Lexicost
AI voice agents / Costs, terms and calculators

Endpointing, turn detection and barge-in

Checked against its sources on

Endpointing is the point at which a speech system decides the caller has finished speaking, usually after a set stretch of silence. Turn detection is the broader decision that the agent may now reply, which some vendors hand to a model. Barge-in is the caller speaking over the agent and making it stop to listen.

What each control decides

Endpointing happens in speech recognition: the recognizer decides the caller's words are complete and finalizes them. The idea is older than AI agents. The W3C's VoiceXML 2.0 standard, a Recommendation since 16 March 2004, defines a property called completetimeout as "the length of silence required following user speech before the speech recognizer finalizes a result". It also states the trade-off: a long value "makes the computer's response slow", while a short one "may lead to an utterance being broken up inappropriately".

The simplest turn detection is a silence timer. Some vendors give the decision to a model that estimates whether the caller is done. ElevenLabs lets a buyer choose the version of its turn detection model, and Vapi's LiveKit option lets a buyer change how long the agent waits at each level of that estimate.

Barge-in is what all five vendors on this page call interruption. VoiceXML 2.0 gives each prompt a bargein attribute and says bargein should be disabled if the application author requires that the user "must hear all of a warning, legal notice, or advertisement".

The wait before the agent speaks is a pause the platform adds after it has decided the turn is over. Two more settings sit at the edges of a turn: who speaks first when the call connects, and how long a caller may stay silent before the agent speaks again. VoiceXML 2.0 gives each prompt a timeout, "the interval of silence allowed while waiting for user input after the end of the last prompt", after which the platform throws a noinput event.

Settings that decide when the agent replies

Two kinds of setting decide when the agent replies after the caller stops: endpointing and the wait before the agent speaks.

Retell. Endpointing: how long the transcriber waits to finalize speech, set per provider in Custom Settings. Endpointing defaults (dashboard): 1,000 ms for Deepgram and Azure, 1,500 ms for Soniox and AssemblyAI. The transcription guide keeps endpointing separate from the agent's own pause. Response Wait time: the agent's wait after the caller stops speaking; none by default, up to 5.5 seconds. The page that describes it says that in the API this is the responsiveness field (responsiveness: how responsive the agent is, 0 to 1; lower waits more before it responds; default 1). enable_dynamic_responsiveness: adjusts reply speed to the user's speech rate and past turn-taking; default false.

Vapi puts most of these settings in an assistant's startSpeakingPlan, where endpointing names the method that decides the caller has finished. Smart endpointing: Off by default; uses the transcriber's end-of-turn detection, else LiveKit (English) or Vapi. waitFunction (LiveKit): how long the bot waits to start speaking, from how likely the user has finished. The speech configuration guide says the Dashboard starts LiveKit’s waitFunction on the Aggressive preset, which the voice pipeline guide gives as 2000 / (1 + exp(-10 * (x - 0.5))), while the API reference gives the default as 20 + 500 * sqrt(x) + 2500 * x^3.

When no smart endpointing plan is set and the transcriber has no endpointing of its own, Vapi waits a minimum time after the transcription, depending on how it ends. onPunctuationSeconds: minimum wait after transcription ending with punctuation, before the model; default 0.1 s. onNoPunctuationSeconds: minimum wait after transcription ending without punctuation; default 1.5 s. The API reference text gives onNumberSeconds a default of 0.4 seconds, while its own @default tag and the voice pipeline guide give 0.5, and the speech configuration guide calls the 0.5-second wait after a number WaitSeconds.

Vapi also has customEndpointingRules: timeouts set by a regex on the customer's speech or the assistant's last message. A Deepgram Flux transcriber has end-of-turn settings of its own. eotThreshold (Deepgram Flux): confidence level required to trigger end-of-turn detection; default 0.7. eotTimeoutMs (Deepgram Flux): maximum wait in milliseconds before forcing the turn to end; default 5,000. The assistant's own pause, which the speech configuration guide calls Wait Time Before Speaking, is a separate field. waitSeconds: how long the assistant waits before speaking; default 0.4 seconds.

Bland describes the wait as adaptive. Resumption: nothing to configure; the agent measures each caller's pace and matches it. The same page says adaptive resumption "is rolling out and may not yet be active in your environment". An older field with interruption in its name sets this wait. interruptionThreshold (legacy, pathway node): a fixed wait in ms after the caller stops, before the AI responds. Bland recommends leaving it unset.

ElevenLabs. Turn eagerness: Eager responds quickly, Normal waits for natural breaks, Patient waits longer (guide). The guide gives no default; the API reference does. The API reference describes the turn eagerness levels as Low, Standard and High, while its own allowed values and the conversation flow guide name them patient, normal and eager. spelling_patience: more patience while the user spells numbers and named entities; auto or off; default auto. turn_model: the version of the turn detection model; default turn_v3. speculative_turn: starts LLM responses in silence before full turn confidence; may add LLM costs; default false.

