Introducing AssemblyAI-TTS-1: text-to-speech built for voice agents
Natural speech for voice agents, streaming in 90 ms. Unlimited concurrency, no annual contract, $15 per 1M characters.
AssemblyAI-TTS-1, our new state-of-the-art text-to-speech model, is now generally available. It gives voice agent developers frontier voice naturalism at ultra-low latency, with audio streaming in 90 ms. At $15 per 1M characters, it is the lowest-priced frontier text-to-speech model for voice agents, and the only one to offer unlimited concurrency without an annual contract.
Callers hang up on anything that sounds like an IVR
Voice agent adoption is exploding. Businesses are moving their existing phone call volume to voice agents that can successfully complete calls, and the models underneath those agents have become accurate and fast at understanding the caller, taking the right steps, and streaming audio back over the phone.
The challenge is overcoming the caller’s expectation that an automated system cannot do what they called to do. For most people, the moment they suspect they’re talking to an IVR phone tree, they pound the 0 key and repeat “representative.” A voice agent must sound natural and respond quickly so the caller stays on the line long enough to realize they’re talking to a new generation of technology that can actually help them.
The second problem is cost. Teams that have already solved naturalness are now paying for it at scale, and a text-to-speech bill that grows with every completed call is a bill that competes with the business case for the agent itself. AssemblyAI-TTS-1 is built for both: it sounds like a person, it starts speaking in under 100 ms, and at $15 per 1M characters, it costs about half of what you pay today.
Natural speech with audio streaming in under 100 ms
We built AssemblyAI-TTS-1 to be the most natural-sounding and fastest text-to-speech model on the market. 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 leads to lower quality and higher latency. AssemblyAI-TTS-1 operates directly on continuous audio representations, which improves the realism of the generated speech and minimizes time to first spoken audio.
AssemblyAI-TTS-1, voice: Amira
Generated audio
So you were saying the latency on your current provider is what, one second? One full second? Oh, come on. That’s not latency. That’s a pause for dramatic effect.
Measured over each provider’s API from the same client, AssemblyAI-TTS-1 starts streaming audio in 47 ms at the median, the fastest of the six models in our head-to-head below, where the next fastest starts at 119 ms and the slowest at 612 ms.
By the metrics, AssemblyAI-TTS-1 scores 99.0% on word accuracy (transcribed audio compared to the input text), 93.7% on entity accuracy (the audio includes and correctly says numbers, codes, abbreviations, and the like), 3.85 out of 5 on average across preference prediction models like UTMOSv2, 85.5 on the normalized TTSDS prosody realism test, and 0.64 on the speaker similarity of cloned audio to the reference recording. The head-to-head below breaks each of these out model by model.
Unlimited concurrency without an annual contract
At AssemblyAI, we’re experts in high-performance inference. Unlike other providers, we never hand the task of serving our speech-to-text models back to our customers once they reach scale, and we have never required a spend commitment per concurrent session. Our speech-to-text customers get unlimited concurrency without an annual contract with us.
AssemblyAI-TTS-1 launches the same way. There is no spend commitment and no concurrency tier to buy. Ten concurrent calls and ten thousand run on the same API at the same per-character price.
AssemblyAI-TTS-1, voice: Aaron
Generated audio
Yeah. So concurrency? No, there’s no limit on pay as you go. Yeah, no, no, we’ve heard. I get it. You’re used to getting on a call with some bloke called Brett to go from five streams to eight. Yeah, we don’t do that. Bit mad, really.
4,000+ voices, instant voice cloning, and native emotion
The voice is the user interface of a voice agent. Our voice library has more than 4,000 voices, so voice agent developers can select the perfect voice for their use case in the dashboard. In many cases, developers prefer to clone a voice rather than select a provided one, so AssemblyAI-TTS-1 ships with instant voice cloning.
Many voice agent developers have found higher call success rates when their agent uses a voice with a different accent than the target caller. AssemblyAI-TTS-1 currently supports various accents for English, Spanish, French, German, Italian, and Portuguese. In the coming weeks, we will be adding more languages and accents.
