HeyGen, known for AI avatars and video, has entered the text-to-speech race with HeyGen Voice, its first in-house voice model. It debuted on October 9, 2026 at number one on Artificial Analysis' Controlled Voice Arena, and this week it arrived on OpenRouter and Design Arena, with a lower-latency Turbo tier and half-price launch pricing through October 31.
The headline is easy to repeat and easy to over-read. "Ranked first" is true for one specific arena. Below we lay out what each source says, what the numbers cover, and how to decide whether to try it. Most figures come from OpenRouter's and Artificial Analysis's posts as relayed by secondary coverage; we could not open the original announcements, so check the leaderboard directly before relying on the exact Elo.
Claimed vs verified
| Claim | Source | Status |
|---|---|---|
| First in-house native TTS model | HeyGen executive on X, secondary coverage | Stated by the company |
| #1 on Controlled Voice Arena, Elo about 1,201 over 1,468 appearances | Artificial Analysis, relayed by secondary outlets | Leaderboard result; figure differs slightly across reports (1,201 vs 1,202) |
| Beats Qwen-Audio-3.1-TTS-Plus and Eleven v4 Turbo | Same | Blind-test result on that arena only |
| $15 per million characters, Turbo $7.50, both 50% off through Oct 31 | OpenRouter post | Vendor-platform pricing, promo |
| 40 characters per second | Reported with the arena data | Single reported figure, method unclear |
| 10th for pronunciation robustness, 83.1% | Artificial Analysis, relayed | Weaker spot |
| #1 overall for voice quality | Not supported | Provider Voices board is led by Eleven v4 Turbo in reports |
What the arena actually measures
The Controlled Voice Arena is built around eight cloned reference voices: two male and two female in US English, plus two male and two female in UK English. Listeners compare outputs in blind tests, and models earn an Elo score. That setup isolates the model from each vendor's own voice library, which makes it a fairer test of the engine, but it also narrows what the score means.
Three caveats follow:
- It is a voice-cloning setting. Results with a vendor's native voices are measured on a different board, and reporting says HeyGen Voice is not yet listed there because it has too few votes.
- Languages are limited to English accents. Strength in UK English and assistant-style roles does not tell you about other languages.
- Pronunciation is a separate axis. HeyGen Voice placed 10th of 29 on pronunciation robustness at 83.1%, though second for preserving the exact text sequence. One analysis notes it is weaker at expanding shorthand such as abbreviations.
For earlier voice-model comparisons on our site, see the ElevenLabs Eleven v4 and v4 Turbo coverage, Cartesia Sonic 3.6 on Artificial Analysis and Nari Labs' Qwen3-TTS speed and cost study.
A microphone sending a green sound wave toward a speaker, standing for a text-to-speech model producing voice output
Pricing: the half-off promo is not the price
OpenRouter lists HeyGen Voice at $15 per million characters against a $30 list price, and HeyGen Voice Turbo at $7.50 per million against $15, both at 50% off through October 31. Plan around the list prices, and remember promos like this are common. We made the same point about OpenRouter's GPT-5.6 Sol 50% promo, which was not a price cut.
One more wrinkle: a figure of $30 per million characters appears in one report as Artificial Analysis's conversion of HeyGen's credit pricing, not a quote from HeyGen. Another review says the base model is free within HeyGen's own platform and API, with a separate $99 per month add-on for professional voice cloning. These are secondary reports and may be out of date. Confirm on HeyGen's pricing page.
Rough cost math
At the OpenRouter promo, a 1,000-word script of about 6,000 characters costs roughly nine cents on the standard model ($15 per million) and four and a half cents on Turbo. At list price those double. A voice agent speaking 10 million characters a month would pay about $150 promo or $300 list on the standard model. These are our calculations from the stated rates, not quoted prices.
Where it fits in the HeyGen stack
HeyGen has been building out beyond avatars: our earlier posts cover HyperFrames, its HTML-to-video framework for agents and its open-sourced real-time avatar stack on GPT-Live-1. Owning the voice model matters for that stack: avatar lip-sync, dubbing and conversational agents all depend on speech quality and latency, and a first-party model reduces dependence on third-party TTS vendors. The OpenRouter and Design Arena listings are the company's push to reach developers who do not use HeyGen's apps.
