Voice, Listening and Turn-Taking

Open in CMS

On this page you choose how your assistant sounds and how it listens: its voice, its speaking style, the model that turns speech into text, and (for Turn-by-turn calls) how it decides when the user has finished talking.

All of these settings are on the Agent voice page (Conversation β†’ Voice in the CMS sidebar), which is one page of sections. They appear only when Voice conversations is on: the voice and speaking style in the Voice section, and speech recognition and turn-taking in the Call flow section, right below Conversation style. Interruptions and push-to-talk sit in the Advanced fold of Call flow.

In this example Atlas, Kirana Group's operations assistant, keeps the defaults: the voice Marin, speech recognition gpt-transcribe, the speaking style Everyday, and Turn-by-turn listening without push-to-talk. The sections below explain each setting and when you would change it.


Before you start

  • Turn on Voice conversations and choose a conversation style. Several settings here are Turn-by-turn only and are hidden when Natural conversation is selected.
  • Optional: write a welcome message, so the voice sample speaks your real greeting.

Voice and listening

The Voice section lets you choose the voice the AI assistant speaks with and how it sounds in conversation. The voices it offers follow where the voice runs, chosen in the Models section: Qlar voice (GPT-Realtime voices) for Turn-by-turn, Qlar voice (GPT-Live voices) for Natural conversation, or the voices of your own account or a connected computer.

Voice section for Atlas: Voices from Qlar voice (GPT-Realtime), Assistant's voice with the voice Marin and a play button to preview a short sample, and Speaking style Everyday (conversational, with an occasional hmm or filler; formality follows the persona)

Assistant's voice

Selects the voice used for the assistant's spoken responses. Each conversation style has its own list and keeps its own choice.

Turn-by-turn: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin (default), cedar.

Click the β–Ά button next to the dropdown to hear a sample in the selected voice before saving. The sample speaks your agent's own welcome message, so you hear the voice saying what users will actually be greeted with. If no welcome message has been written yet, a short built-in English sentence is spoken instead.

Natural conversation: 22 voices, default cedar: the ten above, plus twelve with a regional accent:

VoiceLanguageRegional influencePresentation
beaconEnglishFilipinoMasculine
bossaPortugueseBrazilianFeminine
cinderEnglishSouthern U.S.Masculine
deltaEnglishSouthern U.S.Feminine
gleamEnglishNorth AmericanFeminine
meridianEnglishNorth AmericanMasculine
quartzEnglishAustralianFeminine
rippleEnglishAustralianMasculine
stoneEnglishIrishMasculine
tempoPortugueseBrazilianMasculine
vesperEnglishBritishMasculine
willowEnglishIrishFeminine

Regional influence describes the voice's speaking style; listen to a voice in your agent's language (for example with the β–Ά sample) before choosing it. A voice that appears in both lists still sounds a little different in Natural conversation. For Natural conversation, the β–Ά button plays a fixed sample of the voice rather than your welcome message.

While Natural conversation is selected, the Turn-by-turn list moves to its own Voice-note replies section next to the Voice section, as Voice for voice-note replies: the voice used when the assistant answers with a voice note (for example on WhatsApp). Voices offered only in Natural conversation can't be used for voice notes.

Speech recognition

Turn-by-turn only. The first setting in the Call flow section. Hidden when Natural conversation is selected: it has built-in speech recognition, which is not charged separately.

Selects the model used to transcribe the user's spoken input into text. Most teams should keep the recommended default.

ModelNotes
gpt-transcribeDefault. Billed per minute of audio rather than per token.
gpt-4o-mini-transcribeFaster and cheaper than gpt-4o-transcribe, billed per token.
gpt-4o-transcribeHigh-accuracy transcription, billed per token.
whisper-1Legacy transcription model. Kept for agents already configured with it.

Speaking style

How the assistant sounds when it talks, in either conversation style:

ValueBehavior
Everyday (default)Conversational, with an occasional "hmm" or filler. Formality follows the persona.
ExpressiveLivelier and warmer: reacts to the user and uses fillers more often.
FormalPolite, businesslike spoken answers with no fillers or slang.

