AI/MLSWEMar 2026 – Present
Melody Retrieval System
Creator
A “sing-to-search” project: hum a tune and retrieve ranked matches from a MIDI corpus using learned audio embeddings.
My contribution
Building the embedding and retrieval pipeline, including triplet-loss training and hard-negative mining.
- MERT-v1-95M
- Backbone
- FAISS ANN
- Retrieval
- H100
- Trained on
Highlights
- Building a music search engine that ranks matches for vocal input against the POP909 MIDI dataset using embedding similarity.
- Used a MERT-v1-95M transformer backbone for robust audio embeddings, integrated with FAISS for approximate nearest-neighbor retrieval at scale.
- Fine-tuned a custom MelodyEmbeddingModel on an H100 GPU with triplet loss and hard-negative mining to sharpen embedding separation and retrieval precision.
Stack
PythonMERT-v1-95MFAISSPyTorchHuggingFace