Models - July 28, 2026

Grok Voice API: Multilingual Voices Guide

Quick answer: xAI released 21 additional flagship Grok voices on July 6, 2026, joining five retrained original voices. xAI says all are multilingual and available through the realtime Voice Agent API, Text to Speech API, and Grok Voice Agent Builder.

The right surface depends on the job: use the Voice Agent API for live conversation, Text to Speech for generated speech, and the Builder for configuring and testing an agent experience.

Three product surfaces

SurfaceUse it forWhat to verify now
Voice Agent APIRealtime conversational agentsCurrent models, transports, events, latency, tools, quotas, and pricing
Text to Speech APISpeech generation from textInput limits, output formats, streaming, voices, and usage terms
Voice Agent BuilderConfigure and test a voice agent in the xAI consoleAccount access, deployment options, saved configuration, and current entitlements

The July announcement establishes availability across all three. Implementation details remain current-documentation checks.

What “multilingual” means in the launch

xAI says each new voice is natively multilingual and supports Grok Voice’s 25-plus languages.

That is a platform capability statement, not proof of equal quality in every language, accent, dialect, domain, or noisy environment. Test the actual speech your users will produce and hear.

Include:

  • Names and entity pronunciation.
  • Numbers, dates, currencies, and addresses.
  • Code-switching between languages.
  • Domain terminology.
  • Background noise and low-quality microphones.
  • Long turns, interruptions, and recovery.

Voice casting is part of the interface

xAI describes voices designed for jobs such as support, characters, commentary, advertising, and education.

Choose by the interaction contract, not novelty:

  1. Match pace and tone to the task.
  2. Confirm the voice remains clear under long or technical responses.
  3. Test whether emotional styling obscures critical information.
  4. Provide a non-voice fallback for essential content.
  5. Review regional and cultural fit with native speakers.

Speech tags

The launch article points to speech tags such as [pause] and <whisper> for delivery control.

Treat tag syntax as current API documentation, not permanent prose. Before production, verify:

  • Supported tags for the selected endpoint and model.
  • Whether unsupported tags are ignored, spoken aloud, or rejected.
  • How tags affect latency and streaming.
  • Whether the intended delivery remains accessible.

The original five also changed

xAI says Ara, Eve, Leo, Rex, and Sal were retrained with an improved recipe for more natural pacing, phrasing, and emphasis.

That means an old evaluation may not represent the current voice even when its name did not change. Pin a dated evaluation sample and rerun it after significant model or voice updates.

Voice cloning requires a separate review

xAI’s launch links a voice-cloning path using about a minute of audio. The announcement does not establish that any supplied recording is authorized or that cloning is permitted for every use.

Before cloning:

  • Obtain informed permission from the speaker.
  • Verify identity, region, account, and product requirements.
  • Review current xAI policy and terms.
  • Define how the clone is disclosed and revoked.
  • Protect source recordings and generated audio.
  • Prohibit impersonation and deceptive use.

For a general model comparison route, see Grok 4. For continuous ChatGPT voice architecture, compare the GPT-Live guide.

A practical voice evaluation

Use the same script across several candidate voices:

  1. A greeting and short interruption.
  2. A domain-specific explanation.
  3. Names, numbers, dates, and addresses.
  4. A language switch.
  5. A speech tag supported by current docs.
  6. A fallback when the user is not understood.

Score intelligibility, pronunciation, timing, interruption behavior, emotional fit, latency, and consistency. Keep the source audio and configuration with the evaluation record.

Frequently asked questions

How many new Grok voices launched?

xAI announced 21 new flagship voices on July 6, 2026, joining five original voices that were also retrained.

Are the new voices multilingual?

xAI says all 21 are natively multilingual and support Grok Voice’s 25-plus languages. Test the exact languages, accents, and terminology required by your application.

Where are the voices available?

xAI says they are available in the realtime Voice Agent API, Text to Speech API, and Grok Voice Agent Builder.

Can speech delivery be controlled?

xAI points to speech tags such as pause and whisper. Verify the current tag syntax and endpoint support in official developer documentation.

Can I clone a voice?

xAI links a cloning workflow, but availability does not establish consent or rights. Review current policy, identity, permission, disclosure, and data-handling requirements before using it.

Are these voices available on Flowith?

This guide covers xAI API and Builder surfaces. It does not claim a Flowith voice or model integration.

Primary source