Introduction

On September 23, 2026, Cartesia shipped Multilingual Voices—a voices-and-locale product drop, not a new Sonic base model. After Sonic-3.6 landed on August 27 as the speech engine refresh, this release is about keeping one voice identity intact while it speaks up to 25 languages natively.

If you have ever localized a brand narrator, you know the old trap: five markets meant five actors, and your “house voice” turned into a stranger the moment English ended. Cartesia’s pitch is simpler: pick one voice ID, keep its pitch, warmth, texture, and pace, and let a locale carry that same person into Japanese, Spanish, Hindi, and more.

Primary source: Cartesia — Introducing Multilingual Voices by Akki Shafiqullah and Aiqi Liu.

Pro Tip

Treat Multilingual Voices as a localization layer on top of Sonic, not a replacement model. Keep your existing voice_id. For library voices, pass a supported locale on POST /tts/bytes. For Instant Voice Clones, clone once (~10 seconds of audio), then add up to nine accents/languages via the dashboard or PATCH /voices/{voice_id}/accents before you scale traffic.

What Shipped on Sep 23

Cartesia’s official blog frames the drop around identity continuity:

  • 50+ library voices now speak up to 25 languages natively, with the same voice identity across languages
  • Instant Voice Clones can be made multilingual from about 10 seconds of reference audio
  • Add up to 9 new accents/languages in the dashboard, or via PATCH /voices/{voice_id}/accents
  • Generate with the same voice ID + locale through POST /tts/bytes
  • Every language/accent is scored by native speakers for accent hold and local reading of dates, numbers, and currency

Use cases called out on the blog: customer support, healthcare, and brand localization—anywhere trust collapses if the voice sounds “translated” instead of local.

Same Voice, Many Markets

Cartesia defines a voice as a bundle—pitch range, warmth, texture, pace and emphasis—and argues that identity and recording language used to be tangled. Multilingual Voices is meant to untangle them so a support agent, a housing eligibility bot, or a memorial product can stay one person across the languages your users actually speak.

Customer quotes on the launch post underline the product surface rather than a new codec race:

  • Tony Stoyanov (CTO) on multi-language housing eligibility and recertification at massive call volume
  • Hasan Zalkout (Forevermore CTO) on families typing in their own language while hearing the same familiar cloned voice
  • Dovydas Remeika (co-founder) on generating a guide once, then opening every other language with the same narrator

That is the audio-specialist takeaway: this drop is about locale-aware identity, not another Sonic version bump.

How Builders Call It

Library path

  1. Open the Voice Library, pick a multilingual voice, then More actions → Try in playground
  2. Choose a language from the language menu
  3. Type text and play—same voice, new language

Via API, send text with a locale and the same voice_id on POST /tts/bytes.

Clone path

POST /voices/clone# Instant Voice Clone from ~10 seconds of audio

Then add accents/languages (up to nine) in **Voice Library → More actions → Manage voice → Supported languages & accents → Add**, or:
PATCH /voices/{voice_id}/accents# keep the same voice_id; attach new locales

Generate afterward with the unchanged voice ID plus the target locale. Native-speaker scoring is Cartesia’s quality bar for whether dates, currency, and numbers land the local way—not just whether the phonemes are “close enough.”

How This Fits Cartesia’s Stack

Keep the lanes clear for builders who already shipped on Sonic:

PieceRole

Sonic-3.6 (Aug 27)

Base speech model refresh

Multilingual Voices (Sep 23)

Library + clone identity across up to 25 languages via locale / accents API

ArtRealmAI covers Sonic model bumps when they land; this article is the voices/locale product day.

Conclusion

Multilingual Voices is the localization unlock Cartesia’s library and Instant Voice Clones were missing: one identity, many markets, native-speaker scored accents, and a clean locale path on the same voice ID. If you are shipping support, healthcare, or brand audio across languages, start from the official Multilingual Voices post, clone or pick a library voice once, then expand accents before you rewrite your pipeline around a new model name.

—Anabel ♡