Speech Settings for Voice Agents

The Speech tab (left sidebar, in the agent editor) controls the audio side of a call: how the voice sounds, when the caller can interrupt the agent, and how the agent reacts to background noise or speech it couldn’t understand.

Speech tab showing Speech speed, Number pronunciation, and Speech formatting — Speech tab

Not every field below appears for every agent. The voice sliders depend on the voice’s model, picked on the Prompt tab. Voice Detection and Smart Turn Detection are hidden for realtime models (GPT Realtime and GPT Realtime Mini), which handle audio end to end.

Voice

Setting	What it does	Default
Speech speed	How fast the agent talks.	1.0 (range 0.5 to 2)
Voice consistency	Move it left to fix skipped words, right to fix repeated words.	-
Voice similarity	Higher values match the reference voice more closely; lower values give more flexibility.	-
Voice enhancement	Higher values give cleaner audio at the cost of a little latency.	-
Number pronunciation	The language numbers and special characters are read out in. Auto follows the language the agent is speaking. Pick a language to read numbers that way instead, for example 123 read out as a Hindi speaker would say it.	Auto
Speech formatting	Applies punctuation, paragraphs, and entity formatting (dates, times, numbers) to transcripts to make them easier to read.	On, where supported
Background sound	Ambient audio layered under the agent’s voice: Office, Cafe, Call Center, or Static.	None

Which voice settings you see depends on the voice’s model:

Voice model	Settings shown
Lightning v3.1, Lightning V3.1 Pro	Speech speed, Number pronunciation
Lightning Large, Lightning v2	Speech speed, Voice consistency, Voice similarity, Voice enhancement

Speech formatting only works for agents whose primary language is English or Hindi. Smart Turn Detection, below, has the same restriction.

Pronunciation dictionary

Fix words the voice mispronounces, such as names, brands, or domain-specific terms.

Click Add Pronunciation and fill in two fields:

Field	Description
Word	The word as it normally appears
Pronunciation	How you want it said instead

There’s no phonetic-alphabet requirement. Type it however reads naturally (e.g. Cerebras → sir-EE-brass). Entries show up as removable chips, and you can edit or delete them at any time.

The voice itself, and the languages your agent speaks, are chosen on the Prompt tab, not here. The Speech tab only tunes how that voice behaves.

Turn-Taking & Interruptions

Controls for when the agent lets the caller speak, and when it lets itself be interrupted.

Setting	What it does	Default
Allow interruptions	When off, the caller’s audio is suppressed while the agent is speaking (half-duplex). Useful for agent-to-agent calls or very noisy environments.	On
Wait for user to speak first	The agent skips its first message and waits for the caller to speak.	Off
Mute user until first bot response	Mutes the caller until the agent finishes its first response.	Off
Interruption backoff timer	A short grace period (in seconds) right after the agent starts talking, during which the caller can’t interrupt it. Set to 0 to disable. Helps stop both sides talking over each other.	0 (range 0 to 10 s)
Smart turn detection	Uses a model to predict when the caller has actually finished speaking, instead of only waiting for silence. Cuts down on premature replies. Adds a little latency.	See below

Mute user until first bot response and Wait for user to speak first are mutually exclusive. Turning the second one on turns the first one off.

Turning on Smart turn detection reveals a Wait time slider (1 to 10 s): how long the agent waits, when it isn’t sure the caller has finished, before responding anyway.

Listening

How the agent filters and interprets incoming audio before it reaches the language model.

Voice detection

Tunes how confidently the system decides “this sound is the caller speaking.” Shown only for non-realtime models.

Field	What it does	Default
Confidence	How strict the system is when deciding if a sound is speech. Higher is less sensitive to background noise.	0.7
Min volume	Minimum loudness required to be treated as speech.	0.6
Trigger time	How long to wait after speech is detected before treating it as the caller speaking.	0.2 s
Release time	How long to wait after speech ends before the agent starts responding.	See below

Pushing Confidence or Min volume to the extremes (below 0.1 or above 0.9) can make the agent miss the caller entirely, or trigger on background noise.

Setting	What it does	Default
Personal info redaction (PII)	Redacts personally identifiable information from call transcripts.	On
Denoising	Filters out background noise before it reaches the speech pipeline. Reduces false triggers from environmental sound.	Off
Handle unrecognised speech	If the agent can’t hear the caller clearly, it asks them to repeat instead of going silent.	On

PII redaction removes personal details from transcripts, so turning it on can make transcripts less complete for QA.

Customizing the 'please repeat' response

When Handle unrecognised speech is on, you can write your own responses for when the agent can’t understand the caller (e.g. “Sorry, could you repeat that?”). Leave it empty to use the built-in, language-aware responses, which is the recommended default. Duplicate responses aren’t allowed, and clearing a response’s text deletes it.

Turn-taking defaults

New agents usually start with the values in the Standard column. In some organizations, new agents start with the second set instead. These defaults apply only when an agent is created; existing agents keep their values. Check the Speech tab to see what your agent has.

Setting	Standard	Some organizations
Smart turn detection	Off	On
Wait time	3 s	1.5 s
Release time	0.8 s	0.2 s

These are only starting values. Changing them on one agent doesn’t affect your other agents.

Voicemail

Setting	What it does	Default
Voicemail detection	Detects when a call has reached voicemail instead of a live person.	Off
Voicemail end text	What the agent says before hanging up once voicemail is detected. It’s spoken exactly as written: `{{variables}}` aren’t filled in here, so don’t use them.	A default sign-off message, filled in when you turn detection on

Prompt

Pick your agent’s voice, model, and languages

Agent settings

Timezone, STT model, session length, and silence handling