AssemblyAI-TTS-1: $15 per Million Characters and 90 ms to First Audio

AssemblyAI, a company known for speech-to-text, has launched a voice of its own. In a post dated October 27, 2026, the company says AssemblyAI-TTS-1, its first text-to-speech model, is generally available: "frontier voice naturalism at ultra-low latency, with audio streaming in 90 ms," at $15 per 1M characters, with unlimited concurrency and no annual contract.

If you build voice agents, the interesting part is not another natural-sounding voice. It is the pricing structure, because TTS cost grows with every completed call. That is the argument AssemblyAI makes, and it is also where you should check the details. Below: what is announced, what comes from AssemblyAI's own tests, what third parties show, and how to test it on your own calls.

TL;DR: AssemblyAI-TTS-1 at a glance

Why a speech-to-text company is building TTS

A voice agent is three models in a row: it listens, thinks and speaks. AssemblyAI already sold the listening part, with models such as Universal-3.6 Pro Realtime. Adding the speaking part lets it sell the full loop, and lets it argue on price. Its pitch is that callers hang up on anything that sounds like an IVR phone tree, so a voice must sound natural and respond quickly, and that a bill rising with each call undercuts the business case for the agent.

Voice is a crowded category this month. ElevenLabs launched Eleven v4 and v4 Turbo and two speech-to-speech models for voice agents. Google shipped Gemini 3.8 Flash TTS with voice design in 100 languages. OpenAI has GPT Live API for voice agents, and Mistral has Voxtral 3 with a voice agent SDK. On the open side, there is Gradium's 31M-parameter Phonon. A new closed API has to win on something concrete, and AssemblyAI chose price, concurrency terms and latency.

What AssemblyAI says about how it works

The post says the model was trained on "millions of hours of audio, including real conversational data". The architectural claim: most TTS models represent audio as discrete tokens, which AssemblyAI says leads to lower quality and higher latency, whereas AssemblyAI-TTS-1 "operates directly on continuous audio representations". It does not publish parameter counts or a paper in the post we read, so that claim cannot be independently checked from the page.

Two practical details help builders: the model streams from the first byte, supports a cancel message for barge-in (when a caller interrupts), and returns word timestamps, so it should drop into an existing turn loop. The post says to clone its LiveKit or Pipecat reference agent and point the TTS stage at the new model.

The pricing claim: $15 and no concurrency tiers

At $15 per 1M characters, AssemblyAI says it costs "about half of what you pay today". Its independent-benchmark section cites Cekura's table, in which it says the lowest ElevenLabs price is $40 and the lowest Deepgram price is $30 per 1M characters. Those are AssemblyAI's quoted figures; check the providers' current price pages, since plans differ and volume pricing exists.

A quick way to size it: characters are counted including spaces and punctuation. A rough rule is about 15 characters per second of speech, so one minute of agent speech is on the order of 900 characters. At $15 per million that is roughly 1.4 cents per minute of agent speech, before the speech-to-text and language model costs. Treat that as an estimate from our arithmetic, not an AssemblyAI figure, and measure your own characters per call.

The second commercial claim is concurrency: "Ten concurrent calls and ten thousand run on the same API at the same per-character price." For contact-centre style workloads with spiky traffic, concurrency tiers are often what drives real cost, so this matters more than a few cents per million characters.

Latency: 90 ms and 47 ms are different numbers

The post quotes two figures, and both are AssemblyAI's.

• 90 ms is time to first audio in production, which AssemblyAI calls the conservative end across real networks and both regions.
• 47 ms is the median time to first byte in a lab test: one request at a time from a single client, 30 prompts, three rounds, four stock voices per model. Next fastest was ElevenLabs Flash v2.5 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.

Time to first byte is not the full story for a phone call. Your LLM's time to first token and the telephony hop add to it, and total response time depends on those. The right test is your own pipeline, measured end to end.

Quality: what is vendor-run and what is independent

AssemblyAI's own head-to-head used four English sets (Seed-TTS with 443 sentences, a 580-sentence pronunciation set, 500 recorded customer turns and 165 LibriSpeech passages) and 50 cloned voices shared across models. Its results:

Note two things. The accuracy is scored by AssemblyAI's own speech-to-text model, and the naturalness score is an automated predictor, not human listeners. On entity accuracy, the gap to ElevenLabs v4 is 0.1 points, which is a tie.

Three outside benchmarks are cited, and we have not verified them against their sites:

• Cekura: per AssemblyAI (see the Cekura TTS leaderboard), a "Best Quadrant" placement on accuracy against latency, 91.7% on phone numbers and codes, and 4.50 on naturalness (#4), across 27 models.
• Coval: a 1.7% word error rate, three-tenths of a point behind the leader.
• Artificial Analysis: a controlled-voice arena where listeners compare models speaking through the same eight cloned voices; AssemblyAI says it debuted in the top ten. Check the Artificial Analysis page for AssemblyAI TTS directly, since arena ranks move.

The honest read: it is competitive on accuracy and probably cheap, and not clearly at the top on listener preference. For voice agents that is a reasonable trade.

What is missing at launch

AssemblyAI is candid about gaps. At launch there is no fine-tuned "Pro" voice cloning, no SSML or speed parameter, and no custom pronunciation dictionaries or emotion tagging, which are "coming soon". The model does follow written punctuation: a question mark gets a rising tone, an ellipsis a pause, and a written "um" is spoken. Language coverage is six languages, far fewer than the roughly 100 offered by Gemini 3.8 Flash TTS, so multilingual deployments should check coverage first.

Voice cloning: a consent and fraud point

Instant cloning from a short recording is useful for branded agents and also a fraud tool. AssemblyAI says you must confirm you have the speaker's permission before a cloned voice is made, and the post shows a clone built from 24 seconds of audio. That is a self-attestation, not a technical safeguard. If you ship cloned voices, keep written consent and consider disclosure to callers. Our coverage of Baroness Campbell's AI voice in the House of Lords shows the legitimate end of the same technology.

How to test it on your own traffic

A microphone linked to a speech bubble, the loop a voice agent runs when testing text to speech

AssemblyAI itself says the final arbiter is what callers prefer on production traffic, and suggests running the model on a slice of live calls and comparing completion rates.

1. Pick 50 to 100 real scripts from your agent logs, with numbers, names, codes and URLs.
2. Generate with each candidate using the same voice type or a matched clone.
3. Transcribe and score entities with one neutral ASR, not a vendor's.
4. Measure time to first audio end to end in your pipeline, including the LLM and telephony.
5. Run a 5 to 10% live split and compare call completion and hang-up rates.
6. Price it on your own character counts, including retries and any silence padding.

For the surrounding stack, see our reading of Meta's Muse Voice Transcribe and the Grok Voice 3.0 speech-to-speech launch.

Bottom line

AssemblyAI-TTS-1 is a credible, price-led entry into voice agent TTS, announced with unusually open caveats about missing features. The 90 ms, 47 ms and quality tables are the company's own; the independent arenas place it in the top tier for accuracy, not at the top for naturalness. If your agent speaks millions of characters a month, the pricing and concurrency terms alone justify a test.

Related reading

• ElevenLabs Eleven v4 and v4 Turbo
• ElevenLabs Eleven Realtime for voice agents
• Gemini 3.8 Flash TTS
• GPT Live API for voice agents
• Mistral Voxtral 3 voice agent SDK
• Gradium Phonon open on-device TTS
• Official: AssemblyAI-TTS-1 announcement

Details are accurate as of October 28, 2026 and come from AssemblyAI's launch post; prices, rankings and features may change.