On October 29, 2026, AssemblyAI quietly closed a gap in its own product line. The company that developers already knew for turning speech into text now does the reverse just as seriously: AssemblyAI TTS-1 is its first text-to-speech model, generally available from today, and it is built for one demanding job — giving live AI voice agents a voice that answers fast enough, and accurately enough, that callers stay on the phone. If you have ever hung up on a robotic phone agent that misread your order number, you already understand the problem TTS-1 is trying to solve.
This is not another celebrity-voice toy or a narration tool for audiobooks. AssemblyAI TTS-1 is a developer API priced at $15 per 1 million characters, with more than 4,000 voices, instant voice cloning from a short recording, and no concurrency limit to negotiate on its pay-as-you-go plan. In this guide, we explain what TTS-1 actually is, how it works, what it costs in real call volumes, how its benchmarks hold up once you look past the vendor’s own numbers, and how it compares with ElevenLabs, Cartesia, and Inworld — so you can decide whether it belongs in your voice stack.
Table of Contents
- What Is AssemblyAI TTS-1?
- Why Voice Agents Live or Die on Latency and Pronunciation
- How AssemblyAI TTS-1 Works Under the Hood
- Voices and Instant Cloning in AssemblyAI TTS-1
- AssemblyAI TTS-1 Pricing: What $15 per 1 Million Characters Really Buys
- AssemblyAI TTS-1 Benchmarks: The Honest Read, Vendor vs Independent
- AssemblyAI TTS-1 vs ElevenLabs, Cartesia, and Inworld
- Who Should Use AssemblyAI TTS-1 — and Who Should Not
- Getting Started With AssemblyAI TTS-1
- AssemblyAI TTS-1 Limitations You Should Know Before You Build
- The Verdict on AssemblyAI TTS-1
- Frequently Asked Questions About AssemblyAI TTS-1

What Is AssemblyAI TTS-1?
AssemblyAI TTS-1 is a streaming text-to-speech API from AssemblyAI, the speech-AI company best known for its transcription models. Feed it text — typically the reply your AI agent just generated — and it returns spoken audio in near real time, so a phone caller hears an answer instead of silence. It is a hosted API only: there are no downloadable weights and no self-hosted option. Your application talks to AssemblyAI’s cloud in the United States (us-east-1) or the European Union (eu-north-1), and the audio streams back as it is generated.
That hosting detail matters for two reasons. First, it keeps setup simple: any client that can make an HTTP or WebSocket request can use AssemblyAI TTS-1, with no GPU to provision and no model to serve. Second, it means you are buying into AssemblyAI’s infrastructure and uptime, not running TTS-1 yourself. Teams that need on-premise deployment for compliance reasons will find that door closed at launch — Cartesia, by contrast, offers on-premise and private-cloud deployment on its enterprise plans.
AssemblyAI trained AssemblyAI TTS-1 on millions of hours of audio, including real conversational speech rather than only clean studio recordings. That choice shows up in the product’s priorities: TTS-1 is tuned for the messy reality of phone calls, where an agent must read out a confirmation code, a street address, or a price exactly as written, in a voice that sounds like a person rather than a machine reading a script. Most text-to-speech systems represent audio as discrete tokens, a bit like compressing speech into a sequence of symbols before rebuilding it. AssemblyAI says TTS-1 works directly on continuous audio representations instead, and credits that design with both the natural sound and the short delay before the first audio arrives.
At launch, AssemblyAI TTS-1 speaks six languages: English, Spanish, French, German, Italian, and Portuguese. That is a deliberately narrow list compared with rivals — Inworld claims more than 200 languages, Cartesia 44, and ElevenLabs 32 — so language coverage is the first filter to apply. If your callers speak English or one of the other five launch languages, you are covered. If you need Hindi, Arabic, Japanese, or dozens of other languages, TTS-1 is not your tool today, and you should check the comparison section below before going further.
One more framing point is worth stating plainly, because it explains the whole product. AssemblyAI already sells the other half of the voice conversation: its Universal-3.6 Pro Realtime model transcribes what the caller says, fast. AssemblyAI TTS-1 completes that loop, so a developer can now buy both the ears and the voice of a phone agent from a single vendor, on a single bill. That pairing is the real pitch — not any single benchmark number, but the simplicity of one speech stack.
