SpaceXAI has released Grok Voice Transcribe 2.0, an update to its speech-to-text model that the company calls the most accurate transcription system available. Independent benchmarking from Artificial Analysis backs up at least part of that claim: the new version scores a 2.7% word error rate on final transcripts, down from 3.9% for version 1.0, and currently ranks first among 32 streaming transcription models the firm tracks.
The bigger improvement is in how quickly the model gets a transcript right. One analysis found the word error rate on the model's first partial transcript, the text shown before a speaker has finished, fell from 18.3% on version 1.0 to 3.4% on version 2.0. That matters for live captioning and voice interfaces, where early guesses are often what a user actually sees. The trade-off is latency: the new model takes roughly a third of a second longer than its predecessor to return both partial and final transcripts.
Pricing has not changed: $0.10 per hour of audio for batch processing and $0.20 per hour for streaming. Developers can pick version 2.0 through SpaceXAI's existing Speech-to-Text API, though version 1.0 remains the default when no model is specified. SpaceXAI says that will change, with 2.0 becoming the default and 1.0 being deprecated in the coming weeks. Atlassian's Loom is already using the model to transcribe video content.
Source: MarkTechPost
