--- title: "Smart Turn and End of Turn Detection for Voice Agents" description: "A voice agent ends your turn on silence, on the sound of your sentence or on its words. Compare the three, then run Smart Turn in TypeScript." url: "https://micdrop.dev/blog/end-of-turn-detection-voice-agents" --- [Micdrop](/)›[Blog](/blog) # How Voice Agents Know You Stopped Talking A voice agent ends your turn on silence, on the sound of your sentence or on its words. Compare the three, then run Smart Turn in TypeScript. October 3, 2026 [Godefroy de Compreignac](https://github.com/Godefroy) Key takeaways * A voice agent decides that you stopped talking from one of three signals: a length of silence, the sound of your sentence, or its words. * A silence timer forces a compromise between cutting off people who pause to find a word and slowing down every answer. * Smart Turn is an open model of 8 million parameters that listens to the end of a turn and says whether the sentence sounds finished, in about 25 ms on a graphics card. * Run turn detection where the microphone is. The client can close a turn sooner, while a server can only decide to wait longer. * A held turn needs a deadline, because a model that hears an unfinished sentence where there is none would leave the call hanging. Most voice agents decide that you stopped talking when they hear a set length of silence after your last word. A timer cannot tell the end of a question from a speaker who pauses mid sentence to find a word. End of turn detection checks each pause before the agent answers. A model listens to the sound of the sentence, like Smart Turn or the detector in LiveKit’s agent framework, or a language model reads the transcript and judges whether the thought is complete. Listening to the sound is the faster of the two. Smart Turn, an open model that listens this way, now also runs in TypeScript, in a browser, on a phone or in Node. ## A silence timer cuts people off or slows every answer “I’d like to book a flight to…” The speaker stops to remember the city. A [voice activity detector](/blog/voice-activity-detection-browser) hears only silence there and starts counting. When the count reaches its limit, the turn closes and the agent answers half a question. The user then has to talk over it or start again. A longer wait fixes the cut and costs time on every exchange. [Micdrop](/), our open source TypeScript library for voice agents, can run the Silero voice detector in the browser and close a turn after a set number of Silero windows of silence. The default used to be twenty windows, which closed turns about 1060 ms after the last word. It is eight today, which closes them in 760 ms. We measured both on a MacBook Pro M2, in Chromium, over the French half of the Smart Turn test set. Three hundred milliseconds look small on paper, yet users hear them on every single answer. No silence length suits every speaker. A fast talker in a quiet office wants the short one, while someone dictating an address on a phone needs the long one. The way out is to judge the end of the sentence from something other than silence. ## Three signals can end a turn Signal What it reads Examples Cost at each pause Where it fails Silence How long nobody has spoken Any VAD alone Nothing Cuts off hesitations, or slows every answer Sound of the sentence Intonation, rhythm and the trailing words, in the audio Smart Turn, LiveKit’s audio turn detector 10 to 40 ms on a processor A sentence that is complete yet announces more Words of the sentence The transcript A language model asked through a tool A round trip to the model, and its tokens Waits for the transcript, so it adds latency The audio model and the language model both work on top of the voice detector. Each of them runs when the VAD hears a pause, and decides whether that pause ends the turn. ![The same sentence with a pause in the middle, ended by a short silence timer, a long one and by Smart Turn](/.netlify/images?url=_astro%2Fthree-ways-to-end-a-turn.CMUODpMw.jpg&w=1376&h=768&dpl=6ac135023c329800080a1757) ## Smart Turn listens to whether the sentence sounds finished [Smart Turn](https://huggingface.co/pipecat-ai/smart-turn-v3) is an open turn detection model published by Daily, the company behind the Pipecat voice agent framework, under the BSD 2 clause licence. It is a Whisper Tiny audio encoder with a small classifier on top, 8 million parameters in all. It reads up to the last eight seconds of a turn and answers with one probability that the speaker has finished. The current checkpoints