Why Voice Agents Live or Die on Latency and Pronunciation
A voice agent is not a chatbot with a speaker attached. In a text chat, a two-second pause reads as thinking. On a phone call, a two-second pause reads as a dead line, and the caller starts saying “hello? hello?” or hangs up. Engineers measure this in two related numbers you will see throughout this article, so it is worth defining them now.
Time to first audio is the gap between your application sending text and the first sound reaching the caller. AssemblyAI quotes about 90 milliseconds for AssemblyAI TTS-1. Streaming latency is the ongoing delay while audio keeps flowing, and independent benchmarks measure it differently — from the moment the last word is sent, for example — which is why the vendor’s number and an independent number for the same model can look like they describe different products. They do not contradict each other so much as measure different slices of the same pipeline, and the benchmarks section below shows both.
Pronunciation is the second make-or-break factor, and it is less obvious than speed. A voice agent that says “fifteen” when the order total is “fifty” has not made a cosmetic error; it has created a billing dispute. The hard cases are exactly the strings phone agents read out all day: phone numbers, confirmation codes, tracking numbers, prices, dates, and addresses. AssemblyAI’s own testing leans heavily on this point, with a dedicated set of 580 hard entity targets, and its headline accuracy claim — 93.7% on those entities — is the number to scrutinize most closely in the benchmarks section.
There is also a human factor that no latency chart captures. Callers forgive a slightly synthetic voice if the agent is quick, accurate, and easy to interrupt. They do not forgive an agent that talks over them, cannot be interrupted, or drones on while they are trying to correct it. That is why AssemblyAI TTS-1 includes a cancel message on its WebSocket connection for barge-in — the caller cutting in mid-sentence — and word-level timestamps, which let your application know exactly which word is playing when the caller interrupts. These are small features with outsized effects on whether a call feels human.
How AssemblyAI TTS-1 Works Under the Hood
Strip away the marketing, and AssemblyAI TTS-1 is a pipeline with three doors in and one stream out. Understanding those doors is the fastest way to judge whether it fits your stack.
Three Ways to Connect to AssemblyAI TTS-1: HTTP, WebSocket, and OpenAI-Compatible
The simplest door is the HTTP API: send text, receive audio. It suits one-shot jobs such as generating a voicemail greeting, a notification, or a short prompt. The serious door for live calls is the WebSocket API, which holds one connection open per conversation. Text streams in as your language model generates it, audio streams out as it is synthesized, and a cancel message stops playback instantly when the caller interrupts. The third door is an OpenAI-compatible route, which matters enormously in practice: teams already wired for OpenAI-style speech endpoints can point their existing code at AssemblyAI TTS-1 with minimal rewiring, rather than rebuilding their audio plumbing from scratch.
On the output side, AssemblyAI TTS-1 produces mono audio in PCM, mu-law, or A-law formats, from 8 kHz up to 48 kHz. The native format is 24 kHz PCM, which is the quality tier to use for app-based calls and recordings. Traditional phone systems, however, still run on 8 kHz mu-law, and TTS-1 can deliver that directly, so you do not need a separate transcoding step in your telephony stack. Word timestamps come with the stream, which powers live captions, karaoke-style highlighting in an app, and precise barge-in handling.
AssemblyAI TTS-1 and the Rest of the Voice Stack
No text-to-speech model works alone in a voice agent. The usual chain runs: speech-to-text hears the caller, a language model decides what to say, and text-to-speech speaks it. AssemblyAI now sells both ends of that chain, pairing AssemblyAI TTS-1 with Universal-3.6 Pro Realtime for transcription. For teams that prefer to assemble their own stack, AssemblyAI has also published reference agents for two popular open-source orchestration frameworks, LiveKit and Pipecat, so you can see a working TTS-1 integration instead of starting from a blank file.
