Speech settings

Speech Settings controls how the agent sounds and how it behaves around the caller's speech. This page covers each control in the section. Two of them, Responsiveness and Interruption Sensitivity, have their own page: Turn-taking. Pronunciation is on Pronunciation and expressive tags on Expressive mode.

All fields on this page are on the agent. Set them in the editor or with Update agent.

Background Sound#

Plays a background noise bed under the agent's voice, so a call sounds like it comes from a real place and silence feels less empty.

Background sound volume (ambient_sound_volume, 0 to 2, default 1) appears once a sound is chosen.

Voice Speed, Voice Temperature and Volume#

See Voice for how speed and temperature are applied per provider.

Dynamic voice speed#

Dynamic voice speed (enable_dynamic_voice_speed) matches the caller's pace during the call.

• It measures the caller's speaking rate in words per minute and nudges the agent's speed towards it, by at most 15% either way from the agent's Voice Speed.
• When the caller says "speak faster", "slow down" or similar, the agent changes speed in steps of 0.15.
• The resulting speed stays between 0.7 and 1.2.

Dynamic responsiveness#

Dynamic responsiveness (enable_dynamic_responsiveness) adapts how long the agent waits to each caller. It is covered on Turn-taking.

Enable Backchanneling#

With Enable Backchanneling (enable_backchannel), the agent says short acknowledgements while the caller is speaking at length, the way a person says "mm-hmm" on the phone.

• Nothing is said in the first 2.5 seconds of a caller's turn.
• Backchannel frequency (backchannel_frequency, 0 to 1, default 0.8) sets the minimum gap between acknowledgements: 1 means about every 3 seconds of caller speech, 0.1 about every 12 seconds.
• Backchannel words (backchannel_words) sets the phrases, used in turn. Leave empty for the defaults: "mm-hmm", "yeah", "okay", "right", "I see".
• Backchannels are spoken audio only. They are not added to the conversation the model sees.

Use it for calls where callers tell long stories, such as support or intake. Leave it off for short transactional calls.

Reminder Frequency#

When the caller goes quiet, the agent can nudge them. Reminder Frequency reads "N seconds, up to M times":

The model writes the reminder itself: it is told the caller has been silent and to check briefly whether they are still there, or to repeat its last question, in one short sentence.

Reminders are separate from End Call on Silence, which hangs up after a much longer silence. See Call settings.

Normalize for speech#

Normalize for speech (normalize_for_speech, default on) asks the voice provider to expand numbers, dates and currency into words before speaking. It is applied on Premium and Standard voices. For other voices, ask the model to write numbers as words, or turn on Speech normalization in the Agent Handbook.

Pronunciation#

The pronunciation dictionary is at the bottom of the section. See Pronunciation.

Fields#

{
  "ambient_sound": "call-center",
  "ambient_sound_volume": 0.6,
  "enable_backchannel": true,
  "backchannel_frequency": 0.5,
  "backchannel_words": ["mm-hmm", "right", "got it"],
  "reminder_trigger_ms": 8000,
  "reminder_max_count": 2
}

Per-node overrides in flows#

Conversation flow nodes can override voice_speed, reminder_trigger_ms and reminder_max_count while the node is active. See Conversation flow.