cover 23 languages and weigh about 8 MB quantised for a processor, 31 MB at full precision. The audio holds what a transcript loses. A “to…” spoken with a voice that stays up and trails off sounds unfinished in any language. The model hears it straight from the audio, without waiting for a transcript. It also answers fast enough to run at every pause. Pipecat itself uses Smart Turn as its default way of ending a turn. We checked the two checkpoints on 240 turns of the model’s own test set, half French and half English: Checkpoint Size French English Full precision 31 MB 95.8% 99.2% Quantised 8.3 MB 93.3% 96.7% LiveKit went the same way. Its newer turn detector also reads the audio, covers 14 languages, and works from both its Python and Node SDKs. The full model runs on LiveKit Inference, the company’s hosted service, while a smaller version runs locally on the processor. Both ship under the LiveKit Model License. The older LiveKit detector, built on a small Qwen language model, read the transcript in 50 to 160 ms per turn. LiveKit now marks it as deprecated. The OpenAI Realtime API has its own version, called semantic VAD. A turn detection model scores the end of the sentence. The API turns that score into a wait capped at 2, 4 or 8 seconds depending on an `eagerness` setting. It only exists inside that API. ## Asking the language model whether the thought is complete A language model reading the transcript catches the sentences that sound finished and are not. “There are two things I want to change.” The voice drops at the end and the grammar is complete, yet the user is obviously about to list them. Micdrop’s server can put that question to the agent. `autoSemanticTurn` gives the language model a tool it calls when the last message is an incomplete thought. The server then [holds the answer back and waits for the rest](/docs/server/semantic-turn-detection), which arrives as a second user message. ``` const agent = new OpenaiAgent({ apiKey: process.env.OPENAI_API_KEY || '', systemPrompt: 'You are a helpful assistant', autoSemanticTurn: true,}) ``` The option costs a full round trip to the model and its tokens on every pause. That cost matters little when the agent already takes time to think between turns, which makes those calls the right place for it. When Smart Turn runs in the browser, start without this option and add it once you meet sentences the audio model gets wrong. ## Running Smart Turn in the browser and in Node Pipecat runs Smart Turn from Python. Micdrop publishes `@micdrop/smart-turn`, which runs the same model in TypeScript, on the ONNX runtime of the browser, of React Native and of Node. The package computes the Whisper audio features itself, eighty frequency bands every ten milliseconds, and depends on no other Micdrop package. In a Micdrop call, hand it to the client: ``` import { Micdrop } from '@micdrop/web'import { SmartTurn } from '@micdrop/smart-turn'import '@micdrop/smart-turn/web' await Micdrop.start({ url: 'wss://your-app.com/call', vad: 'silero', turnDetector: new SmartTurn(),}) ``` Any other voice application can call it directly. Feed it the audio while the user speaks, ask it at each pause, and reset it for the next turn: ``` const smartTurn = new SmartTurn()await smartTurn.load() smartTurn.push(samples, 48000) // every chunk of microphone audioconst { complete, probability } = await smartTurn.predict() // at a pausesmartTurn.reset() // once the turn is over ``` The browser runtime picks WebGPU when the browser offers it, and falls back to WebAssembly otherwise. We timed how long the model takes to answer once the speaker pauses, on an Apple M2 in Chromium, and emulated phones with a processor four and six times slower: Where Backend Laptop Mid range phone Slow phone Browser WebGPU 25 ms 26 ms 27 ms Browser WebAssembly, one thread 194 ms 837 ms 1244 ms Browser WebAssembly, four threads 68 ms 232 ms 345 ms Node One core 37 ms Node Four cores 14 ms On WebGPU the time barely changes from one device to the next, since the graphics card does the work. Chrome, Edge, Safari on iOS 26 and Samsung Internet all offer it. On a phone without WebGPU, the WebAssembly path takes between a quarter of a second and more than a second. Two options work better there. A React Native app runs the model on the [native ONNX runtime of the phone](/docs/react-native/turn-detection). A web app can run the same `SmartTurn` object on its server instead: ``` new MicdropServer(socket, { agent, stt, tts, turnDetector: new SmartTurn() }) ``` The server only hears