Grok Voice Transcribe 2.0 Cuts Short-Phrase WER From 20.6% to 6.8%

SpaceXAI has shipped Grok Voice Transcribe 2.0 under the model ID grok-voice-transcribe-2.0, available immediately as a hosted API through its Speech to Text endpoint. The company claims double the accuracy of version 1.0 at identical pricing — $0.10 per audio hour for batch and $0.20 per hour for streaming — with no open weights released, making self-hosting unavailable. For engineers running high-volume transcription pipelines, an unchanged cost structure paired with a reported accuracy jump demands a close look at what the benchmarks actually show.

Benchmark claims and what they cover

SpaceXAI reports a first-place finish among 32 streaming models on the Artificial Analysis leaderboard's AA-WER Streaming benchmark, which uses approximately 8 hours of audio weighted as 50% AA-AgentTalk, 25% VoxPopuli, and 25% Earnings22. Beyond that external ranking, the company publishes results on four internal production-traffic sets: telephony audio at 8 kHz (English customer support calls), conversational audio (English conversations with Grok), credentials (phone numbers, emails, and addresses in English), and short multilingual phrases across 19 languages for voice-assistant utterances. Version 2.0 improves on 1.0 across all four sets.

The only hard versioned figure published is on the short-phrases set: word error rate falls from 20.6% to 6.8%, representing roughly 67% fewer word errors. On telephony, SpaceXAI states it leads every model the company internally tested; no absolute WER figure is given for that set. The internal sets are entirely vendor-reported and not independently reproduced.

SpaceXAI names multilingual transcription as the largest single improvement over 1.0. The model performs automatic language detection and handles mid-recording language switches in a single pass. Written-form formatting — numbers, currencies, units — is documented for 25 languages.

Architecture and feature surface

Grok Voice Transcribe 2.0 is built on the same audio foundation model that powers the broader Grok Voice system, which SpaceXAI says already handles tens of thousands of customer-support calls daily, transcribes millions of hours of video narration, and runs the Grok assistant inside Tesla vehicles. The training corpus is live, noisy, multilingual audio recorded across diverse environments, with post-training applied on top. No additional architectural specifics — layer counts, parameter totals, or attention mechanisms — are disclosed.

Every feature ships without per-feature surcharges. Speaker diarization, word-level timestamps with per-word confidence scores, and key term biasing (up to 100 terms, each up to 50 characters) are included at the base price. The batch endpoint accepts files up to 500 MB across 12 audio formats. The streaming WebSocket endpoint at wss://api.x.ai/v1/stt supports Opus at approximately 4 KB/s, compared to 48 KB/s for raw PCM at 24 kHz — a bandwidth difference relevant for mobile or constrained-network deployments.

Smart turn detection uses an ML confidence model to predict end-of-turn for voice agent workflows. The source documents a confidence threshold of approximately 0.983 in example events, with mid-sentence pauses scoring around 0.005, contrasted with a fixed 400 ms silence timer as the alternative. Requests that omit the model field currently resolve to 1.0, so callers must set model=grok-voice-transcribe-2.0 explicitly.

Atlassian Loom has adopted the model to transcribe every video on its platform, piping results into Cursor for code generation — a workflow SpaceXAI frames as record, transcribe, then code.

Pricing in competitive context

Competitor pricing figures are not present in the source material and are omitted. What is confirmed: SpaceXAI held its price line from 1.0 to 2.0, meaning any accuracy improvement arrives at zero incremental cost to existing users who pin to the new model ID explicitly.

AI Mastery analysis

The architectural story here fits a pattern we have tracked repeatedly: gains that look like model improvements are often driven by infrastructure and training-data decisions rather than scaling model weights. SpaceXAI's foundation is live production traffic — noisy, multilingual, spanning phone lines and in-car microphones — which is a data-infrastructure advantage a larger but more generic model cannot easily replicate. The multilingual short-phrase WER drop from 20.6% to 6.8% is the only hard number comparing versions, and it happens to be the domain where sparse context makes language identification hardest. That specificity of improvement suggests targeted post-training on a known failure mode rather than a broad architecture overhaul.

Two limitations deserve attention. First, the AA-WER Streaming benchmark uses approximately 8 hours of audio; production pipelines processing domain-specific vocabulary, rare accents, or heavily overlapping speech should validate independently before committing. SpaceXAI does not disclose absolute WER figures for telephony or conversational sets — only directional improvement claims. Second, the API-only constraint removes the option to run transcription inside a private network boundary, which pricing alone cannot resolve for teams with data-residency requirements.

For streaming voice agent workflows, Smart Turn is worth examining carefully. Replacing a fixed silence threshold with a per-pause confidence score is a meaningful change for multi-turn voice interfaces, but the documented examples come from a credentials-reading scenario — structurally different from open-ended conversational turns. Engineers should test Smart Turn against their own dialogue distributions rather than assuming the telephony benchmark transfers.

The competitive pressure on speech-to-text APIs has shifted from raw accuracy on clean audio to robustness on degraded, multilingual, and domain-specific inputs — and pricing parity is now table stakes. SpaceXAI's production deployment footprint gives it a training data advantage that is harder to close than a parameter gap, and Grok Voice Transcribe 2.0 is a credible claim on that position, contingent on independent verification.

Primary source

SpaceXAI Releases Grok Voice Transcribe 2.0: A Speech-to-Text API Claiming 2x Accuracy Over 1.0 at $0.10 per Hour — MarkTechPost