Synthflow. Patience Level: how long the agent waits after the caller finishes before it responds: Low, Medium or High. Neither its guide nor its API reference states a default.

How much caller speech stops the agent

A higher number does opposite things at different vendors. At Retell and Bland it makes the agent easier to interrupt. At Vapi and Synthflow the number counts the words a caller must say first, so a higher one makes the agent harder to interrupt.

Retell. interruption_sensitivity: how easily the user can interrupt the agent, 0 to 1; at 0, never; default 1. Interruption Sensitivity (dashboard): new blank voice agents start at 0.9; the API default is 1. Keypad tones (DTMF) have their own switch. allow_dtmf_interruption: DTMF input interrupts the agent even when interruption_sensitivity is 0; default false. enable_backchannel: the agent interjects phrases like "yeah" or "uh-huh"; off if not set. backchannel_frequency: how often the agent backchannels when it can, 0 to 1; default 0.8.

Vapi uses a stopSpeakingPlan. numWords: how many words the customer must say before the assistant stops talking; default 0. voiceSeconds: how long voice activity must be detected before an interruption, if numWords is 0; default 0.2 s. Two lists override the word count. acknowledgementPhrases: phrases that never interrupt the assistant, such as "okay" and "uh-huh"; 22 by default. interruptionPhrases: phrases that always interrupt the assistant at once, such as "stop"; 20 by default. backoffSeconds: how long the assistant waits before talking again after being interrupted; default 1 second. In the speech configuration guide, numWords appears as Words to Stop Speaking, voiceSeconds as Voice Activity Detection and backoffSeconds as Pause Before Resuming.

Bland. interruptibility: how readily the agent yields the floor when the caller speaks over it, 0 to 4; default 2. block_interruptions: when true, the AI will not respond to or process the user's interruptions; default false. Its Send Call page treats the lowest interruptibility level and turning block_interruptions on as the same thing.

ElevenLabs. Interruptions: turned on by selecting interruption as a client event in the agent's Advanced tab. The guide states no default; the tool interruptions page states one. interruption_mode (per tool): whether the user can interrupt while a tool runs; by default they can at any point. interruption_ignore_terms: terms that do not trigger an interruption when the user says them, such as "gotcha". transcribe_on_disabled_interruptions: with interruptions off, still transcribe the user; default false.

Synthflow. Interruption Sensitivity: words the caller must speak before the agent stops and listens; Off disables it. The agent editor docs name the setting Interruption Sensitivity, with an Off choice, while the API reference has a field min_words_to_interrupt (0 to 5) with the same description; neither page says they are one setting. Neither page states a default. Fade Out at Interruption: how many frames the agent's voice takes to fade out once an interruption is detected.

Who speaks first, and how long a silence can last

Retell. begin_message_delay_ms: delays the first message, 0 to 5,000 ms; if unset or 0, the agent speaks at once. reminder_trigger_ms: user silence after agent speech that triggers a reminder to speak; default 10,000 ms.

Vapi. firstMessageMode: whether the assistant speaks first or waits for the user; default assistant-speaks-first.

Bland. wait_for_greeting: when true, the agent waits for the call recipient to speak first; default false.

ElevenLabs. turn_timeout: the longest the agent waits for the user's reply before re-engaging the user; default 7 seconds. initial_wait_time: wait for the user to start when the first message is empty; unset uses turn_timeout. The soft timeout covers a different silence, the wait for the AI model's reply. Soft timeout (timeout_seconds): seconds before the predefined message while awaiting the LLM; default -1 (off). The conversation flow guide says the soft timeout triggers only once per turn, while the API reference has max_soft_timeouts_per_generation (default 1), under which the filler fires every timeout_seconds until that limit is reached.

Synthflow. greeting_message_mode: human waits for the caller to speak first; agent_static or agent_dynamic greet on connect. How long to wait: with Human speaks first, seconds of silence (1 to 10) before the agent starts the conversation. The same page calls one value a balanced default but does not say the slider starts there. Pause Before Speaking: a delay in seconds before the agent begins talking at the start of a call. The agent editor docs describe Pause Before Speaking, while the API reference has two fields, initial_pause_seconds and ring_pause_seconds, and neither page says which one it is. Idle reminders: the agent sends a reminder message after the caller has been silent for a set period.

