Grok Speech (STT + TTS APIs)
A Tier · 8.1/10
xAI's standalone voice APIs -- launched 2026-04-17. Built on the stack that powers Grok Voice, Tesla vehicles, and Starlink customer support. $0.10/hr STT batch, $4.20 per 1M characters TTS, 25+ languages. Now 26 flagship TTS voices (21 new added 2026-07-06), custom voice cloning from ~1 min of audio, and a no-code Grok Voice Agent Builder
Score Breakdown
The Good and the Bad
What we like
- +Published word-error-rate benchmark puts Grok STT ahead of ElevenLabs, Deepgram, and AssemblyAI across phone calls, meetings, podcasts, and telephony (6.9% overall WER vs. ElevenLabs 9.0%, Deepgram 11.0%, AssemblyAI 12.9%)
- +Pricing is aggressive -- $0.10/hr batch STT undercuts Deepgram's $0.0043/min ($0.258/hr) and ElevenLabs Scribe v2 on most real-world workloads, and $4.20 per 1M char TTS is roughly half ElevenLabs' Creator-tier effective rate
- +Speech tags ([laugh], [sigh], [whisper], <emphasis>, <slow>, <pause>) give expressive TTS control without SSML -- genuinely closer to ElevenLabs v3's emotional range than anything else at this price point
- +Same stack that powers Grok Voice, Tesla in-car voice, and Starlink customer support -- this is not a research preview, it's a production-scale system exposed as an API
What could be better
- −Brand-new (launched 2026-04-17) -- expect early rough edges in non-English accents and in the streaming WebSocket API as the infrastructure scales
- −xAI's post-SpaceX-acquisition status (SpaceX bought xAI, announced 2026-02-02) means procurement teams at US-regulated orgs may need to re-run vendor-approval workflows. Bring that up early with your security team if you're enterprise
- −No consumer web UI -- this is API-only. ElevenLabs' Creator/Studio web apps remain easier for one-off podcast or audiobook work
- −Still API/console-first -- no polished consumer web app for one-off podcast or audiobook production the way ElevenLabs Studio offers, even with the new no-code Voice Agent Builder
Pricing
Speech to Text (batch)
- ✓REST API for large audio files
- ✓Word-level timestamps
- ✓Speaker diarization
- ✓Multichannel support
- ✓Inverse Text Normalization (numbers, dates, currencies)
Speech to Text (streaming)
- ✓Lowest-latency WebSocket API
- ✓Real-time speaker ID
- ✓Same accuracy as batch
- ✓Supports 25+ languages seamlessly
Text to Speech
- ✓Natural expressive voices (ARA voice etc.)
- ✓Speech tags: [laugh], [sigh], [whisper], <emphasis>, <slow>, <pause>
- ✓REST + WebSocket streaming
- ✓Usage-based billing, no hidden fees
Known Issues
- THE `grok-voice-latest` REPOINT HAPPENED ON SCHEDULE 2026-08-05 -- AND IT IS A 1.6x PRICE INCREASE IF YOU DIDN'T PIN (verified against xAI's own models endpoint 2026-08-06): the alias migration announced on 7/29 executed as promised. xAI's live model catalogue now attaches **`aliases: ["grok-voice-latest"]` to `grok-voice-think-fast-2.0`** across every cluster it publishes (us-east-1, eu-west-1, us-saltlake-2), and **`grok-voice-think-fast-1.0` now carries no aliases at all**. So any integration still calling `grok-voice-latest` is on Think Fast 2.0 as of 8/05, with no code change and no deploy on your side. **THE PART XAI DID NOT PUT IN THE ANNOUNCEMENT: this alias move raised the per-audio-second rate by exactly 1.6x.** xAI's catalogue prices Think Fast 2.0's realtime audio at **1.6 times** Think Fast 1.0's (13,333,333 vs 8,333,333 in the endpoint's own units). Against the vendor-published **$0.08 per minute** headline for Think Fast 2.0, that puts Think Fast 1.0 at **$0.05 per minute** -- so unpinned callers went from roughly $3.00 to $4.80 per hour of audio overnight. The capability jump is real (AA Speech-to-Speech Quality Index 82.9% vs 75.7%, time-to-first-audio 0.70s vs 1.25s -- see the entry below), and 2.0 uses 0.4x the reasoning tokens of 1.0, so the effective cost story depends on your workload. But if you budgeted voice spend off the 1.0 rate and did not pin, **check your August bill** -- this is a silent 60% unit-price increase on a default alias. Pinning still works: `grok-voice-think-fast-1.0` remains callable, it simply is not what `latest` means any moreSource: xAI models catalogue (docs.x.ai/docs/models -- alias and per-second pricing read directly off the live endpoint payload, fetched 2026-08-06), x.ai/news/grok-voice-think-fast-2 (announcement, 2026-07-29) · 2026-08-05
