03 / Audio / Search

Hum To Search.

A melody search engine that identifies a song from a few hummed notes. It matches relative pitch so a query can still resolve when it is sung in the wrong key.

My work: Retrieval, training, and acceptance logic
Built with: Python · PyTorch · SQLite · pytest

Illustration: comparing a hummed query with a melody’s relative pitch contour.

Melody retrieval

I wrote the query-by-humming search engine and used relative pitch to compare melodic shape across different keys.

Learned representation

I built a retrieval model using a ResNet-18 encoder and NT-Xent contrastive loss, trained on humming examples synthesized from MIDI.

False-positive handling

I traced how low-confidence candidates survived ranking, diagnosed a failure in the retrieval floor, and corrected the acceptance logic.

Next projectScenegraph