Expressive mode

Overview

Expressive mode enables agents to deliver speech that reflects intent, emotion, and emphasis, adapting in real time to how users sound and what they say. It is built on two system-level improvements to the conversational stack:

1. Eleven v3 Conversational — the most emotionally intelligent, context-aware Text to Speech model available in ElevenAgents.
2. A new turn-taking system — more accurately timed responses with fewer interruptions.

Expressive mode is enabled by default when you select Eleven v3 Conversational as your agent’s TTS model.

Eleven v3 Conversational

Eleven v3 Conversational is an ultra-low-latency version of Eleven v3, optimized for live, back-and-forth dialogue. It maintains conversational context across turns and adapts delivery to match the tone and intent of each exchange.

• Context-aware delivery: Agents adapt tone based on conversational context — responding more calmly when a user sounds worried, or more directly when clarity matters.
• Explicit emotional control: Guide delivery through system prompt rules, from precise triggers to broader scenarios, to align with brand voice and compliance requirements.
• 70+ language support: Expanded from ~32 languages in Turbo and Flash models, with improved expressiveness in languages where nuance previously lagged, including Japanese.
• Expressive tags: The LLM can output tags like [laughs], [whispers], or [sighs] to control specific moments of delivery.

Turn-taking system

The new turn-taking system uses real-time signals from Scribe v2 Realtime — including emotional cues and speech patterns — to determine when an agent should speak, pause, or wait. This helps agents respond more naturally, especially in emotionally charged situations.

For example, “yeah” can be a complete acknowledgement or a lead-in to continue speaking. By analyzing how it was said (speech cues like prosody) in addition to the transcript, the system times the agent’s response more naturally.

Turn-taking behavior can be further tuned with the turn eagerness setting.

Configuration

Enabling expressive mode

System prompt examples

Guide your agent’s emotional delivery through natural language instructions in the system prompt. The model interprets these contextually, so explicit tags are not required for every situation.

Broad guidance

You are a customer support agent. When a user sounds frustrated or upset, respond
in a calm, reassuring tone. When delivering good news, allow your tone to reflect
genuine warmth. Maintain a professional but approachable delivery throughout.

Specific triggers

You are a conversational AI agent with expressive speech capabilities.
Tone guidelines:
- When a user expresses frustration, use a calm and empathetic tone
- When explaining technical steps, use a clear and measured pace
- When a user shares good news, respond with warmth and enthusiasm
- When handling complaints, remain composed and solution-oriented

Using expressive tags

In addition to context-aware delivery, the LLM can output explicit tags to control specific moments. Common tags include:

• [laughs] — Adds laughter to the speech
• [whispers] — Lowers volume for whispering
• [sighs] — Adds a sighing quality
• [slow] — Slows down speech delivery
• [excited] — Adds excitement to the delivery

Each tag affects approximately the next 4-5 words of speech before returning to normal delivery.

You can also use expressive tags in your responses for precise control:
- [laughs] for moments of humor
- [whispers] for confidential or intimate moments
- [sighs] for resignation or relief
- [slow] when emphasizing important information
Example: "That's great to hear! [laughs] I'm glad we could sort that out for you."

Best practices

Limitations

FAQ

Related features