As of the launch date, AssemblyAI-TTS-1 does not yet support “Pro” (fine-tuned) voice cloning. Controllability features such as custom pronunciation dictionaries and emotion tagging are also coming soon, but the model natively expresses audio that matches the input text, and prompting techniques reliably produce pauses and disfluencies.
How it compares
For an agent, a word error is a wrong digit in a confirmation number or a dropped character in an email address, and the caller has to ask again. We tested AssemblyAI-TTS-1 against Cartesia Sonic 3.6, Inworld TTS-2, Inworld TTS-2 Flash, ElevenLabs v4, and ElevenLabs Flash v2.5 on four English sets. Seed-TTS is 443 short read sentences. Our pronunciation set is 580 sentences built around one hard target each, such as a phone number, a date, a URL, or a confirmation code. The conversation set is 500 customer turns from recorded calls, and the LibriSpeech set is 165 audiobook passages of 15 to 30 seconds. Every model spoke through the same pool of 50 cloned voices, so the voices are matched and the differences below come from the models. Each figure is a mean across the sets where the measure runs, with 95% bootstrap intervals described in the Methodology.
Accuracy
The pronunciation set is where the hard parts live: a phone number, a date, a URL, a confirmation code, a drug name. AssemblyAI-TTS-1 reads 93.7% of its targets correctly, ahead of ElevenLabs v4 at 93.6%, Cartesia Sonic 3.6 at 82.5%, and ElevenLabs Flash v2.5 at 68.0%. On the plain sentences of Seed-TTS and LibriSpeech it scores 99.0% word accuracy.
Entity accuracy — pronunciation set
Saying the hard parts right
Higher is better
Time to first audio
Measured over each provider’s API from the same client, AssemblyAI-TTS-1 starts streaming audio in 47 ms at the median, the fastest of the six. ElevenLabs Flash v2.5 follows at 119 ms, then 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 still at 54 ms, so the slow requests are not much slower than the typical one.
Time to first byte — median, over the API
Who starts speaking first
Lower is better
Naturalness
Naturalness is harder to put a number on than accuracy, and the final judge is a caller who stays on the line. The closest proxy we have is predicted listener preference, an automated estimate of the mean-opinion score a listening panel would give on a 1 to 5 scale. AssemblyAI-TTS-1 scores 3.85, ahead of Cartesia Sonic 3.6 at 3.82, ElevenLabs Flash v2.5 at 3.70, and ElevenLabs v4 at 3.67, and within a few hundredths of Inworld TTS-2 Flash at 3.88. That is the model speaking from a cloned voice it had never heard before, with no tuning for the panel.
Independent benchmarks
Our own evals are not the only ones. Three independent voice AI benchmarks have scored AssemblyAI-TTS-1 against the field. On Cekura’s text-to-speech benchmark, which reads the same 104 agent sentences through 27 models, AssemblyAI-TTS-1 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. It scores higher than every ElevenLabs and Deepgram model there on reading accuracy (81.5%, against a best of 80.4% for ElevenLabs v4 and 77.4% for Deepgram Flux), on phone numbers and codes (92%, against a best of 89% and 87%), and on naturalness (4.50, against a best of 4.42 and 4.43), at $15 per 1M characters against a lowest of $40 for ElevenLabs and $30 for Deepgram.
On Coval’s text-to-speech leaderboard, which re-measures every model daily over a rolling 30 days, AssemblyAI-TTS-1’s word error rate is 1.7%, three-tenths of a point behind the leader and ahead of ElevenLabs v4 Turbo and Cartesia Sonic 3.6 at 1.8% and Deepgram Aura-2 at 1.9%. And on Artificial Analysis’s controlled-voice arena, where listeners compare models speaking through the same eight cloned voices, AssemblyAI-TTS-1 debuts in the top ten, ahead of ElevenLabs v3.
Pricing
AssemblyAI-TTS-1 is billed per character of input text. The list price is $15 per 1M characters. Every voice, every language, cloning, word timestamps, and both the US and EU data-residency regions are included at the same rate. There are no per-voice fees, concurrency fees, or minimums, and volume pricing is available for committed usage.
