Why latency matters in voice
On a call, silence is much more noticeable than in chat. A pause of more than half a second already feels like a delay, and one longer than a second can sound as if the line has dropped. Several settings in the Calls channel shape how responsive a call feels: the length of the agent’s opening line, the model that generates its answers, how fast the voice speaks and how much the agent tries to say in one turn. This guide shows how to tune them and check the result by ear.
What you will set up and what you need
You will tune an existing voice agent and confirm the result with a test call. You need:
- An agent with instructions and a knowledge base. The voice agent is built from both when you connect the channel.
- A Calls channel on that agent, connected through SIP Trunk, Twilio Account or Web SDK. If you don’t have one yet, start with Build a Voice Agent in 10 Minutes.
- A microphone and speakers for Make a test call, which works with every provider.
Step by step
- Open Channels, select your agent and open the Calls channel.
- Before you change anything, click Make a test call. Note where the pauses happen.
- Shorten the First Message. By default it uses the agent’s welcome message, usually written for chat. One short spoken sentence works better.
- Check Speech Speed. It ranges from 0.7 (slower) to 1.2 (faster), and the default is 1.2. Lower it only if callers struggle to follow.
- Review the LLM, the model that powers the agent’s responses during the call.
- Tighten the agent’s instructions for the phone: short, specific answers instead of long explanations. When the Sync prompt appears in the Calls panel, click Sync Agent to push the new instructions to the voice agent.
- Make another test call and compare it with the first one.
Settings that matter
- LLM: reasoning depth travels with the model. If you move to a faster, lighter model, the agent may lose its step-by-step behavior, so test it on the hardest questions your callers ask.
- Speech Speed: a faster rate means each answer takes less time to say, but callers still need to understand it.
- TTS Model: picked automatically for each language. Adding Additional languages switches the agent to the higher-quality v3 Conversational model, which bills at 5 AI messages per minute of call instead of 3 for standard voices. See Multi-language Voice Agents.
- Max Call Duration (minutes): a hard cap on call length, 10 minutes by default. Set it high enough that useful calls are not cut off.
- Web SDK client tools: whatever a handler returns is spoken back into the conversation, so return a short confirmation rather than nothing.
Common problems and fixes
- The call opens with a long speech. Replace the default welcome message in First Message with one short line.
- Answers run on and callers interrupt. Tighten the instructions: on the phone, short and specific beats long and vague. Then click Sync Agent.
- Your changes don’t show up on calls. New instructions only reach the voice agent after Sync Agent. The knowledge base is a snapshot taken when the channel was created, so changing which integrations feed the agent means creating the Calls channel again.
- Answers lost depth after you changed the LLM. The new model may not reason step by step. Choose a model tuned for reasoning and test again.
- The agent doesn’t answer at all. Make sure Connect agent to this channel is on. New Calls channels start switched off, and connecting a provider alone does not activate calls.
- A Web SDK page tool is never called. Tool parameters must be a flat map where each parameter name has a type, description and required flag. If you write them as a JSON Schema, the call still connects, but the agent never receives the tool.