VAD Settings
Use the vad object in an Advanced Settings resource to configure voice activity detection, turn completion, and interruptions.
Overview
Adjust speech detection to reject background sounds, increase silence delays for callers who pause often, or disable interruptions when the Assistant must finish an utterance.
Note
This page describes VAD behaviour for Assistants using the LLM + TTS + STT architecture. STS (Speech-to-Speech) models detect turns inside the realtime provider, so these fields are translated to the provider's own controls and some are not supported at all. See VAD with STS models below.
Data model
All fields are optional. These are runtime defaults for omitted fields on a pipeline Assistant. Numeric fields also use the runtime default when set to null. Explicit values take precedence.
| Field | Type | Default | Description |
|---|---|---|---|
activationThreshold | number | null | 0.5 | Confidence level required to classify audio as speech (0.0 to 1.0) |
endpointingMinDelay | number | null | 0.3 | Minimum seconds to wait after the last detected speech before declaring end of turn |
endpointingMaxDelay | number | null | 2.5 | Maximum endpointing delay in seconds after the last detected speech |
minSpeechDuration | number | null | 0.05 | Minimum seconds of speech required to trigger a speech event |
disableInterruptions | boolean | false | When true, the user cannot interrupt the Assistant |
Example:
Code
How VAD affects turn-taking
VAD classifies incoming audio frames as speech or silence. The platform uses this signal for:
- Turn completion: VAD, transcription and the turn detector determine when the caller's turn is complete. The endpointing delay adapts within
endpointingMinDelayandendpointingMaxDelay. Response generation can start before the turn is committed. - Interruption handling: when the Assistant is speaking, the platform uses adaptive interruption detection to distinguish an interruption from brief backchannels.
activationThresholdaffects local speech detection, but detected speech does not always stop playback immediately.
minSpeechDuration filters out speech segments shorter than the configured duration.
Note
VAD operates purely on acoustic signals. It detects when someone is talking, not what they are saying. Transcript content is handled by the STT engine. For silence-based endpointing on the transcription side, see endpointing in STT Settings.
Configuring VAD properties
activationThreshold
How confident the VAD engine must be before it classifies incoming audio as speech. The value ranges from 0.0 (extremely sensitive) to 1.0 (extremely strict).
Code
This threshold is used to start local speech detection. Once speech has started, the engine uses a lower exit threshold to avoid rapidly switching between speech and silence. If the threshold is too high, quiet or soft speech may never be recognised, and the Assistant will appear to ignore the user. If it is too low, background noise may be classified as speech, causing the Assistant to interrupt itself or trigger false turns.
Adjust based on your environment:
- Noisy environments (call centres, streets): increase to
0.6to0.7to filter out background noise. - Quiet environments with soft-spoken users: decrease to
0.3to0.4to avoid missing speech.
Value must be between 0.0 and 1.0.
endpointingMinDelay
Minimum time (in seconds) the Assistant waits after the last detected speech before declaring the user's turn complete. It is the configured floor for adaptive endpointing, not a guarantee of the total audible response delay.
Code
When the user stops speaking, the platform considers the VAD signal, the transcript and the turn detector's prediction. Dynamic endpointing can increase the effective minimum based on pauses observed during the call. If the user resumes speaking before commitment, end-of-turn detection is restarted.
A shorter delay reduces waiting but risks cutting off callers who pause mid-sentence. A longer delay gives callers more time to continue and delays the reply.
- Fast-paced conversations: decrease to
0.3to0.4for snappier responses. - Deliberate speakers who pause often: increase to
0.7to1.0to avoid the Assistant jumping in during natural pauses.
Minimum value is 0, expressed in seconds. Must be less than or equal to endpointingMaxDelay.
endpointingMaxDelay
Maximum endpointing delay in seconds after the last detected speech. This caps how long turn detection waits during a pause.
Code
The Assistant operates endpointing in dynamic mode: it adapts the actual delay within the [endpointingMinDelay, endpointingMaxDelay] range based on observed pause patterns in the session. The delay shortens when the user tends to speak in quick bursts and lengthens when they take their time. endpointingMaxDelay caps the adaptive endpointing delay; it does not cap transcript delivery, model processing or total response latency.
Raise this bound for callers who pause during long answers. Lower it to reduce how long turn detection waits during pauses, for example in IVR flows or timed interactions.
