"It says all the right things, but it just doesn't sound human."
That’s the very first reaction from almost everyone who deploys an AI phone agent.
Hey everyone, this is Sonetho. ⚡
We keep hearing the exact same feedback from people who built an AI phone agent after reading Part 1 (Getting Started) and Part 2 (Real Estate Office).
The script is airtight, yet the agent speaks in the exact same flat tone whether dealing with an upset customer or closing a deal.
When a voice fails to adapt in moments where any human naturally would, callers immediately think, "Just another bot," and tune out.
To address this exact issue, ElevenLabs introduced Expressive mode.
⚠️ Before we dive in, one quick clarification: this isn't a brand-new feature.
According to the official blog, it launched back in February 2026.
However, because the official documentation doesn't list dates, many people are discovering it for the first time, and it was by far the most common question we received while writing this series.
Here is how to configure Expressive mode.
🎭 1. Expressive Mode Changes Two Things at Once
This isn't just a simple toggle that injects emotion; it upgrades two parts of the conversational stack at the same time.
① Eleven v3 Conversational (The Speaking Side)
This is a low-latency model optimized for real-time conversation based on v3. The key difference is that it carries conversational context across dialogue turns.
If the caller sounds anxious, the agent responds in a calmer tone; when delivering critical details, it articulates more clearly and deliberately.
② A New Turn-Taking System (The Listening Side)
It decides when to speak and when to wait using real-time signals from Scribe v2 Realtime.
Instead of just parsing transcribed text, it factors in how things are spoken (intonation).
The example in the official documentation is intuitive:
A single word like "Yeah" can mean the caller has finished their answer, or it can be a signal that they're continuing their thought.
Text alone cannot distinguish between the two, but tone of voice makes it clear. The model uses this signal to time when to jump in.
That's why enabling Expressive mode delivers two improvements together: emotional inflection in the voice and fewer awkward interruptions.
⚙️ 2. How to Turn It On: Just Switch One Model
There is no separate toggle. Selecting V3 Conversational as your TTS model enables Expressive mode by default.
In the Dashboard
Open your agent, go to the Agent Voice tab, change the Text to Speech model to V3 Conversational, and save your changes. That’s all it takes.
Via CLI
elevenlabs agents pull --agent "<agent-name>"In the downloaded agent_configs/<agent-name>.json file, update conversation_config.tts.model_id like this:
{
"conversation_config": {
"tts": {
"model_id": "eleven_v3_conversational"
}
}
}elevenlabs agents push --agent "<agent-name>"
Via API
from elevenlabs import ElevenLabs
elevenlabs = ElevenLabs()
elevenlabs.conversational_ai.agents.update(
agent_id="YOUR_AGENT_ID",
conversation_config={
"tts": {"model_id": "eleven_v3_conversational"},
},
)
Get Started with ElevenAgents for Free →
📝 3. Direct Tone Through the System Prompt
If you only swap the model, the vocal range expands, but where and how emotion gets applied is left entirely to the model.
If your brand has a specific voice or customer service protocols, defining them in your system prompt is essential for consistency.
Method ①: Provide Broad Direction
You are a customer support agent. If the user sounds frustrated or upset,
respond in a calm, reassuring tone. When delivering good news, let genuine
warmth come through in your voice. Overall, maintain a professional yet
approachable tone.
Method ②: Define Rules by Scenario
Tone Guidelines:
- If the user expresses frustration: use a calm, empathetic tone.
- When explaining technical steps: speak clearly at a steady, measured pace.
- When the user shares good news: respond warmly and brightly.
- When handling complaints: stay calm and solution-oriented.
The official documentation notes that you do not need to add tags for every single situation.
Writing instructions in natural language as shown above gives the model enough context to interpret appropriately.
In production, Method ② is the safer route. If you have customer service guidelines, spelling out "which tone for which scenario" makes quality checks and audits much easier down the road.
🏷️ 4. Pinpoint Specific Moments with Tags
On top of context-based modulation, you can explicitly prompt specific delivery cues.
The LLM outputs these tags directly within its generated response.
[laughs]adds light laughter[whispers]lowers the voice to a whisper[sighs]adds a slight sigh[slow]slows down the speech rate[excited]delivers an excited, upbeat tone
★This is the single most common point of confusion. A single tag does not modulate an entire sentence.
Each tag only affects the next 4 to 5 words before reverting back to the base tone.
If you drop a tag at the very start of a sentence expecting it to carry all the way through, you won't get the result you want.
You need to place the tag immediately before the exact words you want to alter.
Here is the official example:
"That's great to hear! [laughs] I'm glad we could sort that out for you."
If you are new to audio tags, it's best to start with The Complete ElevenLabs Prompting Guide, where we cover v3 tags in depth.
We break down emotion and sound effect tags alongside pronunciation rules and pacing controls there.
⚠️ 5. Limitations to Review Before Enabling
The official documentation highlights three key limitations, and the first one is especially important.
① Professional Voice Clone (PVC) characteristics are not preserved
If you have invested time and resources into creating a PVC of your CEO or a professional voice actor, this factor will likely decide whether you adopt it.
Eleven v3 Conversational does not currently retain PVC voice nuances reliably.
The resulting audio may not sound like the original voice.
The official recommendation is explicit: if voice identity matters most, stick with Flash v2.
For now, you are faced with a trade-off between expressive emotional range or exact voice identity. You cannot have both just yet.
② Tag duration is strictly 4–5 words
As discussed above, tags are not designed to transform an entire long sentence.
③ Performance varies across voices and languages
Supported languages have expanded to over 70, up significantly from approximately 32 languages in the Flash line.
The documentation specifically highlights noticeable improvements in languages like Japanese, where conversational nuances were previously flatter.
However, the documentation also notes that expressive depth is not uniform across every language.
If you plan to deploy an agent in a language other than English, make sure to run several test calls in your target language with the exact voice you plan to use before rolling it out. A great-sounding English demo does not guarantee identical results across all languages.
💰 6. No Additional Cost
This question comes up frequently, and the answer in the official FAQ is a clear "No."
Eleven v3 Conversational shares the exact same base pricing as other conversational TTS models, starting at $0.08 per minute. There is no surcharge or premium fee for turning on Expressive mode.
You can estimate your monthly total based on expected call volume using the per-minute pricing calculator.
✍️ Wrapping Up
Here is a quick summary:
How to enable: Switch your TTS model to V3 Conversational. That’s all there is to it.
How to tune: Spell out scenario-specific tones in your system prompt.
How to pinpoint: Place audio tags immediately before the target phrase (they only last 4–5 words).
When to hold off: If you rely on PVC, compare both models side by side before deciding.
Thanks for reading! ⚡