AI voice models: xAI Grok Voice

AI voice models, episode 3. Previous: OpenAI vs Gemini TTS. Next: Cartesia vs Hume.

Grok Voice is SpaceXSI’s (formerly xAI) family of voice APIs: a realtime speech-to-speech model for voice agents, a text-to-speech endpoint, speech-to-text, and voice cloning that works across them. The flagship is Grok Voice 3.0, launched on 14 October and now behind the grok-voice-latest alias, so anyone on the alias was moved over automatically. It also powers Grok Bot’s voice mode. Grok Bot helps produce this site and SpaceXSI makes both, so read this with that in mind. We ran it through the same 12-point checklist as ElevenLabs and OpenAI and Gemini. Every figure below is from SpaceXSI’s pages or Artificial Analysis, read on 30 October 2026.

Grok Voice against the checklist

1. Voice cloning. Custom Voices clones a voice from a reference clip of up to 120 seconds; SpaceXSI recommends 90 seconds or more of clean, single-speaker audio recorded in the style you want back, because the model copies delivery as well as timbre. A team can create up to 30 custom voices free, and using them in TTS or speech-to-speech costs nothing extra. The launch post describes a two-stage check in the console: the speaker reads a passphrase that is transcribed and matched live, then speaker embeddings from that clip and the full recording must match. SpaceXSI says you “can’t clone a voice from a pre-existing recording” or someone else’s voice. The API reference also documents a file-upload endpoint, and we could not confirm from the docs how the same check applies there.

2. Languages. The TTS docs list 20 languages plus automatic detection, say more work with varying accuracy, and state that every voice speaks every language. SpaceXSI describes speech-to-speech as 25+ languages, and its newest transcription model, Grok Voice Transcribe 2.0, at 38+ for speech-to-text. There are 26 built-in voices after the July batch of 21, with eve the TTS default.

3. Latency and streaming. SpaceXSI claims “class-leading latency at under one second” for Voice 3.0. Artificial Analysis measures its time to first audio at 0.81 seconds, quick but not the fastest: several models on its board start speaking sooner. Speech-to-speech runs over a WebSocket at /v1/realtime with server-side voice activity detection, ephemeral tokens for browser clients, and LiveKit and Pipecat integrations. TTS can stream over HTTP or a bidirectional WebSocket, with an optimize_streaming_latency setting and cancel-on-interrupt.

4. Emotion and style. TTS takes inline tags such as [pause] and [laugh], and wrapping tags such as <whisper>, <emphasis> and <soft>. Speech-to-speech has an output speed control from 0.7 to 1.5. There is no separate emotion slider: the reference clip sets a custom voice’s expressiveness.

5. Formats and sample rates. TTS outputs MP3 (default 24 kHz at 128 kbps), WAV, raw PCM, mu-law and A-law, at 8 to 48 kHz, with optional character-level timestamps and a replace map that accepts IPA for pronunciation. Speech-to-speech handles PCM at 8 to 48 kHz, mu-law, A-law and Opus, which covers telephony without transcoding.

6. Price (30 October 2026). From the pricing page:

• Speech-to-speech (grok-voice-3.0): $0.08 a minute ($4.80 an hour), plus $0.004 per text input. Same price as Voice 2.0.
• Text-to-speech: $15.00 per million characters.
• Speech-to-text: $0.10 an hour for batch, $0.20 an hour streaming.

On Artificial Analysis’ TTS board, the voice ranked just above SpaceXSI’s lists at $100 per million characters, so this is one of the cheaper commercial options.

7. Speech-to-speech and realtime. This is the product SpaceXSI leads with. Voice 3.0 reasons while it speaks (reasoning effort high by default, or none), calls tools mid-conversation (function calls, MCP servers, web, X and file search), supports session resumption and an idle timeout, and takes a language hint and up to 100 key terms to bias recognition. SpaceXSI says Starlink call centres use it on 20,000 calls a day, with about 70% resolved without a human; those are its own numbers. There is also a no-code Voice Agent Builder, in beta since July, with SIP telephony.

8. Long-form. A single TTS request takes up to 60,000 characters with a 15-minute timeout. For longer scripts, the bidirectional WebSocket has no total text limit (60,000 characters per message) and allows 50 concurrent sessions per team.

9. Against the previous version. Voice 3.0 scores 86.9 on the Artificial Analysis Speech to Speech Index against 81.3 for Voice 2.0, at the same price and with the same API, and 2.0 is now marked deprecated. In SpaceXSI’s launch table, the gains are biggest on human preference (Arena Elo 1070 against 1011) and conversational dynamics (Full Duplex Bench 97.0% against 95.1%). The original five TTS voices were retrained in July for more natural pacing.

10. What users think. Hands-on reports are thin so far. Most posts we found repeat the launch numbers. One builder described building a phone receptionist on Voice 3.0 in a day and called the $0.08-a-minute running cost the selling point. We found no independent long-run reviews of the TTS voices yet.

11. Licence and commercial use. Grok Voice is closed and API-only, with no open weights. Under SpaceXSI’s enterprise terms, updated 27 October, the customer owns the output and SpaceXSI says it will not train its models on customer content. The Acceptable Use Policy bans deceptively impersonating a real person. If you are not on an enterprise agreement, check which terms cover your account.

12. Verdict. For voice agents, Grok Voice 3.0 is one of the strongest options on the independent board at a simple flat price, with telephony codecs and tool calling built in. For narration, the TTS is cheap and flexible but sits well down the listener-preference board.

Where it ranks

On Artificial Analysis’ speech-to-speech page, read on 30 October, Grok Voice 3.0 scores 86.9 on the index, just behind Phonic 1 (High) at 87.1 and ahead of GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live with extended thinking at 82.6. It launched at number one on 13 October data, so the top spot is already contested. Its component scores are 99% on Big Bench Audio, 71.6% on tau-Voice, 97.0% on Full Duplex Bench and a 94.8% task success rate.

On the text-to-speech leaderboard, SpaceXSI TTS has an Elo of 1157 and sits around 28th. That is a long way below the leaders, which is the trade-off for its price.

Summary

	Grok Voice (SpaceXSI)
Best for	Realtime voice agents, phone support
Flagship	grok-voice-3.0 (alias grok-voice-latest)
Speech-to-speech price	$0.08 per minute
TTS price	$15 per 1M characters
Cloning	Up to 120 s clip, 30 free voices, speaker verification
Languages	20 listed for TTS, 25+ for speech-to-speech
Speech-to-speech index (AA)	86.9, second on 30 Oct
TTS Elo (AA)	1157, around 28th
Weights	Closed, API only

We did not generate our own samples for this review. SpaceXSI’s launch posts and console playground have demos. For voices you can run yourself, see our open speech models guide.