There is also a managed option one level up. AssemblyAI’s Voice Agent API bundles speech-to-text, language-model reasoning, and voice generation behind a single WebSocket at a flat $4.50 per hour, billed per second. That bundle is a different product from AssemblyAI TTS-1 itself, and the distinction matters for pricing: if you want full control over which language model answers, you buy TTS-1 by the character and wire it yourself; if you want one bill and one throat to choke, you buy the bundle by the hour. Do not confuse the two when you compare prices with rivals.
Voices and Instant Cloning in AssemblyAI TTS-1
AssemblyAI TTS-1 ships with a library of more than 4,000 voices, spanning languages, accents, and regions, with new voices added over time. In practice, that number matters less than the audition workflow: you need to be able to hear a dozen candidate voices reading your actual scripts — appointment confirmations, not poetry — and pick the one that sounds right for your brand. A voice that sounds warm reading a welcome message can sound wrong reading a debt-collection reminder, and the library’s size is only useful if you test voices against your real prompts.

The headline feature is instant voice cloning. Give AssemblyAI TTS-1 a short recording — the clone in AssemblyAI’s own launch demonstration used just 24 seconds of audio — and it can speak new text in that voice. Two guardrails deserve emphasis. First, cloning requires confirmation of the speaker’s permission, which is the right design and also a legal necessity in several US states with right-of-publicity and deepfake laws. Second, cloning here is instant-only: you get a quick clone from a short sample, not the longer, studio-grade professional cloning process that Inworld and Cartesia offer for brand voices that must stay perfectly consistent across millions of calls. For a small business cloning its owner’s voice for an appointment-reminder agent, instant cloning is plenty. For a national brand whose voice is a trademarked asset, the professional tier elsewhere may still win.
A practical tip for anyone evaluating AssemblyAI TTS-1 voices: test entity-heavy sentences first. Have every candidate voice read a phone number, a six-character confirmation code, a price with cents, and a date, then listen back on an actual phone call, not studio headphones. Phone audio compresses and flattens speech, and a voice that aces clean audio can mangle digits over a real connection. This is also the fairest way to compare TTS-1 against a rival: same sentences, same phone, same ears.
AssemblyAI TTS-1 Pricing: What $15 per 1 Million Characters Really Buys
AssemblyAI TTS-1 costs $15 per 1 million input characters, and that single number covers a lot: every voice in the library, instant cloning, word timestamps, and processing in either the US or the EU region all come at the same rate. Billing counts input characters, including spaces and punctuation. There are no per-voice fees, no concurrency fees, and no minimum spend. AssemblyAI also charges no annual commitment to unlock that rate — you pay as you go, and your concurrency grows with your call volume instead of being a tier you must negotiate.
What does that mean in real money? A typical spoken English sentence runs roughly 15 characters per word once spaces and punctuation are counted, and conversational speech runs about 150 words per minute. That puts one minute of agent speech at roughly 2,250 characters, or about 3.4 cents at the AssemblyAI TTS-1 rate. A ten-minute customer-service call in which the agent speaks for half the time costs roughly 17 cents in voice generation. A business running 10,000 such calls a month would spend around $1,700 on TTS-1 speech — before transcription, language-model, and telephony costs, which are billed separately unless you buy the bundled Voice Agent API.
The honest comparison is not the headline rate alone. Inworld’s TTS-2 Flash also lists $15 per 1 million characters on demand, dropping to $7 on its Growth plan — cheaper per character than AssemblyAI TTS-1 for committed volume — but its on-demand tier allows only 5 concurrent requests, against unlimited concurrency on AssemblyAI’s pay-as-you-go plan. That concurrency gap is where the real cost hides: a contact center that spikes to hundreds of simultaneous calls during an outage or a product launch does not want its voice provider to become the bottleneck, or to pay a higher plan tier just to answer the phone. ElevenLabs Flash v2.5 lists $40 per 1 million characters, and Cartesia Sonic 3.6 works out to roughly $37 to $50 per 1 million through plan credits — both more than double the TTS-1 rate, though each bundles different plan features. AssemblyAI’s claim that TTS-1 is the lowest-priced frontier text-to-speech model for voice agents is therefore defensible on list price, with the Inworld asterisk attached: Inworld matches the $15 rate and undercuts it on a paid plan, but caps your concurrency unless you upgrade.
