이미지: METAL LAB 생성
Summary
- Cartesia's streaming TTS model Sonic-3.6 simultaneously topped both of Artificial Analysis's speech leaderboards
- On the Controlled Voice board, which compares only the engines using 8 identical reference voices, it beat Sonic-3.5 and ElevenLabs' Eleven v3
- It uses a state space model architecture instead of a transformer, and is available only in beta through Cartesia's API
- 모델
- Cartesia Sonic-3.6 (스트리밍 TTS, Sonic-3.5 이후 약 3개월 만)
- 프로바이더 보이스 Elo
- 1,283점, 아티피셜 애널리시스 리더보드 1위
- 컨트롤드 보이스 Elo
- 1,123점, 1위(2위 Sonic-3.5, 3위 ElevenLabs Eleven v3)
- 아키텍처
- 트랜스포머 대신 스테이트 스페이스 모델(SSM) 기반
- 지연시간
- 첫 오디오 생성까지 90ms 미만(제조사 발표치)
- 공개 형태
- 베타, 카르테시아 API로만 제공, 오픈 웨이트·허깅페이스 저장소 없음
- 가격
- 100만 자당 49달러, ElevenLabs Eleven v3(100달러)의 절반, Speechify Simba 3.2(10달러)보다 고가
- 요금제
- Free/Pro 5달러, Startup 49달러/월, Scale 299달러/월(약 10,667분, 동시요청 15건)
Why it won even with voices standardized
Artificial Analysis, which ranks real-time speech synthesis models, runs two leaderboards. One is the Provider Voice board, where each company competes using the voices it actually sells. The other is the Controlled Voice board, where every model is made to clone the same 8 reference voices, stripping away the appeal of a voice catalog to leave only the raw capability of the synthesis engine itself. Cartesia's newly released Sonic-3.6 topped both boards simultaneously — scoring 1,283 on Provider Voice and 1,123 on Controlled Voice. The Controlled Voice result carries more weight because it means the improvement came from the engine itself, not from the voices. On that board, Sonic-3.5 placed second and ElevenLabs' Eleven v3 placed third.
What Sonic-3.6 changed
Sonic-3.6 launched about three months after its predecessor, Sonic-3.5. The biggest change is voice naturalness, a claim that can be independently verified from the outside. MarkTechPost assessed that "the change here is naturalness, and it's independently verifiable." The model runs on a state space model (SSM) architecture rather than a transformer. While transformers require more computation as sentences get longer, SSMs update state sequentially, making them better suited to real-time streaming. Cartesia says this architecture enables a time-to-first-audio of under 90ms. That figure, however, is the company's own measurement of model latency and may differ from the round-trip latency users experience in an actual live service.
Beta status and access conditions
Sonic-3.6 is still in beta and is only available through Cartesia's own API. It is not an open-weight model that can be downloaded and hosted on your own server, nor is it listed on public repositories like Hugging Face — it's a rental-style commercial model. Cartesia's official documentation still lists Sonic-3.5 as the stable version, and most partner companies are still using 3.5.
| Plan | Target | Cost |
|---|---|---|
| Free / Pro | Solo developers, early-stage startups | $0–$5 |
| Startup | Growing companies running call centers | $49/month |
| Scale | Large enterprises needing regulatory compliance | $299/month (approx. 10,667 minutes of TTS, 15 concurrent requests) |
On pricing, Artificial Analysis calculated Sonic-3.6 at $49 per million characters. That's about half the price of ElevenLabs' Eleven v3 ($100), but more expensive than Speechify Simba 3.2 ($10), which scored 1,240.
| Model | Price per million characters | Benchmark Elo |
|---|---|---|
| Speechify Simba 3.2 | $10 | 1,240 |
| Cartesia Sonic-3.6 | $49 | 1,283 (Provider Voice) |
| ElevenLabs Eleven v3 | $100 | 3rd on Controlled Voice |
Features built for practical use
Sonic-3.6 adds several control features designed more for agent conversation scripts than for narration. Writing tags like "[laughter]" directly into the script triggers non-verbal expressions such as laughter during playback. Voice cloning can be done on the fly with as little as about 10 seconds of audio, and the model also supports a pronunciation dictionary that lets users specify the exact pronunciation of words like "subpoena" using IPA notation, such as "<<s|ə|ˈ|p|i|n|ə>>." Speed, volume, and emotion values can be adjusted via the API and the LiveKit Agents plugin, and alphanumeric strings like order numbers or phone numbers are read out accurately without extra preprocessing. Published demo videos show natural pauses and filler-word-laced English, as well as code-switching between Hindi and English.
How to try it
① Start by choosing the Free or Pro tier on Cartesia's website pricing page. ② Creating an account issues an API key, which you connect to your own app or an integration tool like LiveKit Agents. ③ Send script text to the API along with expression tags, pronunciation dictionary entries, and speed/emotion parameters, and the audio streams back in the desired tone. The company lists target industries including finance, healthcare, retail and e-commerce, logistics, recruiting, SaaS customer support, consumer companion apps, and media localization. A call center, for example, could apply it to inbound support automation or outbound confirmation calls, while a media company could use it for dubbing localization. To use it commercially, however, you need at least the Pro tier ($5), and large-scale operation adds separate per-minute call charges ($0.06/minute).
Editor's take
What stands out in this announcement isn't the ranking itself but which board it won on. Winning because your voice catalog is good and winning because your engine is good are two different things, and the Controlled Voice board strips away that distinction to measure raw synthesis quality alone. Cartesia topping that board too suggests the improvement in naturalness comes from the model itself, not from voice marketing.
Anyone who's watched this space for a few years will recognize a familiar pattern repeating. As transformer-based TTS pushed quality higher, latency was always the trade-off holding it back — and companies building on state space models have recently begun breaking that trade-off one piece at a time. Sonic-3.6 topping the leaderboard while touting sub-90ms time-to-first-audio is a case in point: it shows you no longer have to sacrifice either speed or quality.
Domestic call centers and customer support teams should look at the pricing first. At $49 per million characters, it sits between ElevenLabs (more expensive) and Speechify (cheaper), and it's easy to be surprised by the monthly bill if you overlook the additional $0.06-per-minute call charge. The safer approach is to first test the pronunciation dictionary and expression tags on a small scale with the Pro tier before moving up to Startup or higher. The fact that it's a rental API rather than open weights is also a constraint for teams looking to run the speech engine on their own infrastructure.
In the coming weeks, the things to watch are whether Cartesia's documentation switches the stable version from Sonic-3.5 to 3.6, and how ElevenLabs responds to this ranking. It's fair to say the next round in the fight to reclaim the top spot on both leaderboards has already begun.
