Voice conversations are different from text-based conversations. Callers expect quick responses, short answers, and natural turn-taking. Design voice flows specifically for spoken conversations instead of directly converting an existing text flow.

​

Before you start

Before publishing a voice agent, make sure you have:

• A supported voice provider and phone number.
• A voice agent configured in Bot Builder.
• A flow designed for spoken conversations.
• Tested prompts and responses that are short and easy to understand when heard aloud.

​

Telephony

WotNot currently supports Twilio and SIP as the voice channel provider. You can obtain a phone number from it and connect it to WotNot. To configure your voice number:

1. Go to Settings → Channels → Voice.
2. Enter your service provider details.
3. Configure the voice settings.
4. Save the configuration.

Once the number is connected, incoming calls to that number can be routed to your selected WotNot agent.

​

Web-based calling

You can also allow users to talk to your agent from your website using the Voice chat widget. To enable Voice mode:

1. Go to Settings → Channels → Web → Appearance.
2. Under Branding & Visual design, set Widget mode to VOICE.
3. Configure the voice options and default messages.

For more information, see Voice chat widget

​

Connecting a number

​

Provider Setup

To launch the chatbot you built for “Voice” channel, follow these steps:

​

Routing calls to an agent

After connecting the phone number, select the agent that should handle incoming calls.

1. Open Bot Builder and open the voice agent you want to publish
2. Open the Trigger block.
3. Select the number from the dropdown.
4. The agent assigned to this number will handle customer calls.

The configured agent will handle incoming calls received on the connected number.

​

Configuring the voice

​

Voice selection

Choose the voice that your agent uses when speaking to callers. Go to Settings → Channels → Voice→ Settings→ select the voice that best fits your use case. When selecting a voice, test it with the type of content your agent will normally speak. A voice that sounds good

​

Turn timeout (s)

The amount of time the AI agent will wait for a user reply before moving on to the next step in the chat flow. During testing, check:

• Whether the agent detects that the caller has started speaking.
• Whether the agent stops or continues speaking after an interruption.
• Whether the caller’s new response is correctly recognized.
• Whether the conversation resumes from the correct point. Callers may start speaking before the agent has finished responding. Test how the

​

Call silence timeout (s)

The duration of user silence after which the call will automatically disconnect.

​

Max conversation duration (s)

The maximum duration allowed for a conversation with the AI agent.

​

Speech settings

Voice conversations work best when the agent speaks naturally and uses short sentences. When configuring prompts and responses:

• Avoid unnecessary filler.
• Keep sentences short.
• Spell out or explain uncommon abbreviations when needed.
• Avoid excessive punctuation or formatting intended for visual reading.
• Test product names, company names, technical terms, and abbreviations.
• Use pronunciation-friendly wording when the voice repeatedly mispronounces a term.

If a product name is consistently pronounced incorrectly, consider adding a pronunciation-friendly version of the term to the agent’s prompt or response rather than relying on the spelling alone.

​

Latency

Latency is especially noticeable in voice conversations. A caller hears every delay between finishing their sentence and hearing the agent respond. A voice turn typically involves several stages: Caller speech → Speech recognition → Agent processing → Knowledge/API calls → Language model → Speech synthesis → Agent response A delay at any stage can result in noticeable silence. Reduce unnecessary processing: Design the flow to minimize work that must happen before the agent can respond. Consider these practices:

• Keep prompts short and focused.
• Avoid unnecessarily long agent responses.
• Reduce the number of sequential integration or API calls.
• Avoid multiple dependent calls when the same result can be obtained in one operation.
• Use a scripted response when the answer is fixed instead of performing a knowledge base lookup.
• Keep knowledge base searches focused on cases that actually require dynamic information.
• Avoid asking the caller to repeat information unnecessarily.
• Test external integrations for response time before using them in a live voice flow

​

Publishing the agent

Once the phone number, agent, and flow are configured, publish the agent.

• Open the voice agent you want to publish.
• Open the Trigger block and confirm that the correct voice phone number is configured.
• Review the flow and confirm that responses are suitable for spoken conversations.
• Check that integrations, functions, and other required blocks are configured correctly.
• Click Deploy to publish the agent.
• Call the configured phone number and test the published agent.

​

Testing

Always test a voice agent by placing a real call before sharing the number with users. During the call, check the following:

• Latency
• Speech recognition
• Pronunciation
• Pronunciation
• Silence and timeouts

​

Transcripts and handover

​

Voice transcripts

After a call, use the available conversation records and transcripts to review what happened during the interaction. Review transcripts when troubleshooting:

• Incorrect recognition.
• Missing information.
• Unexpected agent responses.
• Repeated questions.
• Failed handovers.
• Long or incomplete conversations.

For more information, see Voice transcripts.

​

Handover to a human

As human handover is not available for the voice call, provide an alternative such as collecting the caller’s details, creating a support request, or directing the caller to a supported channel.

​

Troubleshooting