Processing model

Audio uses planar channels and an explicit sample rate. Chain effects run serially in array order. The caller owns decode, encode, resampling, and buffers. Offline processing starts fresh; streams retain filter, delay, tail, and random history.

Chain is serial only. Branching, merging, and automatic delay compensation are available through the opt-in Graph v1 model, which does not change Chain v1.

Catalog latency badges are qualitative declarations. Python Stream.latency_samples and JavaScript ChainStream.latencySamples expose the same runtime aggregate. Chain.latency_samples() in Python and chain.latencySamples() in JavaScript report the same aggregate without opening a stream, so offline process() output can be phase-aligned. graph.latency_samples() in Python and graph.latencySamples() in JavaScript report the prepared common-max Graph aggregate without opening a stream, and GraphStream.latency_samples / GraphStream.latencySamples expose it on an open stream. EffeTuneNode.latencySamples exposes the cached real-time aggregate. Creation and awaited setParam() or reset() calls update it before returning.

Source-generating effects

The docs overlay explicitly marks types that can intentionally produce non-zero output from zero input at an active setting and sample rate. The candidate-package gate runs all 111 catalog types, using the same canonical assets as the public asset examples where required. It requires the overlay, public catalog, and frozen source-generation-v0.1.json member sets to match exactly, and treats a peak above 1e-7 as generated output. Those effect pages carry a warning; the absence of that mark is a fixture result for the selected settings, not proof that every possible future setting is non-generating.

Adaptive Prediction needs audible input to learn before it can generate sound on its own. Its fixture keeps one stream open while training, switching to Hold, and processing zero input, then checks the final second for sustained output. A freshly created or reset model remains silent in Hold without prior training.

Render a generator that needs no prior training by processing a zero input whose frame count defines the rendered length:

rendered = chain.process(np.zeros((2, frames), np.float32), sample_rate=sr)