The persona's mission and personality from Behavior also shape how the agent speaks during a call. Atlas's persona is "precise, calm, and dependable" and businesslike, so Everyday already sounds professional. Choose Formal if your agent should never use fillers at all.


Conversation flow

Turn-by-turn only. Under Speech recognition in the Call flow section, with Stop assistant on interruption in the Advanced fold. These settings are hidden when Natural conversation is selected: it has no turn detection to tune. It listens the whole time and decides for itself when to speak and when to yield. Its equivalents are under Natural Conversation Settings.

Also called turn-taking or voice activity detection (VAD), this is what tells the system when the user is speaking and when they've finished, so it knows when it's the AI's turn to respond. Without good turn detection, the assistant may cut users off mid-thought, or respond to background noise and filler sounds like "hmm" as if they were a real turn.

Call flow section labelled Turn-by-turn: Speech recognition gpt-transcribe, then Detect when the user has finished Off (notices natural pauses in speech), Noise filter with a slider from Faster response to Ignores noise near the right end, and the closed fold Advanced for interruptions and push-to-talk

Detect when the user has finished

Off by default. When on, turn detection is model-based rather than purely energy-based: it notices natural pauses in speech instead of waiting a fixed amount of time, which is better at telling real speech apart from filler sounds and background noise. The info mark next to the label calls this semantic VAD.

Response speed

Shown when Detect when the user has finished is on. It controls how eagerly the model ends the user's turn: Automatic (default), Slower, Balanced, or Faster. Faster responds sooner but may interrupt more.

Noise filter

Shown when Detect when the user has finished is off. It is a slider from Faster response on the left to Ignores noise on the right. Moving it right ignores more background noise; moving it left responds faster, at the cost of more false interruptions. New agents start near the right end, which is deliberately hard to talk over: background noise and speaker echo are ignored, and the assistant tends to finish its sentence. If you want an agent that is easy to interrupt, turn on Stop assistant on interruption and move the slider toward the middle.

Stop assistant on interruption

In the Advanced fold. Off by default. When on, the assistant stops speaking the instant the user starts talking: responsive, but it may cut off short replies like "ok" or "mhm". When off, it waits about a second before stopping, so it won't cut off short replies like "ok".


Push-to-talk

Turn-by-turn only. In the Advanced fold of the Call flow section, together with Stop assistant on interruption. Hidden when Natural conversation is selected. A Natural conversation call always listens, so it never shows a hold-to-talk button β€” even for an agent that had push-to-talk on before switching.

Push-to-talk requires users to hold down a button while they speak: the microphone only streams audio while it's held. This guarantees background noise or other conversations can never be mistaken for a real turn, which is especially useful in noisy environments or when extra privacy is needed. Atlas leaves it off, but if your users call from a busy project site with machinery running, it is worth turning on.

Require hold-to-talk

Off for new agents. When on, a hold-to-talk button is shown in the voice call UI, and users can switch at runtime between it and automatic listening. When off, the call always runs in automatic mode: the assistant listens continuously, so background noise may be picked up as speech.

Start calls in

Only shown when Require hold-to-talk is on. Which mode a voice session opens in:

ValueBehavior
AutomaticThe server listens continuously and detects turns using the Conversation flow settings above.
Hold-to-talkThe user must hold the talk button to be heard; nothing is streamed otherwise.

When you turn Require hold-to-talk on, this switches to Hold-to-talk automatically; you can change it back to Automatic before saving. Users can switch anytime during the call.


Check the result

  1. Click Save all changes.
  2. Start a voice conversation in the preview (+ menu next to the chat input).
  3. Ask something your agent knows, for example "Who signs the permit for work above two metres?" for Atlas. Listen for the voice and speaking style you chose. For Turn-by-turn, try pausing mid-sentence and talking over the assistant to see whether the turn-taking feels right; adjust Response speed, Noise filter, or Stop assistant on interruption if it cuts you off or ignores you.
  4. If you turned on push-to-talk, check that the hold-to-talk button appears and the call opens in the mode you chose.

Next steps