Often confused with

  • Latency. The time the system needs to produce a reply and start playing it. Most waits on this page are settings a buyer chooses, and a caller hears both as silence. Two Synthflow blog posts give different latency figures for Synthflow, and neither says what it measures: an average of 500 to 700 ms (July 3, 2025) and under 500 ms (May 21, 2025).
  • Voice activity detection (VAD). Deciding whether incoming sound is speech at all. In Vapi's voice pipeline guide, VAD "detects utterance-stop" before the endpointing decision and "detects utterance-start during assistant speech" before the interruption check, and voiceSeconds is described as a VAD duration threshold.
  • Backchannelling. Short sounds such as "uh-huh" that show a listener is following without taking the turn. Retell's agent can make them, and Vapi's acknowledgement phrases and the ElevenLabs interruption ignore terms keep the caller's own from counting as interruptions.

Vendors cited on this page

Bland AI, ElevenLabs Agents (ElevenAgents), Retell AI, Synthflow AI, Vapi. Each profile lists everything that vendor publishes, with its sources. All of them are side by side in the comparison.

Sources

Checked against its sources on 14 Nov 2026. Vendor figures on this page open their source when clicked. Lexicost has no commercial relationship with any vendor named on this site.

AI voice agent cost per minuteAdds up the five parts of a minute from each vendor's own published rates, with a box for any part a vendor does not price, and shows which parts each vendor's headline price leaves out.OpenAI Realtime API cost per minuteWorks out what OpenAI's per-token audio prices come to for a whole call and per minute, which OpenAI does not publish, and shows why a long call costs more per minute than a short one.
Source for every vendor figure on this page (56)
  1. Retell AI, turn-taking setting: Endpointing: how long the transcriber waits to finalize speech, set per provider in Custom Settings (Dashboard Custom Settings transcription mode; only providers with an Endpointing control).Select Custom Settings, then choose a Provider that supports all of the agent’s selected languages. For providers with an Endpointing control, adjust how long the transcriber waits to finalize speech. A longer window can help callers who pause while reading numbers, but can also delay the agent’s response. docs.retellai.com/build/transcription-mode Read 14 Nov 2026.
  2. Retell AI, turn-taking setting: Endpointing defaults (dashboard): 1,000 ms for Deepgram and Azure, 1,500 ms for Soniox and AssemblyAI (Dashboard Custom Settings, when no value is configured; as of October 2026).The dashboard offers these endpointing controls, in milliseconds (as of October 2026): Provider Slider range Default when no value is configured Deepgram 100–10,000 ms 1,000 ms Azure 100–10,000 ms 1,000 ms Soniox 500–3,000 ms 1,500 ms AssemblyAI 100–3,000 ms 1,500 ms docs.retellai.com/build/transcription-mode Read 14 Nov 2026.
  3. Retell AI, turn-taking setting: Response Wait time: the agent's wait after the caller stops speaking; none by default, up to 5.5 seconds (Dashboard Speech settings; from no added wait up to 5.5 seconds).Response Wait time: how long the agent deliberately waits after the caller stops speaking before it responds, from no added wait (the default) up to 5.5 seconds, shown in milliseconds below one second and in seconds above it. docs.retellai.com/build/conversation-flow/global-setting Read 14 Nov 2026.
  4. Retell AI, turn-taking setting: responsiveness: how responsive the agent is, 0 to 1; lower waits more before it responds; default 1 (Range 0 to 1 (API field on the voice agent)).Controls how responsive is the agent. Value ranging from [0,1]. Lower value means less responsive agent (wait more, respond slower), while higher value means faster exchanges (respond when it can). If unset, default value 1 will apply. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  5. Retell AI, turn-taking setting: enable_dynamic_responsiveness: adjusts reply speed to the user's speech rate and past turn-taking; default false (Boolean (API field on the voice agent)).If set to true, the agent will dynamically adjust how quickly it responds based on the user's speech rate and past turn-taking behavior in the call. If unset, default value false will apply. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  6. Vapi, turn-taking setting: Smart endpointing: Off by default; uses the transcriber's end-of-turn detection, else LiveKit (English) or Vapi (Part of startSpeakingPlan).Off: Disabled by default. When smart endpointing is set to “Off”, the system will automatically use the transcriber’s end-of-turn detection if available. If no transcriber EOT detection is available, the system defaults to LiveKit if the language is set to English or to Vapi’s standard endpointing mode. docs.vapi.ai/customization/speech-configuration Read 14 Nov 2026.
  7. Vapi, turn-taking setting: waitFunction (LiveKit): how long the bot waits to start speaking, from how likely the user has finished (Only with LiveKit smart endpointing; maps a probability from 0 to 1 to a wait in milliseconds).LiveKit Smart Endpointing Configuration: When using LiveKit, you can customize the waitFunction parameter which determines how long the bot will wait to start speaking based on the likelihood that the user has finished speaking: docs.vapi.ai/customization/speech-configuration Read 14 Nov 2026.
