Streaming state and sliding inputs

At sample rate f_s, a W-sample window spans W/f_s seconds, while an H-sample hop produces a new window every H/f_s seconds after initialization. Overlap is W−H when H is no larger than W. A detector can emit frequent decisions while each decision still depends on a long history. State whether outputs are causal, whether lookahead is required, and whether timestamps refer to the beginning, end, or center of the observed window.

Use integer sample indices rather than repeatedly rounding wall-clock durations. Clock drift, missing samples, and discontinuities need an explicit policy: drop a decision, insert marked padding, resample, or reset. A circular buffer reduces data movement only if producer and consumer ownership is correct. With DMA capture, do not overwrite samples still being read. Also distinguish raw-sample buffers, feature windows, cached activations, and recurrent state because each has different dimensions and retention requirements.

A streaming network can cache convolution history, maintain a recurrent hidden state, or update a feature extractor’s running statistics. Assign each state to a stream identifier and a compatible model/preprocessor version. Reset at session boundaries and after discontinuities according to the application contract, rather than carrying one person’s audio history into another session. On restart, decide whether state is restored from trusted storage or rebuilt from new samples. Either choice affects startup accuracy, latency, and privacy.

Changing a model can change state shape or meaning even when its output labels remain the same. Do not feed an old hidden state into new weights without a validated migration rule. ExecuTorch 0.7 documents restrictions around exported buffer mutation and backend support, so statefulness is not inferred from a Python module simply having a buffer. A portable deployment may expose state tensors as explicit inputs and outputs, with application-owned lifetimes and a tested invocation order.

Test a long reference recording both as one defined sequence and through the deployed chunk boundaries. Compare outputs after accounting for declared startup context, padding and lookahead; repeat with resets, missing chunks, and multiple interleaved streams. Overlapping windows share signal content and are not independent statistical evidence. Split evaluation by recording, subject, device session, or another defensible unit, and report event detection delay and false alarms over time rather than only an inflated count of nearly duplicate windows.

The modular lab illustrates circular index progression only. Reset, set modulus m to 10 and additive step a to 4, predict residue eight after two Steps and period five, then take those Steps. Compare a fresh Reset with both m at 10 and a at 3 before two Steps: residue six and period ten. Each Step advances one addition. This toy has no audio, window extraction, DMA ownership, cached network state, or evidence of streaming-model equivalence.