I launch developer products, and build with them.

I felt the best way to show I fit this role was to do the job: take the two newest launches, build a working demo, and write what developers need.

I picked Eleven v4 Turbo and turn prediction, because I hit the problem they solve while building property-caller, a voice agent in beta, for a property agent who plans to pay for it and offer it to his colleagues. See it live (work in progress)

A mock-up, written as if I ran this launch.

Positioning

For developers building realtime voice agents: Eleven v4 Turbo and turn prediction give your agent a human sense of timing. It knows when you've finished, and answers before the pause feels long.

Problem

Timing decides whether a voice agent feels like a conversation, and it isn't the only tell.

• People reply within about 200 ms. They predict the end of a turn instead of waiting for silence.¹
• Most agents wait for silence. Set it short and they cut in; set it long and there's dead air.²
• Accent and emotion matter too. Listeners judged the same synthetic statements more credible in one accent than another,³ and added emotion moved them less in a synthetic voice than a human one.⁴

Features chosen

Selected two releases from the last three weeks of the ElevenLabs changelog.

• Eleven v4 Turbo: the realtime variant of Eleven v4; audio starts before the sentence ends.
• Turn prediction: predicts when someone has finished instead of timing silence; turn_v3 in ElevenAgents, turn_prediction on realtime Speech to Text.

Audience

• Voice-agent builders on the API. They feel the cut-offs and the wait most. "Deploy intelligent voice agents for interactive applications." (elevenlabs.io/api)
• Teams automating calls: inbound support, scheduling and reminders. "Automate scheduling, reminders, collections, and internal ops." (elevenlabs.io/agents)
• Telephony platforms that embed ElevenLabs voices, e.g. Twilio, put them behind their own customers. "Developers can start using ElevenLabs' voices in Twilio ConversationRelay today." (ElevenLabs blog)

Value props

01Speaks while it's still thinking.

Your LLM's words stream into Eleven v4 Turbo over the Text-to-Dialogue WebSocket as they're written. Once about eight words have arrived, audio starts coming back, so the agent is already talking while the rest of the reply is still being written. ElevenLabs puts median inference latency at about 100 ms (changelog).

With Eleven v4 Turbo

Starts speaking once the first few words arrive, while the rest is still being written.

Without streaming

Waits for the whole sentence, then speaks.

02Hears the end of a thought, not just silence.

People pause mid-thought to recall a number, check a diary or find a word. A silence timer can't tell that pause from the end of a turn, so it cuts in. Turn prediction uses a model that estimates whether the speaker has finished, and replies only when it predicts they have.

In ElevenAgents that's the turn_v3 turn model; on realtime Speech to Text it's commit_strategy=turn_prediction, and partial transcripts carry a turn_probability (changelog).

On my own agent, a short silence setting cut in on 18 of 18 hesitations I tested.⁵

With turn prediction

Waits through the pause, replies after "I think".

Silence timer

Takes the pause as the end of your turn, and talks over you.

03Sounds like you.

Eleven v4's expressive range, its emotion, emphasis and natural pacing, now comes in the realtime model. It works with a voice you clone yourself, so your agent can sound like your brand, or like you, without giving up speed. The race below runs v4 Turbo in a clone of my own voice.

Model picker

Eleven v4 Turbo: about 100 ms median inference, 90+ languages, the most expressive delivery for agents. Flash v2.5: about 75 ms, 32 languages, lower price per character (models). The voice race (04) compares them on first and full audio.

Moving from Flash v2.5: change model_id to eleven_v4_turbo, move to the Text-to-Dialogue WebSocket, and send flush after each sentence.

Success metrics

• Activation: X% more developers make a first streaming call to eleven_v4_turbo than the pre-launch average.
• Real use: Y% of new callers still call v4 Turbo four weeks later, in production traffic, not tests.
• Adoption: share of realtime text-to-speech traffic on v4 Turbo.
• Time to first call from a human docs visit.

X and Y are set with real baseline data.

The best way to show the value props is a demo where developers can hear the difference and decide for themselves. Press play to hear when each voice starts and how it sounds, including the rounds v4 Turbo doesn't win.

You've heard the speed and the timing. Here's how they come together in real calls, one inbound and one outbound, for two fictional businesses. Talk to each for up to 60 seconds, and try pausing mid-sentence.

See ElevenLabs pricing for what an agent costs.

I hope this page shows how excited I am about ElevenLabs' mission: making content work in any language and any voice, and making talking to software feel like talking to a person.

Here I positioned two new releases for the builders who feel the wait most, built an agent you can talk to in my own voice, benchmarked it against four other providers with the losses left in, and set up two use cases you can call.

I'd love to talk to the team about what's next.