Talk to AssemblyAI-TTS-1 in your voice agent
To get started, sign up for an AssemblyAI account, open the Text-to-Speech playground, select a voice, write what you want the model to say, and listen to the generated speech. For developers building voice agents on open-source voice agent platforms, clone our LiveKit reference agent or Pipecat reference agent to get started with adding AssemblyAI-TTS-1 to your voice agent.
The final arbiter of text-to-speech quality is what callers prefer in a voice agent experience, and that can only be measured on production traffic. Run AssemblyAI-TTS-1 on a slice of your live calls next to your current voice and compare completion rates. Our team will help you set that evaluation up.
Frequently asked questions
What does $15 per 1M characters include?
Everything. Every voice in the library, every language, instant voice cloning, word timestamps, and processing in either the US or the EU are included at the one rate. You are billed on the characters you send, including spaces and punctuation. There are no per-voice fees, concurrency fees, or minimums on pay-as-you-go, and volume pricing is available for committed usage.
Do I need a contract to get more concurrency?
No. Concurrency is unlimited on pay-as-you-go, the same way it is for our speech-to-text models. There is no spend commitment to sign and no concurrency tier to buy. Ten concurrent calls and ten thousand run on the same API at the same per-character price.
Which languages and accents does it speak?
English, Spanish, French, German, Italian, and Portuguese at launch, with a range of accents in each. Set the language on the request and pick a voice, and your agent speaks with that accent from the first word. More languages and accents are coming in the weeks after launch.
How does voice cloning work?
Upload a recording in the dashboard, confirm you have the speaker’s permission, and the cloned voice is ready to use by ID a few seconds later. The clone in this post was made from 24 seconds of audio. Cloned voices cost the same per character as voices from the library.
Can I control pronunciation, emotion, or speaking speed?
Not with dedicated controls yet. Custom pronunciation dictionaries and emotion tags are coming soon, and the model does not accept SSML or a speed parameter today. It does read punctuation and spelling as written, so a question mark gets a rising tone, an ellipsis gets a pause, and a written “um” comes out as one. Prompting your agent’s LLM to write a pause or a hesitation is enough to put one in the audio.
How do I add it to an agent I already run on LiveKit or Pipecat?
Clone our LiveKit or Pipecat reference agent and point the TTS stage at AssemblyAI-TTS-1. The model streams audio from the first byte, supports a cancel message for barge-in, and returns word timestamps, so it drops into an existing turn loop without changes to the rest of your pipeline.
Where does my data get processed?
In the US or the EU, your choice, at the same price. The streaming and HTTP endpoints both have a US host and an EU host, and the voice library and cloning are available in both regions.
Methodology
Test sets. Seed-TTS English (443 rows), the AssemblyAI pronunciation set v0.2 (580 rows, 20 per entity class), customer turns from recorded calls (500 rows, including all 223 turns that contain an entity), and LibriSpeech medium with punctuated and cased text (165 rows). Every row is spoken through one of a fixed pool of 50 cloned voices, half female and half male, the same voice for every model.
Word accuracy. Generated audio is transcribed with Universal-3.5 Pro and compared against the input text after normalization, on Seed-TTS and LibriSpeech, equal-weighted.
Entity accuracy. On the pronunciation set, the share of target entities spoken in an accepted form, scored from a spoken-form transcript.
Predicted MOS. An automated predictor of human mean-opinion-score ratings on a 1 to 5 scale, averaged across preference models including UTMOSv2.
TTSDS prosody. How closely the generated set’s prosody distributions match real recordings, on a 0 to 100 scale.
Speaker similarity. Cosine similarity between speaker embeddings of the cloned output and the reference recording.
Time to first byte. Measured over each provider’s streaming API from a single client, one request at a time: 30 prompts, three rounds, four stock voices per model, from the request to the first byte of audio.
Intervals. 95% bootstrap over texts, 1,000 resamples, resampling voices. Models scored on the same rows share resamples. Eval dated October 28, 2026.