  8. Vapi, where the vendor's own pages disagree: The speech configuration guide says the Dashboard starts LiveKit’s waitFunction on the Aggressive preset, which the voice pipeline guide gives as 2000 / (1 + exp(-10 * (x - 0.5))), while the API reference gives the default as 20 + 500 * sqrt(x) + 2500 * x^3..The Dashboard currently starts with the Aggressive preset when you select LiveKit; choose a preset or enter your own expression in advanced mode. Also: Under the hood, this is parsed into a mathjs expression. Whatever you use to write your expression needs to be valid with respect to mathjs @default “20 + 500 * sqrt(x) + 2500 * x^3” docs.vapi.ai/api-reference/assistants/create Also: Aggressive (Fast Response): "waitFunction": "2000 / (1 + exp(-10 * (x - 0.5)))" docs.vapi.ai/customization/voice-pipeline-configuration docs.vapi.ai/customization/speech-configuration Read 14 Nov 2026.
  9. Vapi, turn-taking setting: onPunctuationSeconds: minimum wait after transcription ending with punctuation, before the model; default 0.1 s (0 to 3 seconds; part of startSpeakingPlan.transcriptionEndpointingPlan, which the API reference says is used only when smartEndpointingPlan is not set and the transcriber has no built-in endpointing).The minimum number of seconds to wait after transcription ending with punctuation before sending a request to the model. Defaults to 0.1. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  10. Vapi, turn-taking setting: onNoPunctuationSeconds: minimum wait after transcription ending without punctuation; default 1.5 s (0 to 3 seconds; part of startSpeakingPlan.transcriptionEndpointingPlan).The minimum number of seconds to wait after transcription ending without punctuation before sending a request to the model. Defaults to 1.5. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  11. Vapi, where the vendor's own pages disagree: The API reference text gives onNumberSeconds a default of 0.4 seconds, while its own @default tag and the voice pipeline guide give 0.5, and the speech configuration guide calls the 0.5-second wait after a number WaitSeconds..Number detection: If the latest message ends with a number, waits for onNumberSeconds (default: 0.5) Also: The minimum number of seconds to wait after transcription ending with a number before sending a request to the model. Defaults to 0.4. This setting exists because the transcriber will sometimes punctuate the transcription ending with a number, even though the customer hasn’t uttered the full number. This happens commonly for long numbers when the customer reads the number in chunks. @default 0.5 docs.vapi.ai/api-reference/assistants/create Also: The system detects the number “123456789” and waits for 0.5 seconds (WaitSeconds) to ensure the customer isn’t still speaking. docs.vapi.ai/customization/speech-configuration docs.vapi.ai/customization/voice-pipeline-configuration Read 14 Nov 2026.
  12. Vapi, turn-taking setting: customEndpointingRules: timeouts set by a regex on the customer's speech or the assistant's last message (Part of startSpeakingPlan; list of rules, each with its own timeout in seconds).These are the custom endpointing rules to set an endpointing timeout based on a regex on the customer’s speech or the assistant’s last message. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  13. Vapi, turn-taking setting: eotThreshold (Deepgram Flux): confidence level required to trigger end-of-turn detection; default 0.7 (Deepgram Flux models only; set on the transcriber, with no smartEndpointingPlan).eotThreshold (Default: 0.7): Confidence level required to trigger end-of-turn detection docs.vapi.ai/customization/voice-pipeline-configuration Read 14 Nov 2026.
  14. Vapi, turn-taking setting: eotTimeoutMs (Deepgram Flux): maximum wait in milliseconds before forcing the turn to end; default 5,000 (Deepgram Flux models only; milliseconds; set on the transcriber).eotTimeoutMs (Default: 5000): Maximum wait time in milliseconds before forcing turn end docs.vapi.ai/customization/voice-pipeline-configuration Read 14 Nov 2026.
  15. Vapi, turn-taking setting: waitSeconds: how long the assistant waits before speaking; default 0.4 seconds (0 to 5 seconds; part of startSpeakingPlan).This is how long assistant waits before speaking. Defaults to 0.4. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  16. Bland AI, turn-taking setting: Resumption: nothing to configure; the agent measures each caller's pace and matches it (The page says the wait can stretch to about two and a half seconds for a caller who pauses mid-sentence).Resumption governs what happens after the caller finishes: how quickly the agent takes its turn. This is adaptive: the agent measures each caller’s pace and matches it automatically. There is nothing to configure. docs.bland.ai/tutorials/agent-speech Read 14 Nov 2026.
  17. Bland AI, turn-taking setting: interruptionThreshold (legacy, pathway node): a fixed wait in ms after the caller stops, before the AI responds (Milliseconds; pathway node modelOptions).Legacy. A fixed wait, in milliseconds, after the caller stops speaking before the AI responds on this node. We recommend leaving this unset; use interruptibility to control interruptions. docs.bland.ai/api-v1/post/update_pathways Read 14 Nov 2026.
