Voice
This section is about how your agent speaks and listens – the TTS voice it uses, how it handles audio, what it does mid-conversation, and how it transcribes what callers say.
It is not about how customers reach your agent. For that, see Phone (Numbers and Web Calling) and Chat.
Raven produces responses that sound natural when spoken aloud. Recommended for any voice deployment.
What lives here
The Voice page has five tabs:
Supporting pages:
How to think about it
Start with Choosing a good voice, then configure the voice in Voice settings, and tune call-runtime behavior in Advanced voice settings.
Programmatic voice configuration
You can also configure voices programmatically using the voice class inside functions – for example, selecting a voice based on conversation context, caller preferences, or other runtime variables.
Voice conversation style guide
These guidelines help your voice agent sound natural rather than robotic. They focus on the linguistic patterns that make spoken conversations feel human.
Social presence markers
Natural conversation includes patterns that acknowledge conversational history and participants. These contribute to a sense of collaboration rather than rote routine-following.
Use progressive tense for active collaboration:
- “I’m not seeing any accounts under that phone number…” conveys active collaboration
- “I don’t see any accounts” sounds too definitive
Reference shared context implicitly – don’t restate what both parties already know:
- “How about Wednesday instead?” (not “How about Wednesday instead of Tuesday?”)
- “In that case, how does Saturday at 2:30 sound?” (not “Since you said you prefer weekends…”)
Vary confirmationals – use a mix of “Great,” “Okay,” “Perfect,” and “Sure” rather than repeating the same one.
Use conversational datives for a collaborative feel:
- “Could you read me your account number?” rather than “Could you read your account number aloud?”
- “Can you log into your account for me?” rather than “Can you log into your account?”
Use face-saving past tense when referencing a user’s request:
- “When were you trying to come in?” rather than “When are you trying to come in?”
Avoid over-explaining
LLMs tend to justify every action in a way humans don’t. Most of the time, the important information and the request can be formed into a single sentence:
- “No problem, what’s your account number?” rather than “To check for outages, I’ll need to look up your account. Could you tell me your account number?”
Walkthrough conversations
When giving multi-turn walkthroughs, don’t end every step with “let me know when you’ve done that.” Provide the instruction and wait – the user will confirm on their own.