Voice activity is evidence for an endpoint policy, not the endpoint itself

An energy threshold can be useful for a controlled signal but is not a universal speech detector: quiet speech can have low energy and loud machinery can have high energy. Learned or statistical VAD systems use other evidence and can still fail under distribution shift. Record channel layout, sample rate, frame size, preprocessing, and detector version. A binary decision, a mode setting, and a posterior score are different API contracts. Do not invent a calibrated confidence value from a detector that only reports active, inactive, or an error.

The pinned WebRTC implementation at commit f54860e9 accepts eight-, sixteen-, thirty-two-, and forty-eight-kilohertz input with ten-, twenty-, or thirty-millisecond frames. Its source validates sample counts against those combinations and returns a binary activity decision or an error. This implementation detail is pinned deliberately because comments in older headers can lag the actual accepted-rate table. Its aggressiveness settings change internal behavior; they are not an application-selected probability threshold. Validate the wrapper and native version together instead of assuming every similarly named VAD exposes the same rates or scores.

An illustrative endpoint rule waits for a fixed number of consecutive inactive frames after observed speech. Longer waiting can protect pauses and trailing consonants but delays the reply; shorter waiting can improve speed while clipping continuing speech. A false active frame can reset the silence counter, while a missed quiet phrase can falsely extend it. Detector-internal hangover, application hangover, encoder lookahead, packet delay, and processing time can stack. Measure their actual composition rather than reporting the configured silence duration as the complete end-to-end endpoint latency.

A conversational endpoint may also use lexical completeness, punctuation hypotheses, prosody, or dialog state, but those signals are uncertain and can depend on future context. Define whether a decision is provisional and revisable or commits the application to generating a response. Keep a short pre-roll where needed to avoid losing the beginning of speech when activity starts, while preserving source indices and access limits. Test turn-internal pauses, fillers, hesitations, different speaking rates, and non-speech sounds instead of tuning solely on clean recordings with obvious gaps.

Measure false endpoints, clipped trailing words, delayed endpoints, false starts, and the final interaction outcome. A single frame accuracy can look strong while errors cluster at precisely the quiet trailing speech that matters most. Use labeled turn boundaries with a stated tolerance and inspect outcomes by speaker characteristics, language, microphone, noise, and conversational style. A change in average latency should be read alongside completion and clipping rates. Preserve rejected and difficult cases rather than excluding them from a success-only latency dashboard.

The network lab illustrates threshold behavior only. Reset, explicitly set BOTH x=1 and w=2, predict output about 0.8176, then Play or Step. Reset and re-set BOTH x=0.5 and w=2; predict about 0.6225. With a paper threshold 0.7 only the first is active. Edits retain t; bias is fixed at −0.5. There is no audio, VAD training, calibration, silence counter, or endpoint policy here. The toy score demonstrates a decision threshold without implying that a sigmoid output or a WebRTC binary decision measures the probability of turn completion.