SpaceXAI Releases Grok Voice Transcribe 2.0, Claiming 2x Accuracy at Same Price

The company says the model targets difficult audio, including noisy phone lines, competing voices, local accents and spoken credentials. It is built on the audio foundation model behind Grok Voice. SpaceXAI says Grok Voice already handles tens of thousands of customer-support calls a day, transcribes millions of hours of video narration and runs the Grok assistant in Tesla vehicles. The training data is live, noisy, multilingual audio recorded across diverse environments, and the company refined the model with post-training.

SpaceXAI reports that Grok Voice Transcribe 2.0 ranks first in accuracy among 32 streaming models on the public Artificial Analysis leaderboard. The AA-WER Streaming benchmark uses about eight hours of audio and weights AA-AgentTalk at 50%, VoxPopuli at 25% and Earnings22 at 25%. The company also measures word error rate on four internal sets drawn from production traffic: English customer-support calls at 8 kHz, English conversations with Grok, English phone numbers, emails and addresses, and short voice-assistant utterances in 19 languages. Version 2.0 improves on 1.0 across all four sets, and SpaceXAI says it leads every model the company tested on telephony. Those internal results are vendor-reported and have not been independently reproduced.

Multilingual transcription is the largest improvement over version 1.0, according to SpaceXAI. The model transcribes dozens of languages, detects language automatically and follows mid-recording language switches in a single pass. On short phrases such as in-car commands, word error rate falls from 20.6% to 6.8%, a reduction of about 67% in word errors. The documentation lists 25 languages for written-form formatting of numbers, currencies and units.

The API includes batch and streaming transcription. Developers can transcribe files and URLs or stream audio over WebSocket at wss://api.x.ai/v1/stt. It returns word-level timestamps with start and end times plus confidence scores, speaker diarization at no additional cost, and multichannel transcription for up to eight channels independently. Key term biasing accepts up to 100 domain terms per request, each up to 50 characters. Text formatting returns numbers, dates, currencies, phone numbers and emails in written form, and filler words such as “um” and “uh” are removed by default. A machine-learning model also provides smart turn detection for voice agents. The batch endpoint accepts files up to 500 MB across 12 audio formats. Streaming accepts Opus at roughly 4 KB/s, compared with 48 KB/s for raw PCM at 24 kHz.

Pricing is identical to version 1.0. Batch transcription costs $0.10 per hour of audio, and streaming costs $0.20 per hour. Diarization, timestamps and key terms are included. SpaceXAI says that equals about $1.67 and $3.33 per 1,000 minutes. To call the model, developers send a POST request to https://api.x.ai/v1/stt with the model set to grok-voice-transcribe-2.0.

Atlassian Loom has adopted Grok Voice Transcribe 2.0 to transcribe every video after finding it more accurate than its existing solution. SpaceXAI describes the workflow as record, transcribe, then code: a user records an action plan in Loom, and the transcript is piped into Cursor, which makes code updates. Sanchan Saxena, SVP of Teamwork Collection at Atlassian, described the process as “closing the loop from context to code.”