"It's the first model that doesn't sound like a model."
— Iris Tanaka, Lead Researcher · Halcyon Labs
svara-global-v1 is a 780-million-parameter voice model built to feel like a person, not a stat sheet. Same voice across English, Hindi, Japanese, and 47 more. Inline emotion. Real-time. Open weights.
First-byte under 200ms from any of nine regions. Streaming TTS that interrupts cleanly when the user does — the way humans actually talk.
The same identity carries from English to Spanish to Japanese to Hindi inside one sentence. No re-prompting, no different voice per locale.
Inline [warm] · [whisper] · [laugh] — 34 emotion, prosody, and nonverbal tags as plain in-text DSL.
Same model. Same API. Different voice, different tone, different language — chosen by the script, not by your wallet. Tap one to hear it.
Always-on voice that listens, remembers, and answers warmly. Replaces stilted assistants.
Front-line voice for support, IVR, and in-app help — confident, never robotic, in your customer's language.
Game NPCs and interactive fiction. Whisper, shout, laugh — staged in plain text.
Audiobooks and long-form content. The same voice carries from chapter one to chapter twenty-six, in any language.
Pick a mood — the same voice changes how it speaks, not who it is. The transcript stays the same; only the feeling shifts.
"It's the first model that doesn't sound like a model."
Open weights, Apache 2.0. 5,000 free characters per month, no card. Self-host on a single A10 — or call our API in fewer lines than it took to read this sentence.
Start a conversation →