AssemblyAI launches TTS-1 voice model at $15 per 1M characters

Built for voice agents that callers do not hang up on

AssemblyAI frames the launch around two problems. The first is caller trust: according to the company, people who suspect they are talking to an IVR phone tree tend to press 0 and ask for a representative. A voice agent therefore has to sound natural and answer quickly enough that the caller stays on the line and realizes the system can actually help.

The second problem is cost. AssemblyAI argues that a text-to-speech bill that grows with every completed call competes with the business case for the agent itself. It says AssemblyAI-TTS-1 costs about half of what customers pay today, and calls it the lowest-priced frontier text-to-speech model for voice agents.

On the technical side, the company says the model was trained on millions of hours of audio, including real conversational data. Most text-to-speech models represent audio as discrete tokens, which AssemblyAI says leads to lower quality and higher latency. AssemblyAI-TTS-1 instead operates directly on continuous audio representations.

Pricing, concurrency and voices

AssemblyAI says there is no spend commitment and no concurrency tier to buy. In its words, ten concurrent calls and ten thousand run on the same API at the same per-character price. The company says its speech-to-text customers already get unlimited concurrency without an annual contract, and TTS-1 launches the same way.

The voice library holds more than 4,000 voices, selectable in the dashboard. The model also ships with instant voice cloning: users upload a recording, confirm they have the speaker's permission, and the cloned voice is ready a few seconds later.

• Supported accents cover English, Spanish, French, German, Italian and Portuguese.
• More languages and accents are planned in the coming weeks.
• "Pro" (fine-tuned) voice cloning is not yet supported.
• Custom pronunciation dictionaries and emotion tagging are coming soon.

AssemblyAI adds that the model natively expresses audio matching the input text, and that prompting techniques reliably produce pauses and disfluencies. It also says many developers have found higher call success rates when the agent uses a voice with a different accent than the target caller.

Benchmark results according to AssemblyAI

AssemblyAI compared the model with Cartesia Sonic 3.6, Inworld TTS-2, Inworld TTS-2 Flash, ElevenLabs v4 and ElevenLabs Flash v2.5 on four English sets: Seed-TTS (443 short sentences), a pronunciation set (580 sentences, each built around one hard target such as a phone number or URL), a conversation set (500 customer turns from recorded calls) and LibriSpeech (165 audiobook passages). Every model used the same pool of 50 cloned voices.

Latency. Measured over each provider's API from the same client, AssemblyAI-TTS-1 starts streaming in 47 ms at the median. ElevenLabs Flash v2.5 follows at 119 ms, Cartesia Sonic 3.6 at 143 ms, Inworld TTS-2 Flash at 244 ms, Inworld TTS-2 at 280 ms and ElevenLabs v4 at 612 ms. At the 90th percentile, AssemblyAI-TTS-1 is at 54 ms.

Accuracy. The model scores 99.0% word accuracy on plain sentences. On the pronunciation set it reads 93.7% of hard targets correctly, against 93.6% for ElevenLabs v4, 82.5% for Cartesia Sonic 3.6 and 68.0% for ElevenLabs Flash v2.5.

Naturalness. Predicted listener preference comes to 3.85 on a 1 to 5 scale. That is ahead of Cartesia Sonic 3.6 at 3.82, ElevenLabs Flash v2.5 at 3.70 and ElevenLabs v4 at 3.67, and a few hundredths behind Inworld TTS-2 Flash at 3.88. The company also reports 85.5 on the normalized TTSDS prosody realism test and 0.64 on speaker similarity for cloned audio.

AssemblyAI also points to Cekura's text-to-speech benchmark, which reads 104 agent sentences through 27 models. There the model lands in the "Best Quadrant" of accuracy against latency, ranks #1 on phone numbers and codes at 91.7%, and places #4 on naturalness at 4.50.

What it means for developers

For teams running voice agents at scale, the pitch is a single per-character price with no concurrency negotiation, plus latency under 100 ms. The model can be tried in a playground with an existing AssemblyAI API key. Since most of the comparison figures come from AssemblyAI's own evaluations, developers may want to test the model with their own voices and call content, especially given the missing Pro voice cloning and the controllability features that are still to come.