Overview
Voice configuration controls speech recognition, spoken replies and phone-session behavior for an Environment. Agent instructions, Playbooks and tools still define what the agent does.
Examples of when to use Voice configuration
- Reduce early interruptions: tune turn-taking for customers who pause while reading order numbers.
- Serve different languages: choose a voice for each conversation
locale. - Collect an account number: enable phone keypad input and define how the caller submits digits.
- Bound phone sessions: set silence and maximum-duration limits.
Choose how to configure Voice
- In Studio — configure the controls and verify a phone call.
- Through the API — update the same Environment settings from your backend.
Speech recognition
Choose how customer speech becomes text.
Speech recognition uses the selected Environment’s speech-to-text provider and language. In Behaviour → Voice input, choose the provider before its model and language; available controls belong to that provider.
For a custom audio client, the input format is negotiated separately from speech-recognition settings. Changing the provider does not convert client microphone frames.
Use the Studio configuration guide or API configuration guide.
Turn-taking and interruptions
Balance complete customer turns with timely replies.
Endpointing determines when the agent treats speech as a finished turn. Interruption controls determine when customer speech can interrupt a reply. These solve different problems.
- Replies start too early: review the selected recognizer’s end-of-turn or silence controls.
- Replies wait too long: test a shorter wait with pauses, dates and account numbers.
- Agent audio continues after interruption: in a custom client, stop and clear playback on
asr.barge-in.asr.interruptalone only reports speech detection.
Provider-specific settings are in the Voice settings reference. Verify changes over the actual phone or browser connection.
Speech output
Choose the voice customers hear and how responses play.
Behaviour → Voice output selects the default voice. Its provider determines the available voice and model controls.
- Sync audio and text output keeps text aligned with speech; enabling it can add latency.
- Pronunciation dictionary controls how selected terms are spoken.
- Background audio and Audio cue add the configured background or waiting sound.
A custom client still owns decoding, playback order and playback acknowledgements.
Language overrides
Use a different voice for each conversation language.
A language override replaces the default voice when the conversation’s locale matches that override. Select the voice for each configured language in Voice output → Language overrides.
Phone and widget sessions initialize locale. API integrations must set it for their session. This selects a voice; it does not translate the agent’s instructions or select the recognizer’s input language.
Follow Set voices by language for Studio setup or configure localeOverrides through the API.
Keypad input
Collect phone digits with an explicit completion rule.
In phone projects, Voice input → Keypad input configures digit collection. Choose a completion key (# or *), an inter-digit timeout, or both. When no timeout is configured, retain a completion key so the caller can submit.
Keypad collection is active while the agent or input step listens. Where the step exposes Listen for other triggers, leave it on; turning it off disables collection there. The current runtime submits at 20 digits even if the delimiter or timeout has not fired. The Environment setting alone does not create a question or store a result. A custom telephony client must implement the dtmf instruction itself.
Use the configuration walkthrough to configure and test collection on a phone call.
Silence and call duration
End unattended or overlong phone conversations deliberately.
Session & timeout contains the phone silence timeout, maximum conversation duration and optional End message. The duration limit applies to the whole call; a silence timeout addresses a call waiting without input.
Set an End message when callers need an explanation before the maximum duration ends the call. Studio’s Inactivity timeout uses session.idleTimeout.timeoutMs. The runtime prefers an enabled positive session timeout; voice.silenceTimeoutMs is a fallback. voice.maxDuration.seconds controls the whole-call duration. Test each condition independently.
Returning callers
Choose whether variables carry over between calls.
Remember callers across calls controls runtime identity for phone sessions. When enabled, the caller’s variables can carry into a subsequent call. When disabled, each call receives a new runtime identity.
The callerIdentity API value is phone-number or call. This is separate from the caller’s presented phone number; neither authenticates the customer. Verify account access through your own identity checks before using private customer data.