- NEW MODEL -- GROK VOICE THINK FAST 2.0, AND `grok-voice-latest` REPOINTS ON 2026-08-05 (2026-07-29, vendor-primary): xAI shipped '**our most capable speech-to-speech voice model**', publishing third-party benchmarks from Artificial Analysis. **AA Speech-to-Speech Quality Index 82.9%** vs 75.7% for Think Fast 1.0, ahead of GPT-Realtime-2.1 High (79.1%) and Gemini 3.1 Flash High (69.5%). Component scores: Big Bench Audio 97.2%, **Full Duplex Bench 95.1%** (up from 77.8%, essentially closing the gap to GPT-Realtime-2.1's 95.7%), tau-voice Bench **56.5%** (vs 45.7% GPT-Realtime-2.1), and **time to first audio 0.70s** vs 1.25s for 1.0 and 2.98s for Gemini 3.1 Flash. Transcription: xAI claims a **1.5-2.0x accuracy improvement over Deepgram Nova 3 and ElevenLabs Scribe v2** across thousands of short phrases in 24 languages, widening to **~10x in noisy settings**. Architecture note worth understanding: these models **reason while speaking** -- 'reasoning in parallel with speech makes the model substantially smarter than other speech-to-speech models with no impact on latency' -- and 2.0 uses **0.4x the reasoning tokens** of 1.0, so tool calls typically fire before the agent finishes its first sentence. **PRICING: $0.08 per minute of audio.** **ACTION REQUIRED FOR API USERS: on August 5, 2026 the `grok-voice-latest` alias moves from `grok-voice-think-fast-1.0` to `grok-voice-think-fast-2.0`.** No action is needed to upgrade; to stay on 1.0 you must **pin `grok-voice-think-fast-1.0` before then**. xAI says the upgrade needs no prompt edits, and cites a Starlink support-line A/B test showing higher sales conversion and support containment. BRANDING NOTE: the vendor page is titled under **SpaceXAI** while the footer still reads (c) 2026 X.AI LLC -- the brand has shifted, the legal entity name has notSource: xAI/SpaceXAI (x.ai/news/grok-voice-think-fast-2, fetched 2026-08-03 via curl -- x.ai 403s WebFetch) · 2026-07-29
- EXPANSION (2026-07-06): xAI released **21 new flagship TTS voices** (Lumen, Castor, Naksh, Atlas, Carina, Zagan, Helix, Orion, Luna, and more), bringing the lineup to 26 -- each cast for a specific job (support, characters, commentary, advertising, education) and natively multilingual across Grok Voice's 25+ languages. The original five (Ara, Eve, Leo, Rex, Sal) were retrained for more natural pacing/phrasing/emphasis. All are available in the realtime Voice Agent API, the Text-to-Speech API, and a **new no-code Grok Voice Agent Builder** in the xAI console. **Custom voice cloning from ~1 minute of audio is now exposed** (closes a prior gap vs ElevenLabs). Speech tags ([pause], <whisper>, <emphasis>, <soft>) control deliverySource: xAI (x.ai/news/new-flagship-voices) · 2026-07-06
- Launched 2026-04-17 -- expect first-week rate-limit surprises and occasional multilingual hiccups in the streaming path. xAI's console rate limits are documented but may be adjusted during the shakedown periodSource: xAI STT/TTS announcement · 2026-04
- The xAI-to-SpaceX acquisition (2026-02-02) means billing, compliance, and procurement flow through SpaceX. For US-regulated customers (healthcare, finance, defense) the vendor-approval pathway is new and may take longer than with a standalone AI vendorSource: xAI announcement, x.ai/news/xai-joins-spacex · 2026-02
- Speech tag control is powerful but underdocumented at launch -- expect trial-and-error to dial in the right [laugh]/[whisper]/<pause> mix for conversational TTSSource: xAI STT/TTS docs, BuildFastWithAI reviews · 2026-04
Best for
Developers building voice agents, real-time transcription tools, accessibility features, or high-volume TTS workloads where the cost per hour of audio actually matters at scale. Strong fit for phone-call and meeting transcription use cases where xAI's published WER advantage (5.0% on phone-call entities vs. ElevenLabs 12.0%) compounds quickly.
