Music is the part of a video pipeline most likely to get it taken down. Platform trending audio cannot be attached through the publishing APIs, so anything posted automatically needs its own track. Stock libraries license per track and per use, which means re-reading terms every time the output changes, and a lapsed subscription can retroactively strip the rights to work already published. Taking audio from a streaming service is infringement, and the reliable outcome is a muted video. What a scheduled pipeline actually needs is a track of an exact length, at a known tempo, that is unambiguously ours.
sonic is an MIT-licensed Python package with numpy and scipy as its only dependencies. There are no samples, no trained model, and no network calls — every sound is built from oscillators, filters, and envelopes at render time, including Karplus-Strong plucked strings and synthesised drums. That is the point: audio generated this way is cleared by construction, and it is not an imitation of any existing record. Twenty-three themes cover pop, electronic, chill, cinematic, urban, and acoustic ground. Each track is composed onto five mix buses — drums, bass, harmony, lead, effects — which an arrangement engine gates section by section, so a track opens, builds, breaks, and resolves instead of looping. The mixdown applies per-bus EQ, sidechain ducking, stereo widening, a shared reverb, glue compression, and ITU-R BS.1770 loudness normalisation. Renders come out at an exact length, as a seamless loop, or as separate stems, with a sidecar of beat, downbeat, and section times for syncing video cuts.
Anyone publishing video on a schedule who needs a soundtrack they are certain they can use, and who would rather not re-check a licence every time a clip goes out.