← robert kirsammer en de
project

Music Fetch

One question — what is this track? — answered for inputs that are usually hopeless: noisy web video, whole playlists, local files. Local and free first; hosted services only if you bring credentials.

noisy inputvideo · playlists · filesvia yt-dlp + ffmpegone engineoverlapping windowsmix + instrumental stemearly-stop · long-mix clusteringcli — analyze · jobslocal http apiterminal uimacos app · swiftui

what it is

A local-first recognition tool. Ingestion follows whatever yt-dlp supports — YouTube, Shorts, Instagram, TikTok, Vimeo — plus anything already on disk. The engine scans overlapping windows from the original mix and, when available, a separated instrumental stem, so a voice-over or crowd noise doesn’t end the search.

Short clips use aggressive early-stop logic; long videos and mixes switch to a clustered mode with request budgets, excerpt caching, and repeated hits fused into timeline segments. Detached jobs can be submitted, watched, and inspected down to per-provider metrics.

the provider chain

1 local catalog chromaprint fingerprints of your own music
2 vibra unofficial shazam matching
3 audd hosted · only with your token
4 acrcloud hosted · only with your keys

vocal / instrumental separation cleans noisy audio before matching.