Grok Voice Transcribe 2.0 drops to a 2.7% error rate in streaming

Grok Voice Transcribe 2.0 improves transcription across conversations, voice commands, and multilingual audio. SpaceXAI's model takes the top spot for streaming accuracy on Artificial Analysis while keeping pricing at $0.10 per hour for batch and $0.20 per hour for streaming.

The biggest gains show up on difficult audio

A poor phone connection, a dictated email address, or a voice command lasting only a few words gives a transcription system very little room for error. Those are the situations SpaceXAI targets with Grok Voice Transcribe 2.0, its new speech-to-text model built on the same audio foundation as Grok Voice.

According to the company, its training includes noisy, multilingual audio recorded across different real-world environments, followed by additional post-training. SpaceXAI says the underlying audio model is already used for customer-support calls, video narration, and the Grok assistant inside Tesla vehicles.

Across its internal evaluations, the company reports improvements in four areas: phone calls, conversations with Grok, spoken information such as phone numbers and email addresses, and short voice commands across 19 languages. On the short-phrase test, Word Error Rate falls from 20.6% with Transcribe 1.0 to 6.8% with 2.0.

The results vary significantly by use case. On English customer-support calls, Transcribe 2.0 records a 7.1% WER compared with 10.6% for its predecessor. On conversations with Grok, the gap widens from 8.7% to 3.3%. These remain company-run evaluations. First for streaming accuracy on Artificial Analysis

An independent benchmark provides a broader reference point. On Artificial Analysis' public leaderboard, Grok Voice Transcribe 2.0 records a 2.7% WER for final streaming transcripts, compared with 3.9% for the previous generation. That placed it first for streaming accuracy at release.

The public improvement is therefore more modest than SpaceXAI's broader “twice as accurate” claim, which draws from the company's own evaluation sets.

Higher accuracy also comes with a trade-off. Artificial Analysis measures 0.49 seconds to a final transcript, compared with 0.37 seconds for Transcribe 1.0. ElevenLabs Scribe v2 Realtime reaches 0.14 seconds on the same metric. Transcribe 2.0 therefore favors lower transcription error over minimum latency. Switching languages mid-recording

Multilingual transcription accounts for one of the largest changes in this generation. The model automatically detects languages and can follow language switches during the same recording without requiring a separate pass.

SpaceXAI's Speech-to-Text documentation also confirms several features aimed at production use: word-level timestamps with confidence scores, speaker diarization, independent transcription across up to eight channels, and support for up to 100 key terms to improve recognition of specialized vocabulary.

Numbers, dates, currencies, phone numbers, and email addresses can be formatted directly in the transcript. The system can also remove filler words and detect when a speaker has finished a turn, a feature designed in part for voice-agent applications. Loom connects speech to code

Atlassian is already using Transcribe 2.0 in Loom after finding it more accurate than its previous transcription solution. That creates a direct workflow: someone records instructions in Loom, then exports the transcript to Cursor, where those instructions can be turned into code changes.

Pricing remains unchanged from Transcribe 1.0 at $0.10 per hour of batch audio and $0.20 per hour for streaming, with diarization, timestamps, and key-term biasing included.

The current API documentation now lists `grok-voice-transcribe-2.0` as the default model, while version 1.0 can still be explicitly selected.