SEGMENTAÇÃO E EXTRAÇÃO DE ESTRUTURAS EMOCIONAIS EM MÚSICAS
This project provides machine learning models for segmenting song lyrics into verses and choruses using BiLSTM and CNN architectures. It incorporates features like repetition detection, rhyme analysis, and boundary-aware loss for improved accuracy.
Clone the repository and install requirements:
git clone [repository-url] cd lyric-segmentation-ml pip install -r requirements.txt
Training is configured via YAML files in configs/training/.
Example: ./train.sh
Specific config: python train_with_config.py configs/training/all_features_boundary_aware_loss.yaml
For CNN: python train_cnn_with_config.py configs/training/cnn_test.yaml
Prediction uses models from training_sessions/.
Example: ./predict.sh
Specific: python predict_baseline.py --session training_sessions/[session_dir]
For CNN: python predict_cnn.py --session training_sessions/[cnn_session_dir]
Client was built using Streamlit, it uses the best 3 models with different training configuration.
To run:
streamlit run client/app.py
YAML configs in configs/ define parameters.
Training configs specify features, model params, loss.
Prediction configs handle inference settings like calibration.
- Models: BiLSTM, CNN
- Features: SSM-based for head, tail, string, phonetic, POS
- Advanced: Attention mechanisms, boundary-aware loss
See documentation/ for more details.
MIT
José Mena ([email protected])
Ricardo Malheiro ([email protected]; [email protected])
MERGE: Music Emotion Recognition - Next Generation (https://mir.dei.uc.pt/projects.html#indextag1)