Not for
Consumer creators who want a polished web studio with voice presets and style sliders -- ElevenLabs Creator/Studio is still easier for one-off podcast or audiobook work. Enterprises in highly-regulated verticals should confirm the post-acquisition (SpaceX) vendor pathway works for them before committing.
Our Verdict
Grok Speech is xAI's clearest 'we are a platform, not just a chatbot' shot at the voice-API category, and on day-one pricing alone it's a credible threat to ElevenLabs, Deepgram, and AssemblyAI for production STT workloads. The published WER numbers are aggressive but plausible given the Tesla / Starlink deployment footprint. TTS at $4.20/1M char with real expressive tags undercuts ElevenLabs on price while narrowing the expressiveness gap. The open questions are (1) how it handles long-tail accents and non-English quality in practice, (2) whether the post-SpaceX procurement pathway slows enterprise adoption, and (3) how ElevenLabs responds on price. For new voice-API buyers shipping in Q2 2026, Grok Speech is now a first-call option alongside ElevenLabs and Deepgram.
Sources
- xAI models catalogue (grok-voice-latest alias now on Think Fast 2.0; per-audio-second pricing, verified 2026-08-06) (accessed 2026-08-06)
- xAI: Grok Voice Think Fast 2.0 (2026-07-29, announced the 2026-08-05 alias repoint) (accessed 2026-08-06)
- xAI: 21 New Flagship Grok Voices (2026-07-06) (accessed 2026-07-07)
- xAI: Grok Speech to Text and Text to Speech APIs (accessed 2026-04-18)
- xAI STT docs (accessed 2026-04-18)
- xAI TTS docs (accessed 2026-04-18)
- xAI joins SpaceX announcement (accessed 2026-04-18)
Explore more Grok Speech (STT + TTS APIs) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for Grok Speech (STT + TTS APIs).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to Grok Speech (STT + TTS APIs)
ElevenLabs
Best-in-class AI voice generation -- now includes 11.ai (MCP-based voice assistant), Eleven v3 expressive speech, and IBM watsonx partnership. $500M raise at $11B valuation (Feb 2026)
Murf AI
Text-to-speech that actually sounds like a real person read your script -- not a robot trying its best
Descript
Edit audio and video by editing text -- the 'Google Docs of media editing' actually lives up to the hype
Speechify
Text-to-speech reader that turns articles, docs, and PDFs into natural-sounding audio
Microsoft MAI-Voice-2
Microsoft's in-house expressive TTS model -- MAI-Voice-2 launched 2026-06-02 at Build: 15 languages (up from English-only), granular emotion-tag control, zero-shot voice cloning from a 5-60s clip, and preferred over MAI-Voice-1 72% of the time. In speaker-similarity tests its speech is 'indistinguishable' from real recordings. On Azure Foundry + integrated into VS Code and Dynamics 365 Contact Center; lower-cost MAI-Voice-2-Flash coming. Original MAI-Voice-1 shipped 2026-04-02
GPT-Live (ChatGPT Voice)
OpenAI's full-duplex voice models (launched 2026-07-08) now powering ChatGPT Voice -- listens and speaks at the same time, backchannels naturally, and delegates hard questions to GPT-5.5 in the background while keeping the conversation going. GPT-Live-1 for paid tiers, GPT-Live-1 mini for Free
Cohere Transcribe
Cohere's first audio model -- launched 2026-03-26 under Apache 2.0, 2B parameters, #1 on Hugging Face Open ASR Leaderboard (5.42 avg WER), 14 enterprise-critical languages. Free API with rate limits; Model Vault for production