Timbre

Overview

Upload a track, paste a link, or play into the mic, and Timbre writes out the notes. ffmpeg normalises the audio, librosa measures tempo and key by chroma correlation, then piano goes to ByteDance's onset/offset model and everything else to Spotify's basic-pitch. music21 quantises the result into MusicXML, LilyPond engraves the PDF, and the browser renders the score while an RQ worker streams five-stage progress over SSE. Around the transcriber sit four smaller surfaces: a Spotify taste profile, a library of analysed uploads, generated writing about why music sounds the way it does, and a daily pick set.

Highlights

• Swappable model registry with a fallback chain, so a missing checkpoint degrades to a weaker model instead of dropping the instrument
• Redis-backed job queue and pub/sub channel streaming live transcription progress to the browser over SSE
• MusicXML, MIDI and PDF on every run, with the score engraved in-browser via OpenSheetMusicDisplay