  18. ElevenLabs Agents (ElevenAgents), turn-taking setting: Turn eagerness: Eager responds quickly, Normal waits for natural breaks, Patient waits longer (guide) (conversation_config.turn.turn_eagerness; can be changed per workflow node).Three modes are available: Eager - The assistant responds quickly to user input, jumping in at the earliest opportunity. Best for fast-paced conversations where immediate responses are valued. Normal - Balanced turn-taking that works well for most conversational scenarios. The assistant waits for natural conversation breaks before responding. Patient - The assistant waits longer before taking its turn, giving users more time to complete their thoughts. Ideal for collecting detailed information or when users need time to formulate responses. Also: Defaults to normal Controls how eager the agent is to respond. Low = less eager (waits longer), Standard = default eagerness, High = more eager (responds sooner) Allowed values: patient normal eager elevenlabs.io/docs/api-reference/agents/create elevenlabs.io/docs/eleven-agents/customization/conversation-flow Read 14 Nov 2026.
  19. ElevenLabs Agents (ElevenAgents), where the vendor's own pages disagree: The API reference describes the turn eagerness levels as Low, Standard and High, while its own allowed values and the conversation flow guide name them patient, normal and eager..Three modes are available: Eager - The assistant responds quickly to user input, jumping in at the earliest opportunity. Best for fast-paced conversations where immediate responses are valued. Normal - Balanced turn-taking that works well for most conversational scenarios. The assistant waits for natural conversation breaks before responding. Patient - The assistant waits longer before taking its turn, giving users more time to complete their thoughts. Ideal for collecting detailed information or when users need time to formulate responses. Also: Defaults to normal Controls how eager the agent is to respond. Low = less eager (waits longer), Standard = default eagerness, High = more eager (responds sooner) Allowed values: patient normal eager elevenlabs.io/docs/api-reference/agents/create elevenlabs.io/docs/eleven-agents/customization/conversation-flow Read 14 Nov 2026.
  20. ElevenLabs Agents (ElevenAgents), turn-taking setting: spelling_patience: more patience while the user spells numbers and named entities; auto or off; default auto (Allowed values: auto, off; conversation_config.turn).Defaults to auto Controls if the agent should be more patient when user is spelling numbers and named entities. Auto = model based, Off = never wait extra elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  21. ElevenLabs Agents (ElevenAgents), turn-taking setting: turn_model: the version of the turn detection model; default turn_v3 (Allowed values: turn_v2, turn_v3; conversation_config.turn).Defaults to turn_v3 Version of the turn detection model to use. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  22. ElevenLabs Agents (ElevenAgents), turn-taking setting: speculative_turn: starts LLM responses in silence before full turn confidence; may add LLM costs; default false (Boolean; conversation_config.turn).Defaults to false When enabled, starts generating LLM responses during silence before full turn confidence is reached, reducing perceived latency. May increase LLM costs. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  23. Synthflow AI, turn-taking setting: Patience Level: how long the agent waits after the caller finishes before it responds: Low, Medium or High (Values: low, medium, high (API); agent editor, Voice section).Patience Level determines how long the agent waits after the caller finishes speaking before it begins its response. Low makes the agent respond almost immediately, Medium adds a natural pause, and High gives the caller extra time to continue. docs.synthflow.ai/voice-configuration Read 14 Nov 2026.
  24. Retell AI, turn-taking setting: interruption_sensitivity: how easily the user can interrupt the agent, 0 to 1; at 0, never; default 1 (Range 0 to 1 (API field on the voice agent)).Controls how sensitive the agent is to user interruptions. Value ranging from [0,1]. Lower value means it will take longer / more words for user to interrupt agent, while higher value means it's easier for user to interrupt agent. If unset, default value 1 will apply. When this is set to 0, agent would never be interrupted. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  25. Retell AI, turn-taking setting: Interruption Sensitivity (dashboard): new blank voice agents start at 0.9; the API default is 1 (New blank voice agents created in the dashboard; range 0 to 1).The API defaults to 1 when interruption_sensitivity is omitted. New blank voice agents created in the dashboard start at 0.9. docs.retellai.com/build/handle-background-noise Read 14 Nov 2026.
