Configure real-time voice agents

[This article is prerelease documentation and is subject to change.]

Configure a real-time voice agent by turning on real-time voice, setting core options, and then configuring features like topics, multilingual support, DTMF, and silence detection.

Set up and enable real-time voice

1. Create a new agent, and configure its basic details, such as a descriptive name and the purpose of the agent in the description.

2. Go to the agent's Voice settings and turn on Enable voice, and then in Voice type, select Realtime voice. Learn more in Choose how to handle speech.

   Important

   This is a one-time selection. After you select Realtime voice, you can't switch back to Basic voice. To use Basic voice, create a new agent.

   

3. Go to the agent's Security settings and select No Authentication.

Knowledge and tools

You can configure your agent to use knowledge and tools. Learn more in Knowledge sources summary, Add tools to custom agents, and Tools, knowledge, MCP, and API.

Nested agents (preview)

Real-time voice agents only support child agents.

Topics

Real-time voice agents support all topics configured in Copilot Studio. Use topics to define deterministic behaviors such as greetings, business rules, and escalation, while the real-time voice model manages conversational responses at runtime. Learn more in Choose how to control the conversation.

Best practices when using topics with real-time voice agents

• Use topics only when deterministic behavior is required.

• Use static text in greeting messages for the fastest first response. Dynamic messages with variables and expressions increase initial latency.

• Conversation Start is enabled by default. If you want the real-time voice model to handle the greeting, disable the Conversation Start topic; otherwise, the greeting configured in the Conversation Start topic plays instead of the voice model greeting.

• Let the real-time voice model handle general conversation and follow-up questions.

• The On Error topic should include an explicit action, such as transfer or end call. Message‑only error handling isn't sufficient. Without a deterministic next step, customers might experience silence or stuck calls, leading to confusion and poor voice experiences.

• Use explicit topic and tool descriptions to declare ownership of data collection. Learn more in Write effective topic and tool descriptions.

Topic node support

The following list describes topic support in real-time voice agents:

Condition node

Message node

Question node

HTTP node

Tool node

Set variable value node

Topic management node

Transfer conversation node

Advanced

System trigger support

Pass variables between topics and the language model

When you use topics in a hybrid conversational flow, understanding how to pass variables between topics and the real-time language model is critical for building reliable, stateful interactions.

This functionality works through the following process:

• You pass input variables defined on a topic into the topic at invocation time, so the language model can provide structured data to the deterministic flow.

• You return output variables defined on a topic to the language model at the end of topic execution as structured key-value pairs. The language model includes these outputs in the conversation context, and you can reference them in subsequent turns.

• Tool call outputs follow the same pattern: you send outputs to the language model at the end of tool execution, and they're available for future use within the conversation context window.

• The language model is populated with conversational context, including tool call output key-value pairs. However, you only return explicitly defined output variables as structured data. You can collect a value inside a topic, such as a verified account number. Define that value as an output. If you don't, the language model can't access it. The agent might ask the caller for the same information again later.

Learn more in Manage topic inputs and outputs.

Multilingual support

Add all secondary languages you want. Localization strings aren't required for any real-time flows. However, for deterministic topic messages, you need to provide the translated messages. Learn more in Configure and create multilingual agents.

The real-time model can understand and respond in many languages. However, Microsoft doesn't formally validate all languages for general availability.

As of April 2026, the following languages are formally validated:

• English (United States) (en-US)
• Spanish (United States) (es-US)
• Arabic
• Portuguese (Brazil) (pt-BR)
• Italian (Italy) (it-IT)
• German (Germany) (de-DE)
• Dutch (Netherlands) (nl-NL)
• French (Canada) (fr-CA)

Microsoft continues to validate other languages and adds them after certification completion. You can add any language supported by Copilot Studio. However, languages that aren't fully certified for GA-level quality should be thoroughly tested before production deployment.

