Accuracy is not the same as snappiness
Grok’s streaming board lead is real. Muse still returns a finished turn faster. CastTranscript shows partial text during speech so the slower finalization does not leave a blank page.
LIVE TRANSCRIPTION
Grok Voice Transcribe 2.0 vs Muse Voice Transcribe
Pick the live pass by streaming accuracy, latency and what becomes an editable conversation, not by a launch headline alone.
Benchmarks and product roles checked September 20, 2026
The short answer
Grok owns the live pass. Muse remains the latency fallback.
Grok Voice Transcribe 2.0 ranks first on Artificial Analysis AA-WER Streaming for both final-transcript and first-partial accuracy. Muse Voice Transcribe returns a finished turn sooner. CastTranscript needs a transcript-first stream that becomes a private document, then an optional MAI publishing pass.
Live workflowGrok Voice Transcribe 2.0Muse Voice Transcribe
CastTranscript role
Default live recording path when an xAI key is configured
Configured fallback when Grok is unavailable
AA-WER Streaming
2.7% final WER and 3.4% first-partial WER at 0.49s after end of speech
3.1% WER at 0.16s after end of speech
Published price
$0.20 per audio hour for streaming, diarization included
Meta lists $0.18 per audio hour
What the room gets
Interim text while speaking, speaker labels and word timing when returned
Speaker turns with sentence timing; word timing is not assumed
Session architecture
Browser PCM16 at 16 kHz through a sticky relay to the xAI STT WebSocket
Browser audio frames through the same relay to the Muse WebSocket
Refinement
Optional, explicit MAI batch pass after the call
The same explicit MAI pass, never a silent second model
Do not stitch the two streams
Each live session uses one provider. The job record stores the model that actually ran. Missing speakers or word times stay visible instead of being patched by a second live model.
Publishing still waits for a choice
A MAI refine pass can add a more deliberate subtitle contract after the room closes. It costs another credit per media minute and only starts when the user asks.
CastTranscript decision
Grok 2.0 is the live default. Batch stays with MAI.
Live conversations use Grok Voice Transcribe 2.0 because the current streaming accuracy and included diarization fit a transcript-first room. Uploaded files stay on MAI-Transcribe-2. Muse remains wired as a fallback, not as a silent overlay.
Start a live transcriptQuestions, answered
Grok and Muse live questions
Which model does CastTranscript use for live recording?
The live path defaults to Grok Voice Transcribe 2.0 when an xAI key is configured. Muse Voice Transcribe remains the fallback. Each session records the provider that actually ran.
Is Grok faster than Muse?
No. On the Artificial Analysis streaming board Grok is more accurate and Muse returns a finished transcript sooner. CastTranscript shows interim text so the live document still updates while someone is speaking.
Does a live transcript include word-level subtitle timing?
Grok can return word timing and speaker labels. CastTranscript stores those fields when they arrive and still treats a separate MAI refine pass as the publishing workflow.
Will CastTranscript mix Grok and Muse in one session?
No. One live session uses one streaming model. A dropped Grok connection does not silently continue on Muse.
Private by default
Capture the room first. Decide on refinement later.
Start a private live transcript with transparent minute reservation and an explicit post-call publishing pass.