Minimum value is 0, expressed in seconds. Must be greater than or equal to endpointingMinDelay.
minSpeechDuration
Minimum duration of audio (in seconds) that must be classified as speech before a speech event is triggered. This acts as a debounce filter: speech segments shorter than this value are discarded entirely, preventing the Assistant from reacting to very short noise bursts such as coughs, clicks, or brief background sounds.
Code
Brief coughs, clicks, or a door closing can trigger false speech events. Increase to 0.1 or 0.2 if the Assistant reacts to these sounds.
Minimum value is 0, expressed in seconds.
disableInterruptions
Set to true to prevent caller speech from cutting off the Assistant's current utterance.
Code
By default, Assistants support barge-in, with adaptive detection to avoid stopping for brief backchannels. Disabling interruptions does not guarantee that speech during playback is discarded.
Use this for scenarios where the Assistant must deliver complete messages without being cut off, such as:
- Legal disclaimers
- Critical instructions or safety information
- Confirmation summaries that must be heard in full
Note
Keep utterances short when interruptions are disabled. Callers cannot stop a long response to correct or redirect the Assistant.
VAD with STS models
With the LLM + TTS + STT architecture, these fields configure local VAD, endpointing and interruption handling. With an STS model, turn detection happens inside the realtime provider instead, and the vad object is translated into whatever equivalent controls that provider exposes. Providers differ in what they accept.
| Field | GPT Realtime | Gemini Live | Ultravox |
|---|---|---|---|
activationThreshold | Yes, applied as a confidence threshold | Approximated, see below | No |
endpointingMinDelay | Yes, as a silence duration | Yes, as a silence duration | No |
endpointingMaxDelay | No | No | No |
minSpeechDuration | Only when interruptions are disabled | Only when interruptions are disabled | No |
disableInterruptions | Yes | Yes | No |
Unsupported fields are ignored rather than rejected: the configuration saves and the conversation runs, but the value has no effect on the call.
Dynamic endpointing is not available
endpointingMaxDelay has no equivalent in any current STS provider. With provider-side turn detection, there is no adaptive [min, max] range. endpointingMinDelay maps to a provider silence duration or, for GPT Realtime semantic detection, to a responsiveness bias. Set it according to the provider behaviour below.
activationThreshold on GPT Realtime and Gemini Live
On GPT Realtime, the default turn detection is semantic: the model decides the user has finished based on what was said, not only on silence. Setting activationThreshold switches turn detection to a purely acoustic mode using that value as the confidence threshold. This is a meaningful behavioural change, so leave the field unset unless you specifically want silence-based turn detection.
If you leave activationThreshold unset and set only endpointingMinDelay, semantic detection is kept and the delay is applied as a responsiveness bias: below 1.0s the model commits turns eagerly, from 1.0s it waits longer, and from 4.0s it waits longest.
On Gemini Live, activationThreshold is approximated rather than applied directly. The provider exposes coarse sensitivity levels instead of a numeric threshold, so values of 0.5 and above map to low sensitivity and values below 0.5 map to high sensitivity. Intermediate precision is lost: 0.55 and 0.95 produce identical behaviour.
disableInterruptions on STS models
On GPT Realtime and Gemini Live, setting disableInterruptions to true turns off the provider's own turn detection and hands turn-taking back to the built-in VAD engine, which is what allows interruptions to be suppressed. As a result activationThreshold and minSpeechDuration become effective again on these two providers when interruptions are disabled, because the built-in engine is now the one detecting speech.
Ultravox does not support disabling interruptions
Ultravox always performs turn detection on its side and offers no way to switch it off. Setting disableInterruptions to true on an Ultravox Assistant has no effect: the request succeeds and the conversation runs with interruptions still enabled. If you need guaranteed uninterrupted playback, use GPT Realtime, Gemini Live, or the LLM + TTS + STT architecture.
Tuning by scenario
For noisy environments, combine activationThreshold 0.6 to 0.7 with minSpeechDuration 0.1 to 0.15 so brief chatter and background bursts are both rejected.
For conversations with deliberate speakers who pause often, raise endpointingMinDelay to 0.7 to 1.0 and, if you need more headroom for very long mid-thought pauses, increase endpointingMaxDelay to 4.0 to 5.0. Tune STT settings alongside these controls; STT finalisation does not by itself force the Assistant to reply.
Both scenarios assume the LLM + TTS + STT architecture. With provider-side STS turn detection, minSpeechDuration and endpointingMaxDelay have no provider mapping, so noise rejection rests on activationThreshold alone and extra headroom for long pauses comes from raising endpointingMinDelay. The STT endpointing field does not apply either, as there is no separate transcription step to wait for.