A new multilingual speech-to-text model from SpaceXAI has achieved the top accuracy ranking for streaming transcripts on the Artificial Analysis leaderboard. This tool, released as Grok Voice Transcribe 2.0, records a 2.7% Word Error Rate (WER) on final transcripts and 3.4% on first partial transcripts. The system supports speaker diarization, multichannel transcription, and smart turn detection for both batch and streaming workloads.

Performance improves over the previous version's 3.9% streaming WER and 4.0% non-streaming WER, the latter of which now stands at 2.3% to place 5th among 59 models. Access through the SpaceXAI API remains at $0.20 per hour for streaming and $0.10 per hour for batch transcription. The model outperforms rivals including Muse Voice Transcribe and ElevenLabs Scribe v2 Realtime in transcription accuracy.