for tuning (optimizers, batch_size, epochs)
• model_cross_validation.py: k-fold or speaker-independent CV using scikit-learn
e. Prediction & Evaluation
• prediction.py: loads saved model (.h5) + features → emotion labels
• utility/networks.py: BLSTM architecture definition
• utility/metrics_util.py: accuracy, confusion matrix
f. Utilities & Configuration
• utility/globalvars.py: global paths, hyperparameter defaults
• LICENSE, README.md, environment (Python, CUDA)
Architectural Patterns
• Pipeline pattern: strictly ordered stages
• Modular code: separate scripts for each pipeline stage
• Reusable utility modules for I/O, feature calculation, network definitions
• Hyperparameter-search sub-pipeline (Hyperas orchestrator calling find_best_model.py)
Technologies
• Python (2.7/3.6)
• Keras 2.1.5 on TensorFlow 1.6
• librosa, pyAudioAnalysis, pydub, webrtcvad
• Hyperas + scikit-learn
• NVIDIA GPU (CUDA)
Essential Elements
• Component labels (script names, utilities)
• Data stores: raw audio, feature pickles, model files