One more pricing warning applies to every provider in this category, including AssemblyAI TTS-1: character counting punishes verbose agents. An agent that says “Absolutely, I would be delighted to help you with that today, and I want to make sure we get this exactly right for you” before answering has burned three times the characters of one that says “Got it — checking now.” Prompt engineering is a cost control lever here, not just a quality lever, and the cheapest voice stack is often the one whose agent speaks least.
AssemblyAI TTS-1 Benchmarks: The Honest Read, Vendor vs Independent
Every AI launch comes with benchmarks, and every benchmark comes with a sponsor. AssemblyAI TTS-1 is no exception, so this section separates what AssemblyAI measured itself from what an independent firm measured, and tells you which numbers to trust for which decision.
AssemblyAI’s own evaluation, dated October 28, 2026, tested AssemblyAI TTS-1 against five rival models on four English test sets, using 50 matched cloned voices so the comparison would be apples-to-apples. The headline result: 93.7% accuracy on 580 hard entity targets such as phone numbers and codes, effectively tied with ElevenLabs v4 at 93.6% and well ahead of Cartesia Sonic 3.6 at 82.5% and ElevenLabs Flash v2.5 at 68.0%. Word accuracy came in at 99.0% on Seed-TTS and LibriSpeech sentences. Predicted mean opinion score — a model-estimated rating of how natural speech sounds — was 3.85, behind Inworld TTS-2 Flash at 3.88 and ahead of Sonic 3.6 at 3.82. Median time to first byte was 47 milliseconds, against 119 milliseconds for Flash v2.5, 143 milliseconds for Sonic 3.6, and 244 milliseconds for Inworld TTS-2 Flash. AssemblyAI also reports an 85.5 prosody score on the TTSDS test, 0.64 speaker similarity for cloned audio, a 1.7% word error rate on the Coval leaderboard, and a top-10 debut in Artificial Analysis’ controlled-voice arena.
Those are strong numbers, and they are also self-reported, on test sets AssemblyAI chose, using its own measurement setup. That does not make them false — vendors usually measure carefully, because rivals will check — but it does mean you should treat them as a best-case picture. The independent picture comes from Cekura, a benchmarking firm that timed 27 models. On Cekura’s tests, AssemblyAI TTS-1 scored 81.5% reading accuracy and 4.50 on naturalness (UTMOS, a standard automated quality metric), the highest naturalness score in that comparison — ahead of Inworld at 4.46, Cartesia at 4.41, and ElevenLabs Flash v2.5 at 4.33. But Cekura’s median streaming latency for TTS-1 was 159 milliseconds, measured from the last word sent, placing it 8th of 19 timed models; Inworld posted 45 milliseconds and 93.2% reading accuracy on the same test, at the same $15 list price.
How do you reconcile 47 milliseconds from AssemblyAI with 159 milliseconds from Cekura? By noticing what each measures. AssemblyAI’s time-to-first-byte counts the gap until the first chunk of audio arrives under its own test conditions. Cekura’s streaming latency counts from the last word sent through sustained playback under its conditions. Different start lines, different tracks, different results — and both can be true at once. The practical takeaway is not that one side is lying, but that AssemblyAI TTS-1 looks like a top-tier naturalness and entity-accuracy model whose real-world latency advantage over rivals is smaller than its own headline suggests, and that Inworld is the sharper latency competitor on independent measurement. If your use case is a phone agent where a caller reads back a confirmation code, AssemblyAI’s entity accuracy is the number that matters most. If your use case is rapid back-and-forth conversation with constant interruptions, Cekura’s streaming latency deserves more weight — and you should run your own test, on your own traffic, before committing, exactly as AssemblyAI’s own buyer’s guide advises for any voice vendor.
AssemblyAI TTS-1 vs ElevenLabs, Cartesia, and Inworld
The table below summarizes the head-to-head facts, drawn from AssemblyAI’s launch coverage and the vendors’ published pricing. Treat benchmark rows as labeled: AssemblyAI’s own test and Cekura’s independent test measure different things, as explained above.

