Our friends and colleagues have been encouraging us to release it publicly, so here it is! We recently presented it at Sonar+D in Barcelona.
We’ve trained an AI model to understand the correspondence between music and language. We combine a machine listening system for audio signal processing with transformers for text embeddings.
Once trained, we index a huge catalogue of unseen audio, ensuring that the search system can efficiently scale to millions of tracks.
At the moment, our model is optimised for our preferred music; electronic and techno, ambient and relaxing music. We’re currently fine-tuning to handle all kinds of genres and moods.
Our model handles two types of discovery: (a) natural language search and (b) similarity search.
(a) Speak: search for music using freeform natural language prompts, describing the mood, aesthetic, texture, setting and context of a track.
(b) Music: discover tracks that are acoustically similar to ANY reference track from Apple Music.