the pause after the browser has already closed its turn, so it can only decide to wait longer. Prefer the [client side detector](/docs/client/turn-detection) wherever the browser offers WebGPU. ![Time Smart Turn takes to answer at a pause, in the browser on WebGPU and WebAssembly, in Node and in React Native](/.netlify/images?url=_astro%2Fwhere-smart-turn-runs.D3zB7L5z.jpg&w=1376&h=768&dpl=6ac135023c329800080a1757) ## Combining Smart Turn with Silero VAD The two models split the work. [Voice activity detection](/docs/client/vad) decides which audio is worth sending, so silences stay out of the stream and the speech to text bill stays the same. Smart Turn decides when the server is told to answer. During a pause in the middle of a sentence, the client sends nothing and the stream on the server stays open, so the agent receives one user message instead of two. graph TD A\[User speaks\] --> B\[Silero VAD hears a pause\] B --> C\[Smart Turn reads the turn\] C -->|Sounds finished| D\[Turn closes, the agent answers\] C -->|Sounds unfinished| E\[Turn stays open\] E -->|User speaks again| A E -->|No speech for turnMaxWait| D A VAD reports a pause only after a stretch of silence. If Smart Turn heard that silence, it would take it for a finished sentence, even in the middle of a hesitation. So `@micdrop/smart-turn` gives the model the audio only up to shortly after the last word. On the test set, the model flagged 94% of the unfinished sentences with that trimmed audio, against 82% when it also heard half a second of silence and 78% with a second and a half. Since the model handles hesitations, the VAD can stop covering for them. That is why Micdrop’s Silero default came down from twenty windows of silence to eight, which closes each turn 300 ms sooner. When you turn the detector off, put the longer wait back: ``` import { Micdrop, SileroVAD } from '@micdrop/web' // Without turn detection, the silence has to cover the hesitations aloneawait Micdrop.start({ url: 'wss://your-app.com/call', vad: new SileroVAD({ redemptionFrames: 20, minSpeechFrames: 8 }),}) ``` ## Tuning the threshold and the wait limit A turn held too long only delays the answer. A turn closed too soon cuts the speaker off, which users remember. When in doubt, let the detector wait. `SmartTurn` takes a `threshold`, 0.5 by default, the value the model’s own benchmarks use. Any threshold between 0.3 and 0.7 came within half a point of its accuracy in our runs, so change it only to favour one side. A lower threshold answers sooner and cuts people off more often, a higher one lets hesitations run longer. `turnMaxWait` limits how long a wrong verdict can hold the call. Once the model asks to wait, the turn closes on its own after four seconds by default and the agent answers. The clock only runs while nobody speaks, so a long answer full of hesitations never runs out of time. With the defaults, a sentence the model wrongly holds gets its answer just under five seconds after the last word, the 760 ms of silence plus the four seconds. Lower it if your users ask short questions, raise it for interviews and dictation. A lone “hmm” is speech that sounds finished, so turn detection closes the turn and the agent answers a noise. With [noise filtering](/docs/server/noise-filtering) turned on, the agent recognises those interjections and ignores them. Background noise is the VAD’s job rather than the turn model’s. In a noisy room, run the volume detector and Silero together, so a turn only opens when both hear a voice. ## Frequently asked questions ### What is end of turn detection? End of turn detection decides when a user has finished speaking, so a voice agent can answer. The simplest version waits for a fixed amount of silence. Better versions add a model that listens to the sound of the sentence, like Smart Turn, or a language model that reads the transcript, and keep the turn open while the sentence is unfinished. ### What is Pipecat Smart Turn? Smart Turn is an open turn detection model from the Pipecat team at Daily, under the BSD 2 clause licence. It is a Whisper Tiny encoder with a classifier, 8 million parameters, that reads up to eight seconds of audio and returns the probability that the speaker has finished. It covers 23 languages and runs in a few tens of milliseconds on a processor. ### Can Smart Turn run in JavaScript? Yes. The `@micdrop/smart-turn` package runs Smart Turn in TypeScript, in a browser on WebGPU or WebAssembly, in React Native on the native ONNX runtime, and in Node. It works with any voice