Grok's 26 voices across 25+ languages vs how Sume picks a voice

How many voices does Grok have, and does Sume match them?

xAI says 21 new flagship voices joined the original five on July 6, 2026, with the five originals improved, for 26 in total. Sume does not publish a voice-name list in its TTS contract, so you cannot ask for a voice called "Atlas"; you select a voice by avatar or by id.

This matters when you port a script. A request that names a vendor voice will not work on Sume, and Sume returns a clear error before charging anything.

What does the xAI voice release say?

The announcement says each voice was cast for a specific job: support, characters, commentary, advertising and education. The voices are available in the realtime Voice Agent API, the Text to Speech API and the new Grok Voice Agent Builder in the xAI console, and the page lists 25+ languages.

Named voices on the page include Naksh, Atlas, Aurora, Liora, Carina, Zagan, Helix, Orion, Luna and Wellness. The TTS docs add that you can list voices programmatically, and that eve is the default.

How does voice selection work on Sume?

The Sume TTS 1.0 contract offers two routes. The first is an avatar: send avatar_id or avatar_handle, from GET /v1/avatar-1.0/avatars, and Sume resolves that avatar's TTS voice, which works when the avatar's voice status is ready. The second is voice.id, which must be a TTS voice UUID or a Voices library id starting voi_.

If you send both, they must resolve to the same voice or the request fails with 400. Names from other ecosystems are rejected synchronously with invalid_voice_id.

Which job should each voice do?

xAI casts voices by job, which is a sensible way to choose even when the catalog is not portable. Use the same five jobs as a checklist and decide what each needs before you audition anything.

• Support: calm, clear, steady pace; test with numbers, order ids and addresses.
• Characters: strong personality; expect to keep one voice per character for the whole project.
• Commentary: faster, energetic delivery; set speed in generation_config rather than rewriting the script.
• Advertising: short reads with brand names; use pronunciation_dict_id for names that mis-read.
• Education: neutral and warm; check long passages for drift in tone.

What should you do when moving a Grok script to Sume?

Map each Grok voice to a Sume avatar voice by listening, not by name. Generate the same two sentences with each candidate and keep the closest. Record the avatar handle in your project config, so every later render reuses it.

Then set language explicitly. Sume treats a missing language as English at the provider, with a Hangul or kana-only fallback for Korean and Japanese, so a Spanish script without language: "es" will be read as English. The language default post shows the failure.

Finally check pricing before a batch. The page for xAI does not list a price, so compare against the Grok per-million-character quote and use the Sume dry-run cost preview on your own script length.