Accompanist

Every 40 ms the runtime builds one 144-integer conditioning vector: twelve style tokens, a 128-slot pianoroll of which pitches are masked, off, sustaining, attacking or simply on, one drums slot, and three discretized guidance scales. A small encoder graph turns that into a conditioning embedding, which the temporal decoder cross-attends to. The pianoroll is the whole control surface — velocity has no slot in it at all, which is why the module ignores how hard you play and cares only about what is held.

The temporal decoder is twelve layers over a sliding 41-frame window, roughly twenty seconds of context, with a fixed-size key/value cache that never grows or reallocates. Its output feeds a depth decoder that runs twelve sequential steps, one per residual-vector-quantization level, each producing logits from which a code is drawn with temperature and top-k using a seeded Gumbel-max — the same seed gives the same codes on any execution provider. Those twelve codes are the frame.

A SpectroStream decoder turns codes back into 1920 stereo samples at 48 kHz. The decoder here is the stateless windowed variant, so it re-decodes a context window per emitted chunk; that is why decode chunk size is a control rather than an implementation detail. The finished PCM is written into a lock-free single-producer ring, and a tiny stream-source AudioWorklet on the audio thread copies out of it. No inference ever runs in the render callback and no JavaScript callback is invoked from it.

Because generation and playback are decoupled by that ring, the module can be truthful about its own speed. It measures p90 frame time and classifies itself: comfortably under budget is 'live', keeping up on average is 'buffered', slower than realtime is 'offline'. The reservoir target follows, and the control latency it implies is displayed as a number in milliseconds rather than hidden.