Interruption stops owned output while echo control follows what was actually rendered

Associate synthesis work and audio chunks with a session and generation identifier. When an interruption commits, retire the old generation and prevent its late results from entering the active playback queue. Propagate cancellation to upstream generation, synthesis, transport, and client buffers where supported, while still rejecting stale chunks that arrive after the signal. A producer can finish an already launched task after cancellation; that outcome is safe only if downstream ownership checks prevent its obsolete audio from being played. Handle resources and acknowledgments without double release.

Separate application-queued samples from samples already handed to an output device. The application can often discard its queue immediately, while device or operating-system buffering may need a separate stop or flush operation. Some sound is already physical and cannot be undone. Measure interruption-detection time, cancellation propagation, and final audible tail on the actual platform. Keep current-generation ordering and sample coverage valid after a restart so the new response does not inherit a stale index or accidentally mix with the old one.

A microphone can contain desired near-end speech, acoustic reflections of the system's output, and noise. An echo canceller estimates the render-to-capture path using an appropriately aligned reference, not merely the text that the synthesizer intended to say. Account for gain, resampling, channel mixing, device buffering, and acoustic delay. An output chunk discarded before playback should not be treated as if it was rendered, while an old device tail may still need a reference after logical cancellation. Separate capture and render clocks can drift and require compatible alignment handling.

The pinned WebRTC audio-processing interface illustrates distinct reverse/render and capture processing and explicit delay and drift concerns. That historical interface is a concrete contract, not a promise that every modern wrapper exposes the same calls. Double talk means desired speech and echo occur together; overly aggressive suppression can reduce echo while damaging the user. Nonlinear speakers, clipping, changing acoustic paths, and incorrect delay can also leave residuals. Evaluate near-end speech preservation and actual interruption success alongside echo metrics rather than maximizing suppression alone.

Replay tests with quiet users, overlapping syllables, room changes, headsets and speakers, different volumes, and late stale chunks. Label which waveform was actually rendered and when the user's interruption began. Report false interruptions, missed interruptions, audible stop latency, clipped user speech, and recognition outcomes. A low residual-energy score can result from suppressing both echo and desired speech, so it is not a sufficient quality guarantee. Test reconnects and generation-ID reuse as well as the ordinary case, because stale ownership bugs often appear only after delayed completion.

The waves lab illustrates an ideal cancellation mismatch. Reset, explicitly set BOTH phase=180 degrees and ratio=1, predict resultant amplitude zero, then Play or Step. Reset and re-set BOTH phase=180 degrees and ratio=0.8; predict residual amplitude 0.2. Edits retain t. These are equal-frequency analytic signals, not an adaptive echo canceller, microphone recording, or double-talk detector. The example shows that matching a reference's phase and level matters while leaving the actual room response, clocks, nonlinearities, and speech-preservation tests outside the toy model.