Silero VAD End-of-Turn: 700ms of Silence Cut Off 118 of 1,000 Answers A developer at Preterview, a voice interview practice app, replayed 1,000 recorded answer turns through Silero VAD end-of-turn logic and found that a 700ms silence threshold triggered 118 premature turn endings, where speech resumed within three seconds. A threshold sweep showed 1,500ms cut false endings to 24 but added unacceptable reply latency, while an adaptive 700/2,000ms scheme reduced premature cutoffs to 21 with a 700ms median wait. The candidate said, "The main reason we moved off Redis was..." and then stopped to think. Seven hundred milliseconds later, my voice bot decided they were done talking. It generated a follow-up question, started speaking it, and got talked over two seconds later when the person finished their sentence. Barge-in handling cancelled the audio. The LLM call was wasted. And the transcript that got scored was half an answer. That was my Silero VAD end-of-turn logic working exactly as configured. I replayed 1,000 recorded answer turns through it offline to see how often it happened. The answer was 118 times. Here's the test, the threshold sweep, and the fix that got it down to 21 without making the bot feel slow. This is from Preterview, a voice interview practice app I run full disclosure: I built it . You talk to an AI interviewer out loud, it asks follow-ups, and at the end it scores the answers and writes a report. Pauses are the whole problem there, because people being interviewed pause constantly. You can see what a session looks like at preterview.com/en https://preterview.com/en . Casual voice chat is forgiving. Interview answers are not. People stop mid-sentence to recall a number, pick a word, or decide whether to admit something went badly. Those are exactly the pauses a silence timer reads as "I'm finished". Silero VAD returns a speech probability for each small audio chunk 512 samples at 16 kHz, so 32ms . End-of-turn is a rule you layer on top: if probability stays under a threshold for N milliseconds, the speaker is done. The VADIterator helper bakes this in with min silence duration ms . My original version was about this simple: python from silero vad import load silero vad, VADIterator model = load silero vad vad = VADIterator model, threshold=0.5, sampling rate=16000, min silence duration ms=700, for chunk in mic chunks 512 : 32ms of 16 kHz audio event = vad chunk if event and "end" in event: end turn send transcript to the LLM, start the reply There is nothing wrong with this code. Silero is doing its job: the person really was silent for 700ms. The bug is the assumption that 700ms of silence means the turn is over. I didn't want to hand-label 1,000 recordings, so I used a mechanical definition: A premature end-of-turn is any point where the rule fires end , and speech resumes within 3 seconds. If you're "done" and then keep talking within 3 seconds, you weren't done. This undercounts a little some people resume after 4 seconds and overcounts a little some people really do add a second thought . But it's consistent across every setting, which is what a sweep needs. The dataset was 1,000 answer turns from opt-in sessions and my own test accounts. I replayed the raw audio through the same chunking and VAD code, offline, with no network or TTS in the loop. That matters for the caveat at the end. There isn't a good fixed one. Every value trades cutoffs against reply latency, because the silence window gets added to the wait on every turn, including the ones that really are over. Here's the sweep: | Silence window | Premature end-of-turns of 1,000 | Wait after a real ending | |---|---|---| | 500ms | 231 | 500ms | | 700ms what I shipped | 118 | 700ms | | 1,000ms | 62 | 1,000ms | | 1,500ms | 24 | 1,500ms | | Adaptive 700 / 2,000ms | 21 | 700ms median | 1,500ms looks great on cutoffs. In practice it feels broken. You finish a sentence, and the interviewer just sits there for a second and a half before the LLM even starts. Add model time-to-first-token and TTS startup on top, and the gap is long enough that people say "hello?" Because the cut happens at the moment the speaker is most likely to pause: right before the hard part of the sentence. I pulled the streaming STT partial transcript at the instant each of the 118 cutoffs fired and looked at the last few words. That's 100 of 118 where the words themselves were shouting "not done yet". Nobody ends an answer on "because". The 18 complete-sentence cases are the honest limit. "We migrated in two weeks." followed by a long breath and "Actually, the rollback took another week." No rule on the text catches that, and I'm not sure it should. Interrupting after a complete sentence is something a human interviewer does too. Instead of one silence window, I pick the window from the tail of the partial transcript. Silero only tells me when silence starts. My own timer decides when it's long enough. INCOMPLETE TAIL = { "and", "so", "because", "but", "or", "um", "uh", "like", "which", "that", "the", "a", "to", "of", "first", "then", } def silence budget ms partial: str - int: text = partial.strip .lower if not text: return 2000 nothing said yet: they're still starting if text.endswith ",", "..." : return 2000 last = text.split -1 .strip "., ?" if last in INCOMPLETE TAIL: return 2000 return 700 vad = VADIterator model, threshold=0.5, sampling rate=16000, min silence duration ms=100 silence started = None for chunk in mic chunks 512 : event = vad chunk if event and "start" in event: silence started = None elif event and "end" in event: silence started = now ms if silence started is not None: if now ms - silence started = silence budget ms stt.partial : end turn silence started = None The VAD now reports silence starts after 100ms. My code decides whether that silence is a full stop or a breath. Results on the same 1,000 turns: Lowering the Silero threshold. I tried 0.35 to make soft trailing syllables count as speech. It barely moved cutoffs, because the problem was real silence, not quiet speech. It did start counting keyboard clicks as speech. Asking a small LLM "is this sentence finished?" It was more accurate than my word list on paper. But it runs in the critical path at the worst possible moment, and the round trip ate more latency than the adaptive window saved. A set lookup takes microseconds. Just letting barge-in clean it up. Barge-in stops the TTS, but the damage is done. The LLM already answered half an answer, and my scoring pipeline was grading the fragment. Split answers scored lower than the same person's unsplit answers, which is a scoring bug caused by an audio bug. This was an offline replay, not live traffic. In production, the STT partial lags the audio by a few hundred milliseconds, so the transcript tail is sometimes stale when the timer checks it. Those were the 3 misses in my run, and live it will probably be a few more. The word list is also English-only. A language where the verb comes last breaks the assumption that the last word tells you much. Not for long-form spoken answers. In my replay of 1,000 answers, a fixed 700ms silence window for Silero VAD end-of-turn detection cut people off 118 times, mostly right after words like "and", "because" and "so". Making the window longer fixes cutoffs only by slowing every reply. The better approach is to keep Silero for detecting when silence starts, then choose the silence budget from the partial transcript: short when the last words form a finished sentence, around 2 seconds when they end in a conjunction, filler or comma. That took my premature end-of-turns from 118 to 21 with no change to median reply latency. Written by the developer behind Preterview https://preterview.com/en , an interview prep platform.