  26. Retell AI, turn-taking setting: allow_dtmf_interruption: DTMF input interrupts the agent even when interruption_sensitivity is 0; default false (Boolean (API field on the voice agent); can be overridden per conversation or subagent node).If set to true, DTMF input will interrupt the agent even when interruption_sensitivity is 0. Can be overridden per conversation or subagent node. Default to false. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  27. Retell AI, turn-taking setting: enable_backchannel: the agent interjects phrases like "yeah" or "uh-huh"; off if not set (Boolean (API field on the voice agent)).Controls whether the agent would backchannel (agent interjects the speaker with phrases like "yeah", "uh-huh" to signify interest and engagement). Backchannel when enabled tends to show up more in longer user utterances. If not set, agent will not backchannel. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  28. Retell AI, turn-taking setting: backchannel_frequency: how often the agent backchannels when it can, 0 to 1; default 0.8 (Range 0 to 1 (API field on the voice agent); only applies when enable_backchannel is true).Only applicable when enable_backchannel is true. Controls how often the agent would backchannel when a backchannel is possible. Value ranging from [0,1]. Lower value means less frequent backchannel, while higher value means more frequent backchannel. If unset, default value 0.8 will apply. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  29. Vapi, turn-taking setting: numWords: how many words the customer must say before the assistant stops talking; default 0 (0 to 10 words; part of stopSpeakingPlan; at 0, voiceSeconds is also used).This is the number of words that the customer has to say before the assistant will stop talking. Words like “stop”, “actually”, “no”, etc. will always interrupt immediately regardless of this value. Words like “okay”, “yeah”, “right” will never interrupt. When set to 0, voiceSeconds is used in addition to the transcriptions to determine the customer has started speaking. Defaults to 0. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  30. Vapi, turn-taking setting: voiceSeconds: how long voice activity must be detected before an interruption, if numWords is 0; default 0.2 s (0 to 0.5 seconds; part of stopSpeakingPlan; used only when numWords is 0).VAD duration threshold when numWords = 0. Determines how long voice activity must be detected before triggering an interruption. Range: 0-0.5 seconds (Default: 0.2) docs.vapi.ai/customization/voice-pipeline-configuration Read 14 Nov 2026.
  31. Vapi, turn-taking setting: acknowledgementPhrases: phrases that never interrupt the assistant, such as "okay" and "uh-huh"; 22 by default (Part of stopSpeakingPlan; list of strings).Defaults to ["i understand","i see","i got it","i hear you","im listening","im with you","right","okay","ok","sure","alright","got it","understood","yeah","yes","uh-huh","mm-hmm","gotcha","mhmm","ah","yeah okay","yeah sure"] These are the phrases that will never interrupt the assistant, even if numWords threshold is met. These are typically acknowledgement or backchanneling phrases. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  32. Vapi, turn-taking setting: interruptionPhrases: phrases that always interrupt the assistant at once, such as "stop"; 20 by default (Part of stopSpeakingPlan; list of strings).Defaults to ["stop","shut","up","enough","quiet","silence","but","dont","not","no","hold","wait","cut","pause","nope","nah","nevermind","never","bad","actually"] These are the phrases that will always interrupt the assistant immediately, regardless of numWords. These are typically phrases indicating disagreement or desire to stop. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  33. Vapi, turn-taking setting: backoffSeconds: how long the assistant waits before talking again after being interrupted; default 1 second (0 to 10 seconds; part of stopSpeakingPlan).This is the seconds to wait before the assistant will start talking again after being interrupted. Defaults to 1. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  34. Bland AI, turn-taking setting: interruptibility: how readily the agent yields the floor when the caller speaks over it, 0 to 4; default 2 (Integer 0 to 4: 0 Block interruptions, 1 Difficult, 2 Balanced, 3 Easy, 4 Always interrupt. Set per call in the /v1/calls request body; pathway nodes can override it).How readily the agent yields the floor when the caller speaks over it. Set with the interruptibility field (integer, 0-4, default 2). docs.bland.ai/tutorials/agent-speech Read 14 Nov 2026.
  35. Bland AI, turn-taking setting: block_interruptions: when true, the AI will not respond to or process the user's interruptions; default false (Boolean, /v1/calls request body; the Send Call page says interruptibility 0 is the same as block_interruptions: true).block_interruptions booleandefault:"false" When set to true, the AI will not respond or process interruptions from the user. docs.bland.ai/api-v1/post/calls Read 14 Nov 2026.
  36. ElevenLabs Agents (ElevenAgents), turn-taking setting: Interruptions: turned on by selecting interruption as a client event in the agent's Advanced tab (Dashboard Advanced tab, Client Events).Interruption settings can be configured in the agent’s Advanced tab under Client Events. To enable interruptions, make sure interruption is a selected client event. elevenlabs.io/docs/eleven-agents/customization/conversation-flow Read 14 Nov 2026.
