Explore Cartesia Sonic 3.6, including instant and professional voice cloning, low-latency speech, multilingual voices, localization, streaming and API features.
Most voice AI tools are built for creators who generate audio and download it. Cartesia is built for a different job: speech that streams in real time, fast enough for live phone agents and interactive apps. Its Sonic model family adds voice cloning and a voice changer on top of very low latency.
This guide covers what Cartesia Sonic is, how its voice cloning and voice conversion work, how to get started, and what to check on consent and pricing. To compare it with other tools, browse our Voice Cloning & Voice Conversion category.
Sonic is Cartesia's flagship text-to-speech model. Unlike most TTS systems, which are transformer-based, Sonic is built on state-space-model (SSM) architectures, which is how it reaches roughly 90 ms time to first audio. The company was founded by researchers behind the Mamba and S4 state-space papers.
The Sonic line has evolved quickly. Sonic 3 added emotion and laughter and expanded coverage to 42 languages, and third-party sources describe Sonic 3.5 as generally available since May 2026. Cartesia may have shipped newer versions since, so check the official site for the current default model.
Cartesia is an API-first platform. Sonic is not open source and runs through the API and SDKs, though the Enterprise tier is reported to offer on-premise or VPC deployment.
Voice cloning creates a reusable voice from a sample of someone speaking. You then send text and Sonic speaks it in that voice. Cartesia offers two levels: Instant cloning, and Professional cloning, which trains a speaker model from submitted audio.
Voice conversion takes existing recorded speech and re-speaks it in a different voice. On Cartesia this is the Voice Changer endpoint, which transforms audio instead of generating it from text.
A common workflow is to clone a voice once, then use it for both text-driven narration and live or recorded voice transformation.
| Capability | What it does |
|---|---|
| Instant Voice Cloning | Clones a voice from as little as 3 seconds of audio |
| Professional Voice Cloning | Trains a dedicated speaker model for higher fidelity |
| Voice Changer | Converts recorded speech into a target voice |
| Real-time streaming TTS | Starts streaming audio in well under a second, with sub-90 ms model latency |
| 40+ languages | Native-accent voices across many languages |
| Emotion and delivery controls | Adjust pace and emotion, including laughter and emotional cues in Sonic 3 |
| Infill | Edit or fill gaps in existing audio |
| Speech-to-text and voice agents | Ink for transcription, Line for managed voice agents |
| Playground and API | Test in the browser, then call the same model through the API and SDKs |
You can try the model without signing up in Cartesia's voice generator, and see the cloning feature on the official voice cloning page.
| Instant Voice Cloning | Professional Voice Cloning | |
|---|---|---|
| Audio needed | A few seconds (Cartesia cites 3 seconds) | More audio, submitted for training |
| Speed | Available immediately | Requires a training step |
| Fidelity | Good, preserves accent, style, and emotion | Higher, built for long-term brand voices |
| Plan access | Paid plans | Higher tiers (reported on Startup and above) |
| Rights requirement | Consent required | You must own or have rights to the audio |
Plan availability changes, so verify it on the official pricing page.
Sonic suits developers and product teams building real-time, speech-first experiences. If your priority is latency and API reliability, it is a strong fit. It is less ideal if you want an open, self-hosted model or a heavily visual studio for casual editing.
| Need | Better fit |
|---|---|
| Ultra-low-latency voice for live agents | Cartesia Sonic |
| Large creator ecosystem and studio tools | ElevenLabs |
| Fast cloning with open models available | Fish Audio |
| Fully open-source, self-hosted cloning | Chatterbox |
Cartesia's own comparison pages claim advantages over ElevenLabs on latency and the amount of audio needed for cloning. Treat vendor comparisons as claims and test with your own scripts.
1. Confirm you have the rights.
Clone only your own voice or one you have explicit permission to use. Professional cloning requires that you own or have rights to the audio.
2. Create an account.
Sign up on the Cartesia platform and open the playground.
3. Choose a model.
Select the current Sonic model. Check the model list for the latest default.
4. Record or upload a clean sample.
Use a quiet room and one speaker with natural delivery. In the playground you can upload a longer recording and select a short segment of roughly 10 seconds.
5. Create the clone.
Instant clones are created right away. Test it by typing a short script and playing it back.
6. Tune delivery.
Adjust speed, volume, and emotion. Add emotional cues in your text where the model supports them.
7. Use voice conversion if needed.
Send recorded audio to the Voice Changer endpoint and choose your cloned voice as the target.
8. Move to the API.
Use the official SDKs, such as the Python library, to stream speech in your app. Test latency in your real deployment region.
Voice cloning can be abused for impersonation and fraud. Third-party reviews note that Cartesia applies consent checks to cloning, which may limit some use cases. Whatever the platform checks, you remain responsible for how you use a voice:
This guide is not legal advice. Consult a qualified professional for commercial or regulated deployments.
Cartesia uses a credit-based model, where Sonic TTS uses roughly one credit per character. Voice agent usage is billed separately. Third-party sources report these tiers:
The lower figure in each range appears to be annual billing and the higher one monthly billing, but confirm this on the official pricing page. The Voice Changer is reported to draw from your credit allowance by the second of audio, so budget for it separately from text generation.
Cartesia Sonic is one of the strongest choices when speed matters most. Its streaming latency, 40+ languages, instant cloning from a few seconds of audio, and Voice Changer make it a good foundation for live voice agents and interactive products. The platform also covers transcription and managed agents, so teams can build a full voice stack in one place.
The trade-offs are clear. Sonic is closed-weight and API-only outside Enterprise, credit-based costs need modeling, the free tier is not for commercial use, and consent and disclosure remain your responsibility. Start with a short, clean sample of your own voice, test the real-time performance in your own setup, and scale once the results meet your requirements. Ready to try it? Explore Cartesia.
Sonic is Cartesia's real-time text-to-speech model family. It focuses on very low latency, many languages, and voice cloning through an API. Learn more at cartesia.ai.
Cartesia says Instant Voice Cloning works from as little as 3 seconds of audio. Professional cloning uses more audio to train a dedicated speaker model. See the official voice cloning page.
Yes. The Voice Changer endpoint converts recorded speech into a target voice.
Sonic 3 supports 42 languages according to third-party sources. Check the official site for the current list.
Cartesia reports about 90 ms model latency for Sonic and lower for its Turbo variants. Real-world latency depends on your network and region.
No. Sonic is closed-weight and available through Cartesia's API and SDKs. On-premise deployment is reported as an Enterprise option.
Cartesia has a free tier with a limited monthly credit allowance. It is intended for testing, and commercial use requires a paid plan. See the official pricing page for current details.
Only with their explicit permission. Using a cloned voice to impersonate or mislead people is unethical and may be illegal.