AssemblyAI TTS-1 vs rivals at a glance:
- AssemblyAI TTS-1: $15 per 1M characters; 6 languages; ~90 ms claimed time to first audio; 47 ms median time to first byte (vendor test); 159 ms median streaming latency (Cekura); 93.7% entity accuracy (vendor test); 4,000+ voices; instant cloning only; unlimited concurrency on pay-as-you-go; hosted US/EU only.
- Inworld TTS-2 Flash: $15 per 1M on demand, $7 on Growth plan; 200+ languages; under 25 ms P99 server-side time to first byte claimed; 45 ms median streaming latency (Cekura); 93.2% reading accuracy (Cekura); instant and professional cloning; 5 to 500 concurrent requests by plan.
- ElevenLabs Flash v2.5: $40 per 1M characters; 32 languages; ~75 ms claimed time to first audio; 119 ms median time to first byte (vendor test); 177 ms median streaming latency (Cekura); 68.0% entity accuracy (vendor test); voice cloning available; concurrency not disclosed for self-serve.
- Cartesia Sonic 3.6: roughly $37 to $50 per 1M characters via plan credits; 44 languages; under 90 ms claimed time to first audio; 143 ms median time to first byte (vendor test); 270 ms median streaming latency (Cekura); 82.5% entity accuracy (vendor test); instant and professional cloning; 2 to 15 concurrent requests by plan; on-premise and private-cloud deployment on enterprise plans.
AssemblyAI TTS-1 vs ElevenLabs: Price and Accuracy Against the Biggest Brand
ElevenLabs is the name most buyers already know, and TTS-1 undercuts its Flash v2.5 model on price by more than 60% while beating it on entity accuracy in AssemblyAI’s test (93.7% against 68.0%) and on naturalness in Cekura’s test (4.50 against 4.33). That is a striking gap, but keep it in proportion: Flash v2.5 is ElevenLabs’ speed tier, not its flagship quality tier — AssemblyAI’s own data ties TTS-1 with ElevenLabs v4 on entity accuracy, and v4 costs substantially more. The fair summary is that TTS-1 gives you near-flagship entity accuracy at a speed-tier price, while ElevenLabs still wins on language breadth (32 languages) and on its mature creator ecosystem, which matters more for content production than for phone agents.
AssemblyAI TTS-1 vs Inworld: The Closest Fight on Price and Speed
Inworld is the rival to watch. Its TTS-2 Flash matches TTS-1 at $15 per 1 million characters, beats it to $7 on a paid plan, wins the independent latency race (45 milliseconds against 159 milliseconds on Cekura’s measure), and covers more than 200 languages. TTS-1 answers with unlimited pay-as-you-go concurrency against Inworld’s 5-request on-demand cap, a larger launch voice library framing, and the one-vendor pairing with AssemblyAI’s transcription. If you run spiky call volumes and hate negotiating tiers, AssemblyAI wins. If you run steady high volume in many languages and can commit to a plan, Inworld’s $7 rate and faster independent latency are hard to ignore.
AssemblyAI TTS-1 vs Cartesia: Simplicity Against Enterprise Control
Cartesia’s Sonic 3.6 is the enterprise-control option: more languages (44), professional cloning, and the only on-premise or private-cloud deployment in this group on enterprise plans. It also costs roughly two to three times the TTS-1 rate and posts the slowest independent streaming latency of the four (270 milliseconds). For a regulated enterprise that must keep audio processing inside its own cloud, Cartesia may be the only viable choice regardless of price. For everyone else, TTS-1 is cheaper, faster on independent measurement, and simpler to buy.
Who Should Use AssemblyAI TTS-1 — and Who Should Not
TTS-1 is a strong fit for teams building live phone agents: appointment scheduling, order-status calls, customer-support triage, delivery updates, and lead qualification, where the agent must read numbers correctly, answer quickly, and survive volume spikes without a concurrency negotiation. It also fits developers already using AssemblyAI transcription who want one vendor, one dashboard, and one bill for the full speech loop. Small and mid-sized businesses benefit most from the pricing shape: no minimum, no annual contract, and no penalty for a quiet month followed by a busy one.
