Is Cartesia free?

Start with a short browser speech test

Use the public language demo to audition a voice before opening a longer project. Cartesia's current model documentation names Sonic-3.6 and points to its Playground. The public recordings are demonstrations, not proof that your account can generate every voice or model for free.

1. Pick a language and a recorded voice

   Open one of the language links and listen to a provider recording. Note the voice and locale displayed for that particular recording; choosing another example can replace text you edited.

2. Prepare the exact spoken words

   Use a short original passage in the chosen language, within the demo's documented 500-character limit. Translate and review it beforehand: text to speech reads your script rather than translating it for you.

3. Check the actual model and funding route

   For Playground work, confirm the selected model is Sonic-3.6 and the account is on Free before generating. Keep paid upgrades and overages off. If the current model is unavailable, record the limitation rather than treating an older model as the same test.

4. Review a complete take

   Compare its words and ending with your script. Keep the text, selected voice, locale, settings and any displayed cost together. A successful request is not a pronunciation rating.

Practice: open a miniature night market

This fictional announcement is an original practice script, not a generated Cartesia result. Try a clear narration voice first. Keep creative directions outside the words to be spoken unless you intentionally use a documented speech tag.

• Decide how seven and nine should be spoken before comparing voices. Listen for skipped words in clock tower and blue ribbons, and check that the final sentence finishes.
• Change one setting at a time. Keep the same text for a voice comparison; use a separately reviewed translation when testing another language.
• Keep one distinct voice per language for a small multilingual comparison. Six languages are a useful sample, not coverage of the model's entire language list.

Budget credits without promising a fixed recording length

The current Free plan lists 20,000 model credits per month. Speech synthesis and transcription use that same pool; the separate $1 Managed Agents balance is not another narration allowance. The price table's roughly 27 TTS minutes is an estimate, not a guaranteed finished-audio duration.

Pricing describes one TTS credit per character. As a planning example, a 250-character passage and one equally long retry would budget about 500 credits, before any other usage. Check the actual displayed cost and remaining balance; formatting, other products and repeated requests matter.

Monthly does not mean midnight on the first day. Paid renewal and paid rollover wording do not establish a Free account's exact renewal time or rollover. Keep those Free details unconfirmed until the account or an explicit official rule verifies them.

Match language, accent and delivery separately

A model's supported language list does not mean every voice has every native accent. Inspect the voice's language and supported accents, then choose the recording that matches your listener. A voice can fall back to its original accent when the requested locale is unsupported.

For a multilingual comparison, use one language per request and a different suitable preset voice for each. The six provider listening links on this page cover different languages and voices. They are not a controlled comparison, and their exact recording model, generation cost and Free funding are not supplied here.

Speed and volume are guidance rather than exact duration multipliers. Start with ordinary punctuation and neutral controls. The current emotion guide describes emotion tags as English-only; do not promise the same tagged delivery in every language.

Check downloads, rights and private scripts

The public language pages direct longer scripts and downloads to Playground. API documentation supports WAV and MP3 through its Bytes endpoint; that does not prove a Free browser account's download menu, resolution or format. Inspect an actual saved file before relying on those properties. No Cartesia export was tested for this listing.

Free-generated outputs are restricted to noncommercial use, and pricing adds a commercial-use license on Pro. Do not treat attribution, a downloadable file or the absence of a visible mark as permission to use Free audio in monetized videos, client work or an ad-supported showcase. Our provider listening links leave the recordings on Cartesia's own pages.

The terms describe possible use of inputs, outputs and interactions to improve models and a request-based training opt-out. Use fictional, nonsensitive text for practice. Cloning someone else's voice requires permission; a built-in voice avoids a separate upload-and-consent exercise.

Choose browser practice or a separate API integration

The browser is enough to evaluate a short narration. A streaming API is useful when an application must request speech dynamically, but it adds credentials, response handling and its own playback pipeline. Keep API keys on your server and separate model time-to-first-audio from a complete conversation's response time.

No Cartesia latency benchmark was run here. A provider's advertised model latency is not a measured delay on your phone, network or full voice-agent workflow. Use the same script and playback setup if you later compare services.

When to choose Google AI Studio & Gemini API instead

Try Google AI Studio when you want a browser speech workflow with six original multilingual samples you can play on this site. Their transcripts, selected voices and free funding observations are recorded together.

Google's terms and account/model capacity differ from Cartesia's. Our tested access does not promise unlimited requests, exclusivity or permission to reuse someone else's input. Compare the complete samples yourself instead of treating either guide as a voice-quality ranking.