  37. ElevenLabs Agents (ElevenAgents), turn-taking setting: interruption_mode (per tool): whether the user can interrupt while a tool runs; by default they can at any point (Values: allow, disable_during_tool, disable_during_tool_and_turn; on webhook, client and system tools and MCP servers).By default a user can interrupt the agent at any point, including while a tool is executing. For some tools this is undesirable: a payment confirmation or a compliance disclaimer read out after a lookup should be delivered in full. The interruption_mode field controls this per tool. elevenlabs.io/docs/eleven-agents/customization/tools/tool-configuration/tool-interruptions Read 14 Nov 2026.
  38. ElevenLabs Agents (ElevenAgents), turn-taking setting: interruption_ignore_terms: terms that do not trigger an interruption when the user says them, such as "gotcha" (List of strings, case-insensitive exact match; conversation_config.turn).List of terms that should not trigger an interruption when spoken by the user (e.g. ‘gotcha’, ‘understood’). Uses case-insensitive exact matching. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  39. ElevenLabs Agents (ElevenAgents), turn-taking setting: transcribe_on_disabled_interruptions: with interruptions off, still transcribe the user; default false (Boolean; conversation_config.turn).Defaults to false When interruptions are disabled, still transcribe what the user says so it can carry into the next turn. When off, user speech during a non-interruptible turn is ignored and won’t trigger a turn. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  40. Synthflow AI, turn-taking setting: Interruption Sensitivity: words the caller must speak before the agent stops and listens; Off disables it (Agent editor, Additional Settings).Interruption Sensitivity controls how easily the caller can interrupt the agent mid-sentence. The value represents how many words the caller must speak before the agent stops and listens. Setting it to Off disables interruptions entirely; lower values make the agent yield quickly, while higher values let it finish longer phrases before pausing. Also: You can set how many words the human must speak before the AI will stop speaking and listening. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant docs.synthflow.ai/additional-settings Read 14 Nov 2026.
  41. Synthflow AI, where the vendor's own pages disagree: The agent editor docs name the setting Interruption Sensitivity, with an Off choice, while the API reference has a field min_words_to_interrupt (0 to 5) with the same description; neither page says they are one setting..Interruption Sensitivity controls how easily the caller can interrupt the agent mid-sentence. The value represents how many words the caller must speak before the agent stops and listens. Setting it to Off disables interruptions entirely; lower values make the agent yield quickly, while higher values let it finish longer phrases before pausing. Also: You can set how many words the human must speak before the AI will stop speaking and listening. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant docs.synthflow.ai/additional-settings Read 14 Nov 2026.
  42. Synthflow AI, turn-taking setting: Fade Out at Interruption: how many frames the agent's voice takes to fade out once an interruption is detected (Frames of 20 ms each (API reference: 5 frames is a 100 ms fade)).Fade Out at Interruption sets how many frames the agent’s voice takes to fade out once an interruption is detected. docs.synthflow.ai/additional-settings Read 14 Nov 2026.
  43. Retell AI, turn-taking setting: begin_message_delay_ms: delays the first message, 0 to 5,000 ms; if unset or 0, the agent speaks at once (Milliseconds, range 0 to 5,000 (API field on the voice agent); only when the agent speaks first).If set, will delay the first message by the specified amount of milliseconds, so that it gives user more time to prepare to take the call. Valid range is [0, 5000]. If not set or set to 0, agent will speak immediately. Only applicable when agent speaks first. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  44. Retell AI, turn-taking setting: reminder_trigger_ms: user silence after agent speech that triggers a reminder to speak; default 10,000 ms (Milliseconds, positive number (API field on the voice agent)).If set (in milliseconds), will trigger a reminder to the agent to speak if the user has been silent for the specified duration after some agent speech. Must be a positive number. If unset, default value of 10000 ms (10 s) will apply. docs.retellai.com/api-references/create-agent Read 14 Nov 2026.
  45. Vapi, turn-taking setting: firstMessageMode: whether the assistant speaks first or waits for the user; default assistant-speaks-first (Allowed values: assistant-speaks-first, assistant-speaks-first-with-model-generated-message, assistant-waits-for-user).This is the mode for the first message. Default is ‘assistant-speaks-first’. Use: ‘assistant-speaks-first’ to have the assistant speak first. ‘assistant-waits-for-user’ to have the assistant wait for the user to speak first. docs.vapi.ai/api-reference/assistants/create Read 14 Nov 2026.
  46. Bland AI, turn-taking setting: wait_for_greeting: when true, the agent waits for the call recipient to speak first; default false (Boolean, /v1/calls request body).wait_for_greeting booleandefault:"false" By default, the agent starts talking as soon as the call connects. When wait_for_greeting is set to true, the agent will wait for the call recipient to speak first before responding. docs.bland.ai/api-v1/post/calls Read 14 Nov 2026.