It is a weaker fit in four cases. First, multilingual products: six launch languages cannot serve a global audience, and adding a second vendor later for other languages complicates the stack you were trying to simplify. Second, content production: if you are producing audiobooks, video narration, or marketing voiceovers, you want the speed controls, pronunciation dictionaries, SSML, and emotion direction that TTS-1 does not yet offer — and you should look at the broader voice-generator market, including our guide to the best AI voice generators of 2026. Third, regulated on-premise deployments: hosted-only access rules out teams that must process audio inside their own infrastructure. Fourth, brand-voice perfectionists: instant cloning is fast and good, but a flagship brand voice that must sound identical across millions of calls may still need a professional cloning tier elsewhere.
A useful rule of thumb: if your product is a conversation, TTS-1 deserves a serious trial. If your product is a performance — a narration, a character, an advertisement — wait for the missing creative controls, or choose a tool built for that job.
Getting Started With AssemblyAI TTS-1
Getting a first voice out of TTS-1 is a short job, and the steps below reflect the documented integration path rather than any single tutorial.

- Create an AssemblyAI account and get an API key. The text-to-speech API is billed on the same pay-as-you-go account as AssemblyAI’s transcription products, so existing AssemblyAI customers add voice generation without a new contract.
- Pick a voice from the library — or clone one. Audition voices with your real scripts, especially entity-heavy lines such as order numbers and prices. For cloning, prepare a short, clean recording and the speaker’s permission confirmation; AssemblyAI’s launch example used 24 seconds of audio.
- Choose your interface. Use HTTP for one-shot audio, WebSocket for live conversations with barge-in support, or the OpenAI-compatible route if your code is already built around that shape. Request 24 kHz PCM for app audio, or 8 kHz mu-law for traditional telephony.
- Pair it with transcription. Combine TTS-1 with Universal-3.6 Pro Realtime for the listening half, or drop both into a LiveKit or Pipecat reference agent if you use those frameworks. Keep the language model, transcription, and speech layers separately measurable so you can tell which layer caused a bad call.
- Test on real phone audio before launch. Run your entity-heavy test sentences through an actual phone call, measure your own time to first audio and interruption behavior, and compare against a rival on the same sentences. Vendor benchmarks are a shortlist, not a decision.
- Control verbosity to control cost. Because billing counts every input character including spaces, trim your agent’s prompts and replies. Short, direct answers are cheaper and, on the phone, usually better.
For context on where TTS-1 sits in the wider voice-agent race, it is worth reading our coverage of Mistral Voxtral 3 vs ElevenLabs Realtime, which compares two full voice-agent stacks launched days earlier, and our guide to Nvidia Nemotron Transcribe, the free open model tackling the transcription half of the same problem. The pattern across all three stories is the same: voice AI in late 2026 is splitting into a speed-and-accuracy race for live calls, and a creative-quality race for content — and TTS-1 is firmly a contender in the first race.
AssemblyAI TTS-1 Limitations You Should Know Before You Build
No launch-day product is complete, and TTS-1 has a clear missing list: no SSML support for fine-grained pronunciation markup, no speed control to slow down or speed up delivery, no pronunciation dictionaries for brand names and unusual terms, and no emotion tags to direct tone beyond the voice itself. If your agent must pronounce a pharmaceutical brand name correctly every time, or slow down when reading a legal disclaimer, you will feel these gaps immediately. AssemblyAI has not published a timeline for these features, so treat them as absent, not imminent.
The other limitations are structural rather than missing features. Language coverage is six languages at launch. Deployment is hosted-only, in the US or EU. Cloning is instant-only. And the benchmark story, while strong, mixes vendor and independent measurements that disagree on latency — a reminder to test on your own traffic rather than quote a single number to your stakeholders. None of these limitations makes TTS-1 a bad product; they define the shape of the product it is, and the buyers it serves best, in its first weeks.