How to test it fairly
- Use your real scripts. Include numbers, dates, abbreviations, product names and code-like strings, since that is where the 83.1% pronunciation score suggests risk.
- Run blind A/B tests against your current vendor with at least a dozen listeners and several voices.
- Measure time to first audio, not just throughput. The 40 characters per second figure is a speed, not a latency.
- Compare standard and Turbo on the same script to see what quality you give up for speed.
- Check licensing for voice cloning. Professional cloning needs consent and rights for the voice; read HeyGen's terms.
- Price at list, then treat the promo as a bonus. For routing and gateway context see our OpenRouter enterprise guide.
Reading the Elo gap
Elo in an arena is relative. The reports give HeyGen Voice about 1,201 against roughly 1,182 to 1,186 for Qwen-Audio-3.1-TTS-Plus and 1,166 to 1,167 for Eleven v4 Turbo, with the exact numbers differing by a point or two between outlets. A gap of fifteen to thirty-five Elo points means listeners preferred HeyGen's clip in a modest majority of head-to-head comparisons, not that it is dramatically better. With 1,468 appearances the confidence interval is also not tiny, so the top spots could swap as votes accumulate. New models often debut high on novelty and settle, which is another reason to rerun your own test in a few weeks.
The listing on the OpenRouter platform (a Turbo page also exists) matters for a different reason: it gives developers a single API key and billing relationship across providers, so trying a second voice vendor costs minutes. The Design Arena listing is aimed at conversational assistants and customer-service use. For ranking context, the Artificial Analysis site hosts the arenas and lets you listen to samples yourself, which is the quickest way to judge whether the preference gap is audible to you.
Which job, which voice
| Job | What the evidence suggests | What to verify |
|---|---|---|
| English assistant or support bot | Strong blind-test showing, especially UK English and assistant roles | Latency to first audio, number and abbreviation handling |
| High-volume narration on a budget | Turbo at $7.50 per million characters promo, $15 at list | Quality loss versus standard; promo ends Oct 31 |
| Multilingual products | No multilingual evidence in the sources reviewed | Test every language you ship |
| Voice cloning for brand voices | Controlled arena uses cloned reference voices; professional cloning is a paid add-on in one report | Consent, rights, cost of the add-on |
| Maximum native-voice quality | Provider Voices board reportedly led by Eleven v4 Turbo | HeyGen's standing once it has enough votes |
What this means for builders
If you build voice agents, audiobooks or video narration, a new strong contender is useful leverage in price negotiations and as a fallback provider. If you need many languages, reliable abbreviation handling or the best native voices today, the evidence so far does not show HeyGen Voice ahead. Treat the arena result as a reason to run your own test, not as a ranking of record.
Questions people are asking
Is the model open source? Nothing in the sources we reviewed says so. It is offered through HeyGen's v3 API and through OpenRouter, which indicates a hosted, proprietary model. If you need local or on-premise speech, look at open options such as the ones in our VoiceStudio local ElevenLabs alternative guide.
Why does a video company build a voice model? Avatars, dubbing and agents all need speech. Controlling the engine lets HeyGen tune lip-sync, latency and cost together, and sell the voice separately to developers who never touch its video tools.
Why do reports disagree on the Elo and the price? Leaderboards update as votes arrive, so an outlet that copied the figure an hour later may show a different number. Price quotes differ because platform pricing, credit conversions and promotions are not the same thing. When two sources disagree by a point or two, trust the live leaderboard; when they disagree on dollars, trust the seller's own page.
What happens on November 1? The OpenRouter promo ends and the listed rates double to $30 and $15 per million characters, per OpenRouter's own post. If you are prototyping now, record your usage so you can model the real bill.
Does a TTS arena win translate to agent quality? Only partly. A phone agent also depends on speech recognition, interruption handling and turn-taking, covered in our write-ups of OpenAI's GPT-Live-1 voice agent API. A pleasant voice with a slow first response still feels broken on a call.
Bottom line
HeyGen Voice is a credible first-party TTS model with a strong showing in blind tests on one controlled arena, a Turbo tier, and a cheap promo window. It is not yet shown to be the best voice model overall. Check the Artificial Analysis boards yourself, confirm pricing with HeyGen, and run your own listening test before switching production traffic.
Figures reflect secondary coverage of Artificial Analysis and OpenRouter posts as of October 14, 2026 and may change.