Speech recognition vendors tend to sell accuracy and price as a trade-off. SpaceXAI is holding the second flat while claiming more of the first.
Grok Voice Transcribe 2.0 arrived on September 18 with batch pricing unchanged at $0.10 per hour of audio and a company claim that it is twice as accurate as version 1.0. It inherits its audio foundation model from Grok Voice, already used for support calls, video narration and voice agents inside physical products, including the assistant in Tesla vehicles.
For a ranking, SpaceXAI points to Artificial Analysis, where it says the model leads 32 streaming competitors on accuracy. Its own benchmarks are the more revealing ones. Four production sets measure word error rate: 8 kHz telephone audio from support lines, conversations with Grok, spoken account codes and email addresses, and brief multilingual commands. The company reports improvement on all four.
Multilingual speech shows the widest gain. Across a short-phrase set of assistant utterances in 19 languages, word error rate drops to 6.8 percent from 20.6 percent. Language detection is automatic, and the model follows speakers who switch languages mid-recording without a second pass.
The Speech-to-Text API handles uploaded files and live streams alike. Rivals named in the comparisons include Gemini 3.5 Transcribe, MAI-Transcribe-2, Whisper Large v3, Deepgram Nova-3 and ElevenLabs Scribe v2.