  47. ElevenLabs Agents (ElevenAgents), turn-taking setting: turn_timeout: the longest the agent waits for the user's reply before re-engaging the user; default 7 seconds (Seconds; conversation_config.turn; the conversation flow guide gives a range of 1 to 30 seconds).Defaults to 7 Maximum wait time for the user’s reply before re-engaging the user elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  48. ElevenLabs Agents (ElevenAgents), turn-taking setting: initial_wait_time: wait for the user to start when the first message is empty; unset uses turn_timeout (Seconds; conversation_config.turn).How long the agent will wait for the user to start the conversation if the first message is empty. If not set, uses the regular turn_timeout. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  49. ElevenLabs Agents (ElevenAgents), turn-taking setting: Soft timeout (timeout_seconds): seconds before the predefined message while awaiting the LLM; default -1 (off) (Seconds; the conversation flow guide gives a range of 0.5 to 8.0 seconds and recommends 3.0).Defaults to -1 Time in seconds before showing the predefined message while waiting for LLM response. Set to -1 to disable. elevenlabs.io/docs/api-reference/agents/create Read 14 Nov 2026.
  50. ElevenLabs Agents (ElevenAgents), where the vendor's own pages disagree: The conversation flow guide says the soft timeout triggers only once per turn, while the API reference has max_soft_timeouts_per_generation (default 1), under which the filler fires every timeout_seconds until that limit is reached..Soft timeout triggers only once per turn to prevent multiple fillers in succession. Also: Defaults to 1 Maximum filler messages while waiting for a single LLM response. Fires every timeout_seconds until the LLM streams content or this limit is reached. elevenlabs.io/docs/api-reference/agents/create elevenlabs.io/docs/eleven-agents/customization/conversation-flow Read 14 Nov 2026.
  51. Synthflow AI, turn-taking setting: greeting_message_mode: human waits for the caller to speak first; agent_static or agent_dynamic greet on connect (Allowed values: human, agent_static, agent_dynamic; voicemail messages need human).Controls how the agent delivers the initial greeting. human: the agent waits for the human to speak first. agent_static: the agent speaks greeting_message once the call is established. agent_dynamic: the agent generates a greeting from the prompt and speaks it once the call is established. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant Read 14 Nov 2026.
  52. Synthflow AI, turn-taking setting: How long to wait: with Human speaks first, seconds of silence (1 to 10) before the agent starts the conversation (1 to 10 seconds; only when the greeting is set to Human speaks first).The How long to wait slider sets a fallback timeout (1 to 10 seconds). If the caller stays silent past that point, the agent steps in and starts the conversation on its own. Three seconds is a balanced default. docs.synthflow.ai/greeting-message-behavior Read 14 Nov 2026.
  53. Synthflow AI, turn-taking setting: Pause Before Speaking: a delay in seconds before the agent begins talking at the start of a call (Agent editor, Additional Settings; seconds).Pause Before Speaking adds a delay (in seconds) before the agent begins talking at the start of a call. Also: Duration before the agent starts speaking. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant Also: Ring time before the agent answers the call. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant docs.synthflow.ai/additional-settings Read 14 Nov 2026.
  54. Synthflow AI, where the vendor's own pages disagree: The agent editor docs describe Pause Before Speaking, while the API reference has two fields, initial_pause_seconds and ring_pause_seconds, and neither page says which one it is..Pause Before Speaking adds a delay (in seconds) before the agent begins talking at the start of a call. Also: Duration before the agent starts speaking. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant Also: Ring time before the agent answers the call. docs.synthflow.ai/api-reference/platform-api/agents/create-assistant docs.synthflow.ai/additional-settings Read 14 Nov 2026.
  55. Synthflow AI, turn-taking setting: Idle reminders: the agent sends a reminder message after the caller has been silent for a set period (Message and Idle Duration threshold are set by the customer; agent editor, Call Configuration, Advanced settings).When enabled, the agent sends a reminder message after the caller has been silent for a set period. You can customize the Message the agent says (for example, “I’m still here. Do you have any questions?”) and the Idle Duration threshold that triggers it. docs.synthflow.ai/call-configuration Read 14 Nov 2026.
  56. Synthflow AI, where the vendor's own pages disagree: Two Synthflow blog posts give different latency figures for Synthflow, and neither says what it measures: an average of 500 to 700 ms (July 3, 2025) and under 500 ms (May 21, 2025)..Average latency: 500 to 700ms Also: Synthflow offers latency under 500ms, dynamic emotion mapping, and multilingual voice cloning. synthflow.ai/blog/parloa-review synthflow.ai/blog/how-can-i-use-ai-to-reduce-the-number-of-call-center-agents Read 14 Nov 2026.