application, Micdrop or not. ### What is the difference between VAD and turn detection? Voice activity detection answers “is someone speaking right now”, frame by frame. Turn detection answers “has this person finished”, once per pause. A voice agent with turn detection runs both: the VAD finds the pauses, and the turn detector decides which of them end the turn. ### What is semantic VAD? Semantic VAD is the name OpenAI gives to turn detection in its Realtime API. The API turns a model’s score of whether the user has finished into a wait of up to 2, 4 or 8 seconds, depending on the `eagerness` setting. It only works inside that API, while Smart Turn runs next to any speech to text and any language model. ### Should turn detection run on the client or on the server? On the client when the device can run it, since the client is what closes the turn and only it can answer sooner. On a browser without WebGPU, run the same model on the server. It costs a round trip and can only make the agent wait longer, but it spares slow phones the work. ## Getting started A silence timer forces you to pick between cutting people off and slowing every answer. Smart Turn takes that choice away for about 25 ms per pause on WebGPU, and the language model stays available for the sentences that sound complete and are not. Install the model next to the browser package: Terminal window ``` npm install @micdrop/web @micdrop/smart-turn ``` Then [set up turn detection in the browser](/docs/client/turn-detection), and add [semantic turn detection on the server](/docs/server/semantic-turn-detection) for browsers without WebGPU. If Micdrop is new to you, [run a first voice call in about five minutes](/docs/getting-started). ![Smart Turn and End of Turn Detection for Voice Agents](/.netlify/images?url=_astro%2Fthumbnail.BKm2_58N.jpg&w=1200&h=630&dpl=6ac135023c329800080a1757) On this page [1\. A silence timer cuts people off or slows every answer](#a-silence-timer-cuts-people-off-or-slows-every-answer)[2\. Three signals can end a turn](#three-signals-can-end-a-turn)[3\. Smart Turn listens to whether the sentence sounds finished](#smart-turn-listens-to-whether-the-sentence-sounds-finished)[4\. Asking the language model whether the thought is complete](#asking-the-language-model-whether-the-thought-is-complete)[5\. Running Smart Turn in the browser and in Node](#running-smart-turn-in-the-browser-and-in-node)[6\. Combining Smart Turn with Silero VAD](#combining-smart-turn-with-silero-vad)[7\. Tuning the threshold and the wait limit](#tuning-the-threshold-and-the-wait-limit)[8\. Frequently asked questions](#frequently-asked-questions)[9\. Getting started](#getting-started) On this page1\. A silence timer cuts people off or slows every answer2\. Three signals can end a turn3\. Smart Turn listens to whether the sentence sounds finished4\. Asking the language model whether the thought is complete5\. Running Smart Turn in the browser and in Node6\. Combining Smart Turn with Silero VAD7\. Tuning the threshold and the wait limit8\. Frequently asked questions9\. Getting started Build your own voice agent Micdrop handles the microphone, the streaming and the turn taking. Bring your own API keys and ship a voice mode in an afternoon. [Get started](/docs/getting-started) ## Keep reading [![Voice Activity Detection in the Browser: Silero vs WebRTC](/.netlify/images?url=_astro%2Fthumbnail.BUmw0JPI.jpg&w=1200&h=669&dpl=6ac135023c329800080a1757) August 13, 2026 ## Voice Activity Detection in the Browser: Silero vs WebRTC A browser voice agent needs voice activity detection to tell when the user speaks. Compare volume thresholds, WebRTC VAD and Silero, then tune your choice. ](/blog/voice-activity-detection-browser) [![Pipecat Alternative for TypeScript and Node.js Web Apps](/.netlify/images?url=_astro%2Fthumbnail.DvxNJo7U.jpg&w=1200&h=669&dpl=6ac135023c329800080a1757) March 3, 2026 ## Pipecat Alternative for TypeScript and Node.js Web Apps Micdrop runs the whole voice AI pipeline in TypeScript, browser and Node.js side, with provider fallback and semantic turn detection built in. ](/blog/alternative-to-pipecat) [![LiveKit Agents Alternative: Voice AI in TypeScript](/.netlify/images?url=_astro%2Fthumbnail.CVhkI8e6.jpg&w=1200&h=630&dpl=6ac135023c329800080a1757) August 13, 2026 ## LiveKit Agents Alternative: Voice AI in TypeScript Micdrop runs a voice call in TypeScript over a WebSocket to your own Node server. LiveKit Agents works in Node too, as a worker that joins a WebRTC room. ](/blog/alternative-to-livekit-agents)