Cartesia's Sonic-3.5 Beats Google in Blind Human Voice Quality Test
Cartesia's Sonic-3.5 claimed the top spot on the Artificial Analysis Speech Arena leaderboard, beating Google and Inworld on naturalness — at a speed and price tradeoff worth knowing.
Takeaways
- Leaderboard milestone: Artificial Analysis reported Cartesia's Sonic-3.5 reached #1 on the Speech Arena with an Elo of 1,218, edging out Google and Inworld.
- Naturalness wins, speed trades off: Sonic-3.5 hits 105.5 chars/sec — faster than Google (26.3) but slower than Inworld (205).
- Premium pricing: At $39/1M characters, it costs more than twice Google's Gemini 3.1 Flash TTS ($18.3/1M).
- 42 languages, 500+ voices: Includes 9 Indian languages and native alphanumeric handling with no preprocessing.
- SSM architecture: Built on State Space Models (not Transformers), enabling sub-90ms time-to-first-audio latency.
- Live rankings shift: The live leaderboard is dynamic — the top cluster remains tightly contested as votes accumulate.
Artificial Analysis, the independent AI benchmarking outfit behind the Speech Arena leaderboard, announced that Cartesia's Sonic-3.5 had climbed to the #1 position in their blind human evaluation rankings, surpassing both Google's Gemini 3.1 Flash TTS and Inworld's Realtime TTS 1.5 Max. The Speech Arena works like a taste test: real users listen to two unlabeled audio clips generated from the same text and vote for whichever sounds more natural , no brand loyalty, no benchmarks, just ears.
How the Elo math shook out
Elo is the same rating system used in chess , a model gains points by winning head-to-head comparisons and loses points when it doesn't. Sonic-3.5 reached an Elo score of 1,218 based on 1,144 arena appearances, placing it ahead of Inworld Realtime TTS 1.5 Max at 1,194 and Gemini 3.1 Flash TTS at 1,209. The margin over Google was narrow , just 9 Elo points , which matters because these rankings shift as more votes come in. Indeed, the live leaderboard has since continued to move, with the top cluster remaining tightly contested.
What Sonic-3.5 actually is
Sonic-3.5 is Cartesia's fastest, most natural text-to-speech model, with sub-90ms latency and native support for 42 languages. Cartesia is a research company that spun out of the Stanford AI Lab and built its technology on State Space Models (SSMs) , a different architecture from the Transformer models that power most large language models, and one that is significantly more efficient, enabling the sub-100ms latency the company is known for.
SSMs process sequences by maintaining a compact hidden state rather than attending over the full context window, which means they can stream audio tokens with much lower overhead. That architectural choice is what makes real-time voice agents actually feel real-time.
Where it shines
The model follows transcripts faithfully, voices confirmation codes and heteronyms correctly without preprocessing, and supports 42 languages at native quality , including English, Hindi, Spanish, French, German, Japanese, Hebrew, and 35 more. A few standout capabilities:
- Alphanumeric handling: Order IDs, phone numbers, and email addresses are spoken naturally in every supported language with no pre-processing step required.
- Context-aware pronunciation: Heteronyms , words like
read,bass, orbowthat change pronunciation based on context , are resolved correctly from surrounding words. - Indian language depth: The model supports 9 Indian languages, a notable differentiator for voice agent deployments in South Asian markets.
- 500+ voices out of the box, with instant voice cloning from as little as 10 seconds of audio.
The speed-quality-cost triangle
This is where the tradeoffs get real. Sonic-3.5 is priced at $39 per million characters, compared to Gemini 3.1 Flash TTS at $18.3/1M characters and Inworld Realtime TTS 1.5 Max at $35/1M characters. You're paying a meaningful premium over Google for the quality win.
On speed, the picture is more nuanced. Sonic-3.5 processes 105.5 characters per second, compared to 205 characters per second for Inworld Realtime TTS 1.5 Max and 26.3 characters per second for Gemini 3.1 Flash TTS. Inworld is nearly twice as fast at throughput, while Google is four times slower. Sonic-3.5 sits in the middle , but its sub-90ms time-to-first-audio is the metric that matters most for conversational agents, where the user is waiting for the first word, not the last.
| Model | Elo Score | Price ($/1M chars) | Speed (chars/sec) |
|---|---|---|---|
| Sonic-3.5 | 1,218 | $39.00 | 105.5 |
| Gemini 3.1 Flash TTS | 1,209 | $18.30 | 26.3 |
| Inworld Realtime TTS 1.5 Max | 1,194 | $35.00 | 205.0 |
What enterprise users are saying
The reception from production deployments has been notably strong. One enterprise customer reported a 2.9% lift in conversion and a 12.2% increase in customer engagement after switching to Sonic-3.5. Another runs 20 million outbound calls per month on Cartesia, with peak concurrency of 5,000 calls in a single minute and 100ms time-to-first-byte , which they describe as 2x faster than every other voice provider they tested.
Cresta described Sonic-3.5 as delivering refined prosodic rhythm, natural intonation, superior pacing, and a wider emotional range compared to previous TTS models. Sierra noted it had become one of their top-performing models specifically for multilingual coverage, where other models often fall short.
When to use it , and when not to
Sonic-3.5 is a strong fit for:
- Real-time voice agents where sub-100ms first-audio latency is non-negotiable
- Multilingual deployments, especially those covering Indian, Middle Eastern, or Southeast Asian markets
- High-volume outbound calling at enterprise concurrency
- Applications where naturalness directly drives conversion , customer support, sales, recruiting automation
It's less compelling if cost efficiency is the primary constraint. At $39/1M characters, it costs more than twice what Google's Gemini Flash TTS charges for a model that scores only 9 Elo points lower. Sonic-3.5 is also an API that gives you audio , it doesn't come with business logic for knowledge retrieval, helpdesk integration, ticket triage, or workflow automation. You're buying a very good vocal layer, not a full agent stack.
How to start using it
Sonic-3.5 is available now via the Cartesia API. The model ID to use is sonic-3.5, which automatically tracks the latest stable snapshot. For production deployments where reproducibility matters, pin to a dated snapshot like sonic-3.5-2026-05-04. For testing the bleeding edge, sonic-latest gives you access to beta releases with no stability guarantees.
import cartesia
client = cartesia.Cartesia(api_key="your-api-key")
audio = client.tts.bytes(
model_id="sonic-3.5",
transcript="Your table reservation at Corner75 for Saturday at 7:30 p.m. is confirmed.",
voice={"id": "db6b0ed5-d5d3-463d-ae85-518a07d3c2b4"}, # Skylar - en-US
output_format={"container": "mp3", "sample_rate": 44100},
language="en"
)You can also listen to samples directly in the Artificial Analysis Speech Explorer or vote on models yourself in the Speech Arena to form your own opinion before committing to an API key.