{
  "id": 7758644,
  "title": "Five hundred milliseconds of silence: how a voice agent decides you have finished",
  "url": "https://urgent.news/2026/09/16/five-hundred-milliseconds-of-silence-how-a-voice-agent-decides-you",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-09-16T09:57:39.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/whywave/five-hundred-milliseconds-of-silence-how-a-voice-agent-decides-you-have-finished-4jjn"
  },
  "original_language": "en",
  "account": "The core issue for voice assistants is determining when a person has finished speaking. A delay in detecting the end of a turn can lead to cutting off the speaker mid-sentence. Conversely, waiting too long could cause the user to feel unheard. This dilemma presents a window of just a few hundred milliseconds where most systems close it by implementing a timer. At WhyWave, a voice companion for children aged three to five, this timer became a point of contention. The research behind this decision starts with observing human conversation patterns. A 2009 study showed that across ten languages, the most common situation occurred when the answer began exactly as the question ended, with a median offset of zero milliseconds. However, the average delay was around two hundred and eight milliseconds, with Japanese speakers being the fastest at seven milliseconds and Danish speakers the slowest at four hundred and sixty-nine milliseconds. This indicates that people tend to answer before the other person has fully stopped talking, likely due to their predictive abilities. The timer approach, while common, has limitations. It fails in several ways for children. They answer more slowly than adults, with a median response latency of 625 milliseconds compared to 371 milliseconds for adults. Moreover, younger children are more likely to pause mid-sentence, with five-year-olds stopping nearly twice as often as adults. These nuances make the timer less effective for younger users, as their pauses often coincide with the timer triggering. Recent advancements in machine learning offer alternative solutions. Models like OpenAI's semantic_vad analyze the words spoken to predict when a turn might end, adjusting the timeout dynamically based on the context. Another approach, Pipecat's Smart Turn v2, uses a smaller, more efficient model to process the raw waveform and detect intonation cues, providing a faster and more accurate response. Despite these improvements, the challenge of accurately identifying the end of a speaker's turn remains complex, particularly with children whose speech patterns differ significantly from adults.",
  "summary": "Every voice assistant makes one decision more often than any other and discusses it less than any other: the person went quiet, have they finished the thought or are they taking a breath? You can get it wrong in both directions and both are obvious. Move too early and you cut the person off mid-sentence. Play it safe and they sit in silence wondering whether they were heard at all. Between those…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}