ElevenLabs voice models: Turbo v4, Conversational v3, Flash v2.5 and Multilingual v2

/agents → your agent → Voice Pipeline → Synthesized Voice. When the selected voice is an ElevenLabs voice, its row carries a Model: control with four choices. The voice stays the same. The model decides how that voice is performed: how lively it sounds, how quickly it starts talking, which languages it can speak, and what each character costs.

The control appears only on the Standard (STT + LLM + TTS) pipeline with an ElevenLabs voice. Realtime pipelines bring their own voices, and the other text-to-speech providers have no model to choose.

The four models

“First audio” is the time from the agent’s text being ready to the first sound of its voice. We measured it on live test calls through the real phone path, with the same agent, voice, script and language model on every call. Only the voice model changed. Each call had two English turns and two Chinese ones.

• Turbo v4 was the fastest on every reply. Across 21 replies it ranged from 0.09 to 0.12 s.
• Flash v2.5 came second, at 0.11–0.13 s.
• Conversational v3 came last, at 0.16–0.20 s.
• The order never changed between English and Chinese. None of these models starts slower in one language than in another.

Multilingual v2 is in a different class. It took about half a second to start on production calls, and over a second in direct tests. That gap is long enough to hear.

What the choice does, and does not do, to speed

The voice model is a small part of how long a reply takes. On the calls above, a whole reply mostly took 1.3–1.7 s, from the caller stopping to the agent’s first sound. The voice model’s share of that was 0.1–0.2 s. If your agent feels slow, switching between Turbo v4, Flash v2.5 and Conversational v3 moves it by less than a tenth of a second.

The exception is Multilingual v2. It adds a noticeable pause before every answer, so do not use it on a live phone agent unless the sound matters more to you than the wait.

Replies in Chinese did take longer than replies in English, but the voice was not the cause. The language model took a little longer to begin its answer. The pause detector was also more likely to decide a Chinese caller had not finished speaking. Changing the voice model will not close that gap.

Which to pick

Leave it on Turbo v4 unless you hear a reason to change. It is the fastest to start, speaks the most languages, and is the most expressive. Choosing nothing gives you Turbo v4.

Try Flash v2.5 if Turbo v4 sounds like too much. Expressive cuts both ways. On our calls Turbo v4 came across as someone trying hard, and Flash v2.5 as a safe bet that cannot go far wrong. For an agent reading out order numbers and opening hours all day, plain may be what you want.

Try Conversational v3 for something in between. It sounds natural, a little calmer than Turbo v4. It is also the slowest of the three fast models to start speaking, by less than a tenth of a second.

Keep Multilingual v2 for agents where the sound has to be its best. It has the richest audio of the four, costs twice as much per character, and makes every caller wait longer for each answer. It was built for narration, not for conversation.

Listen before you choose, then ring it

Under the selected voice, Hear this voice on plays the same voice on each model: a short greeting, in that voice’s own language. Play all four before you decide. The differences are in the delivery, and you hear them faster than you can read about them.

Then call the number from a real phone. A phone line carries far less of the sound than your speakers do. Some of what separates Multilingual v2 from the others on a laptop is lost on a call. The liveliness of Turbo v4 is not.

Languages each model cannot speak

Turbo v4 and Conversational v3 speak every language an agent can be set to. The older two do not:

• Flash v2.5 cannot speak Persian, Nepali or Thai.
• Multilingual v2 cannot speak Persian, Nepali, Thai or Vietnamese.

Under a model that cannot speak your agent’s language, the picker says “Doesn’t speak” followed by the language’s own name (for a Thai agent, “Doesn’t speak ไทย.”), and you cannot choose it.

The agent’s other languages count too. If your agent answers in English but can switch to Thai mid-call, Flash v2.5 could not speak once it switched. So the call uses Turbo v4 instead of your pick, even though English alone would have allowed it. A pick that cannot cover every language the call might need is not honoured.

Cantonese: only Turbo v4 lists Cantonese as a language of its own. The other three list Chinese. For an agent that should sound Cantonese, stay on Turbo v4.

Four things that happen underneath

No pick means “follow the default”. An agent with no model chosen uses the platform default, which is Turbo v4 today. If the default changes later, that agent changes with it. Choosing Turbo v4 by hand is saved the same way as choosing nothing. To stay on a model whatever the default becomes, choose one of the other three.

An agent switched to “Multilingual” earlier is still on Multilingual v2. Before this control existed, the only choice was a Multilingual switch for higher quality, and that switch meant Multilingual v2. Agents that had it turned on keep it, with its pause before every reply. They are not moved to Turbo v4. If you turned it on for one of yours, listen to Turbo v4 and decide whether the wait is still worth it.

Switching to a voice from another provider clears the pick. Only ElevenLabs has models. Pick an Azure or OpenAI voice and the model choice is dropped. If you come back to ElevenLabs, the agent starts on the default again.

Octo cannot change the model. Octo can change the voice and the pipeline for you, but not this. Set it on the agent page and press Save.

What it costs

Turbo v4, Conversational v3 and Flash v2.5 are billed at the same rate per character spoken. Multilingual v2 is billed at twice that rate. Only the words the agent speaks are billed, so a long call on Multilingual v2 costs twice as much in voice as the same call on any of the other three. The language model and the phone line cost the same on every model.

Models you will see elsewhere but not here

ElevenLabs’ own site lists more models than this picker shows. They are left out on purpose.

• Eleven v4 and Eleven v3, the full versions, cost twice as much as Turbo v4 and Conversational v3. ElevenLabs built the Turbo and Conversational versions for live conversation, which is what a phone agent does.
• Turbo v2.5 is an older model at the same price as Flash v2.5. An agent that still carries it uses the default instead.
• The English-only v2 models speak one language. Turbo v4 already covers English, along with the 93 other languages.