Voice agents have one chance to keep a caller engaged. AssemblyAI-TTS-1, the company’s first text-to-speech model, is now generally available as a hosted API built for agents handling live phone calls, where latency and pronunciation can decide whether someone stays on the line.
The launch gives developers a new way to close the voice loop, especially teams already using AssemblyAI for speech-to-text. AssemblyAI-TTS-1 costs $15 per 1M characters, has no annual contract, and does not require developers to buy a concurrency tier.
A Voice API Built for Live Calls
AssemblyAI-TTS-1 runs on AssemblyAI’s cloud in the US region, us-east-1, or the EU region, eu-north-1. It is a hosted API model with no open weights and no on-prem option, so teams must use AssemblyAI’s infrastructure to run it.
Developers can connect through any client that supports HTTP or WebSocket. The service offers three interfaces: an HTTP API, a WebSocket API, and an OpenAI-compatible route. AssemblyAI also published reference agents for LiveKit and Pipecat, giving developers starting points for voice-agent integrations.
The WebSocket interface keeps one connection open for each conversation and accepts a cancel message for barge-in. That matters during phone calls, where a caller may interrupt an agent before it finishes speaking. Instead of waiting for the current response to end, the agent can cancel the output and respond to the new input.
Audio output comes in mono PCM, mu-law, or A-law at rates from 8 kHz to 48 kHz. The native format is 24 kHz PCM, while telephony systems can request 8 kHz mu-law directly. Word timestamps are included as part of the output.
Voice Cloning, Six Languages, and 4,000 Voices
The launch includes a library of 4,000+ voices and instant voice cloning. A clone requires a short recording and confirmation that the speaker gave permission. The clone shown in the launch post used 24 seconds of audio.
AssemblyAI-TTS-1 supports six languages at launch: English, Spanish, French, German, Italian, and Portuguese. That combination gives voice-agent teams a way to build phone experiences across several major markets without switching between multiple text-to-speech systems.
The model also targets the details that can break a real conversation. AssemblyAI tested it against five other models across four English sets, using 50 matched cloned voices. The evaluation included phone numbers, codes, and other difficult entities, not only ordinary sentences.
- Entity accuracy reached 93.7% across 580 hard targets.
- ElevenLabs v4 scored 93.6%, Cartesia Sonic 3.6 scored 82.5%, and ElevenLabs Flash v2.5 scored 68.0% on that measure.
- Word accuracy reached 99.0% on Seed-TTS and LibriSpeech sentences.
- Predicted MOS was 3.85, behind Inworld TTS-2 Flash at 3.88 and ahead of Sonic 3.6 at 3.82.
- The TTSDS prosody score was 85.5, while speaker similarity for cloned audio was 0.64.
AssemblyAI says TTS-1 tied ElevenLabs v4 on entity accuracy in its tests while costing under one-fifth of its price. All of those figures are self-reported from an evaluation dated October 28, 2026, so developers will still want to test the model against their own calls, voices, and telephony systems.
Latency Takes Center Stage
AssemblyAI’s API test measured a 47 ms median time to first byte across six models. In the same comparison, ElevenLabs Flash v2.5 took 119 ms, Cartesia Sonic 3.6 took 143 ms, and Inworld TTS-2 Flash took 244 ms.
An independent Cekura benchmark offers a different view of streaming performance. TTS-1 posted 159 ms median streaming latency, measured from the last word sent, placing eighth among 19 timed models in that test. On Cekura’s benchmark of 27 models, it reached 81.5% reading accuracy and 4.50 naturalness.
Inworld TTS-2 Flash posted 45 ms and 93.2% accuracy on that benchmark at the same $15 price. AssemblyAI also cites a 1.7% word error rate on Coval’s leaderboard and a top-10 debut in Artificial Analysis’ controlled-voice arena.
Those results point to a model designed around the practical pressure of voice agents: start speaking without delay, pronounce critical details correctly, and keep the conversation moving. For a caller, the difference between a smooth response and a long pause can shape the entire experience.
AssemblyAI-TTS-1 is available now as a generally available hosted API. With HTTP, WebSocket, and OpenAI-compatible access, direct telephony formats, voice cloning, and pricing at $15 per 1M characters, AssemblyAI is positioning the model as a single-vendor path from incoming speech to spoken response.
Based on
- AssemblyAI Ships AssemblyAI-TTS-1: A $15 per 1M Character Text-to-Speech API for Voice Agents — marktechpost.com
- Block’s new voice model sounds eerily human | TechCrunch — techcrunch.com
- Agent Assurance by TestMu AI Is Now Generally Available, Letting Teams Test What Their AI Agents Actually Change | AFP.com — origin-zh9p7m-www.afp.com
- DeepMind’s Genie 4 world model adds audio and more realistic people to simulations | Ars OpenForum — arstechnica.com
- The Harness Makes The Agent — forbes.com