There is also a compliance note worth carrying into any deployment. Instant voice cloning lowers the barrier to impersonation as well as to legitimate personalization, and US states are moving quickly on voice likeness and AI disclosure rules. AssemblyAI’s permission-confirmation step is a start, not a compliance program: get written consent from anyone whose voice you clone, tell callers when they are speaking with an AI agent where the law requires it, and keep records. The cheapest voice API can become the most expensive mistake if it is deployed without those basics.
The Verdict on AssemblyAI TTS-1
TTS-1 arrives as a focused, honestly priced answer to a specific question: what does a live phone agent’s voice need to sound like in late 2026? The answer it gives is fast enough, accurate where it counts — names, numbers, and codes — cheap at $15 per 1 million characters, and free of the concurrency negotiations that quietly tax growing contact centers. Its independent naturalness score leads the comparison set, its entity accuracy ties the flagship tier, and its one-vendor pairing with AssemblyAI transcription removes real integration friction.
The case against it is equally clear: six languages, no creative controls, hosted-only deployment, and an independent latency figure that trails its sharpest rival. None of that disqualifies it. It simply means TTS-1 is the right default for US and European phone agents speaking its launch languages, and the wrong default for multilingual products, narration studios, and on-premise enterprises. Run the phone test, count your characters, and let your own callers decide — that is the evaluation this product invites, and it is the one it is most likely to pass.
Frequently Asked Questions About AssemblyAI TTS-1
What is AssemblyAI TTS-1?
TTS-1 is AssemblyAI’s first text-to-speech model, generally available from October 29, 2026. It is a hosted streaming API that turns text into spoken audio for live voice agents, priced at $15 per 1 million input characters.
How much does AssemblyAI TTS-1 cost?
TTS-1 costs $15 per 1 million characters, billed on input characters including spaces and punctuation. All voices, instant cloning, word timestamps, and US or EU processing are included at that rate, with no per-voice fees, concurrency fees, or minimum spend on pay-as-you-go.
How fast is AssemblyAI TTS-1?
AssemblyAI quotes about 90 milliseconds to first audio, and its own test measured a 47-millisecond median time to first byte. Independent benchmarker Cekura measured a 159-millisecond median streaming latency under its own methodology. The two figures measure different parts of the pipeline, so test on your own traffic before deciding.
Which languages does AssemblyAI TTS-1 support?
At launch, TTS-1 supports six languages: English, Spanish, French, German, Italian, and Portuguese. If you need other languages, consider Inworld (200+), Cartesia (44), or ElevenLabs (32) instead.
Does AssemblyAI TTS-1 support voice cloning?
Yes — TTS-1 offers instant voice cloning from a short recording, with confirmation of the speaker’s permission required. AssemblyAI’s launch example used 24 seconds of audio. Professional, studio-grade cloning tiers are not offered at launch.
Can I run AssemblyAI TTS-1 on my own servers?
No. TTS-1 is a hosted API only, running in AssemblyAI’s US (us-east-1) or EU (eu-north-1) cloud. There are no open weights and no on-premise option at launch.
How does AssemblyAI TTS-1 compare with ElevenLabs?
TTS-1 costs $15 per 1 million characters against $40 for ElevenLabs Flash v2.5, beats it on entity accuracy in AssemblyAI’s test (93.7% vs 68.0%), and leads it on naturalness in Cekura’s test (4.50 vs 4.33). ElevenLabs still leads on language coverage and its creator ecosystem.
Is AssemblyAI TTS-1 good for audiobooks or narration?
Not yet. TTS-1 lacks SSML, speed control, pronunciation dictionaries, and emotion tags at launch, so narration and content-production users should choose a tool built for that job, such as those in our best AI voice generator guide.
Sources: AssemblyAI’s text-to-speech product page and voice-agent solutions page (pricing, voices, languages, and interface details); MarkTechPost’s October 29, 2026 launch coverage (vendor benchmark figures, comparison table, and missing-feature list); Cekura’s independent benchmark results as reported in that coverage (reading accuracy, naturalness, and streaming latency). Vendor benchmark figures are AssemblyAI’s own and are labeled as such in this article; independent figures are Cekura’s. Prices were checked on October 29, 2026, and may change — confirm the current rate on AssemblyAI’s pricing page before buying.