Context variables

A real-time voice agent supports context variables that allow it to behave more intelligently by carrying information about the call, the caller, and the current conversation. The system automatically provides a limited set of call and conversation context to the model at runtime. This set includes:

For all other context variables, follow the steps outlined in Configure context variables for agents.

Agent voice

Select the voice your agent uses by selecting your agent and go to Settings > Voice > Select voice. Real-time voice agents support the following voices:

• Alloy
• Ash
• Ballad
• Coral
• Echo
• Sage
• Shimmer
• Verse
• Marin
• Cedar

Speech sensitivity

Speech sensitivity voice activity detection (VAD) determines when the agent should respond after the caller finishes speaking.

Understanding VAD types

Real-time voice agents support two VAD approaches:

Server-based VAD - Based on sound (silence)

• Detects end of speech based on audio signals (silence duration, volume)

• Responds quickly once silence is detected

• Deterministic, predictable behavior

• Best for structured interactions, short responses, noisy environments

Semantic VAD - Based on sentence context

• Determines turn completion based on meaning of what was said

• Evaluates whether caller completed their thought

• Adapts to natural pauses, filler words, trailing speech

• Best for: Conversational interactions, complex questions, open-ended discussions

Select the right VAD

Use server-based VAD when all of the following conditions are true:

• Interactions are structured (IVR-style menu navigation).

• Responses are short and predictable.

• Background noise is a concern (semantic VAD might wait too long).

• You want fast, crisp turn-taking.

Use semantic VAD when all of the following conditions are true:

• Conversations are open-ended.

• Callers might hesitate or use filler words ("um", "let me think...").

• Questions are complex (callers explain situations).

• Natural conversation flow is prioritized.

Configure server-based VAD

Go to Settings > Voice > Phone Setup > Speech input > Sensitivity > Based on sound (silence).

Threshold

• Lower (0.3-0.4): More sensitive; picks up quiet speech, might trigger on background noise.

• Higher (0.7-0.9): Less sensitive; requires louder speech, reduces false triggers.

• Recommended: Start with 0.5; increase if background noise causes false triggers.

Prefix padding

• Captures audio before speech detection (prevents cutting off first word).

• Lower (200 ms): Faster response; might miss first syllable.

• Higher (500 ms): Safer capture; slight delay.

• Recommended: 300 ms (good balance).

Silence Duration

• How long the caller must be silent before agent responds.

• Lower (500 ms): Fast turn-taking; might interrupt if caller pauses mid-thought.

• Higher (1000 ms): More patient; might feel slow.

• Recommended: Start with 750 ms.

Configure Semantic VAD

Go to Settings > Voice > Phone Setup > Speech input > Sensitivity > Based on sentence context.

Parameter: Eagerness (how quickly the agent responds after semantic completion)

DTMF configuration

Dual-Tone Multi-Frequency (DTMF) allows callers to enter information by using their phone keypad.

You can turn on DTMF for your agent at both the topic and global levels. To set it at the global level, select your agent and go to Settings > Voice > Conversation behavior > DTMF.

To set it per topic node, learn more in Turn on DTMF support for your voice-enabled agent.

To support reliable input completion, you can configure DTMF timing and termination behavior. This configuration includes an inter‑digit timeout, which defines how long the system waits between key presses, and an optional termination character (such as # or *) that explicitly signals the end of input. When you use a termination character, the system processes input immediately without waiting for a timeout.

Silence detection

Silence detection allows real-time voice agents to recognize when a caller provides no input for a specified period. Set up silence detection as a global voice setting for the agent by going to Settings > Voice > Conversation Behavior > Silence detection.

The silence timer starts when the agent finishes speaking and detects no speech or DTMF input from the caller. If the silence timeout is reached, the agent follows the configured silence detection topic.

Latency messaging

Add latency message or music to your agent when background operations take longer than expected. To configure latency messaging, go to Settings > Voice > Conversation behavior > Latency messaging.

Real-time voice agent evaluation (preview)

Real-time voice agents support sending text during evaluation, however, audio processing isn't supported.