September 19, 2026
Streaming transcription has a new leader: Grok Voice Transcribe 2.0 delivered 2,7% errors
On September 18, 2026, Grok Voice Transcribe 2.0 ranked first for streaming transcription accuracy on AA-WER: 2.7% errors 0.49 seconds after speech ended. Its first partial result was 3.4%, and in batch mode it ranked 5th out of 59 models with 2.3% AA-WER. Streaming costs $0.20 per hour, batch costs $0.10, and the price matches the previous version.
Artificial Analysis
@artificialanlys
SpaceXAI released Grok Voice Transcribe 2.0 and took first place for Final Transcript and First Partial Transcript accuracy in AA-WER Streaming: 2.7% WER 0.49 seconds after speech ended. Grok Voice Transcribe 2.0 is a new speech-to-text model from @SpaceXAI that replaces Grok Voice Transcribe 1.0, previously known as Grok Speech to Text. Final Transcript streaming WER improved from 3.9% to 2.7%, while non-streaming AA-WER improved from 4.0% to 2.3%. The model is available through the SpaceXAI API for streaming and non-streaming transcription at the same price as its predecessor. Key points ➤ Final Transcript: Grok Voice Transcribe 2.0 reaches 2.7% WER 0.49 seconds after speech ended. It is more accurate but slower than Muse Voice Transcribe, which scored 3.1% with 0.16-second latency, and ElevenLabs Scribe v2 Realtime, which scored 3.6% with 0.14-second latency. It is also more accurate but slightly slower than Cartesia Ink-2 with semantic end-of-turn detection, which scored 3.4% and 0.43 seconds. ➤ First Partial Transcript: the model reaches 3.4% WER in 0.49 seconds. It is slightly more accurate than Muse Voice Transcribe and ElevenLabs Scribe v2 Realtime, both at 3.6%, but slower, as both models produce a result in 0.13 seconds. It is more accurate but slower than Cartesia Ink-2 with external end-of-turn detection, which scored 4.9% and 0.17 seconds, as well as Cartesia Ink-2 with an external endpoint, which scored 4.0% and 0.07 seconds. ➤ Non-streaming transcription: Grok Voice Transcribe 2.0 achieved 2.3% AA-WER and ranked 5th out of 59 models. Above it are Alibaba's Fun-Realtime-ASR-preview and StepFun's StepAudio 3 ASR at 1.7%, Microsoft AI's MAI-Transcribe-2 at 2.0%, and ElevenLabs Scribe v2 at 2.2%. Its speed factor is about 160x. ➤ Price: Grok Voice Transcribe 2.0 costs $0.20 per hour of streaming transcription, or $3.33 per 1000 minutes. This is slightly more expensive than Muse Voice Transcribe at $3, cheaper than Cartesia Ink-2 at $4, and about half the price of ElevenLabs Scribe v2 Realtime and Deepgram Flux, which each cost $6.50. Non-streaming transcription costs $0.10 per hour, or $1.67 per 1000 minutes. Congratulations to @elonmusk and the @SpaceXAI team on the launch! More below ⬇️
· 119.4K views
Grok Voice Transcribe 1.0 had 3.9% errors in final streaming text and 4.0% in batch transcription. Version 2.0 reduced those figures to 2.7% and 2.3%, so completed transcripts need fewer manual corrections.
Workflow. Atlassian Loom chose 2.0 after comparing it with its previous solution and uses it for every video. In the described workflow, a transcript from Loom is passed to Cursor, which makes changes to the code.
Integration. Batch transcription runs through `/v1/stt` with the `model=grok-voice-transcribe-2.0` parameter, while streaming uses WebSocket. The API accepts files up to 500 MB and up to 8 channels, while diarization, timestamps, and key terms are included in the price. Model 1.0 currently remains the default, so 2.0 must be specified explicitly.
In the coming weeks, SpaceXAI plans to make 2.0 the default model and mark 1.0 as deprecated.
