Grok Voice Transcribe 2.0 vs Muse Voice Transcribe: A 17-Day Reign and a Quarter Second

On September 1, 2026, Meta said Muse Voice Transcribe had taken first place in Artificial Analysis's streaming speech recognition evaluation with a 3.1% final word error rate, and that claim stood unchallenged for seventeen days. On September 18, SpaceXAI shipped Grok Voice Transcribe 2.0 at 2.7% on the same board and took the slot. Two models, one ranking, seventeen days apart — and the more interesting comparison is not the 0.4-point gap in accuracy, because Meta's model is still roughly three times faster to finalize a sentence: 0.16 seconds after end of speech against Grok Voice Transcribe 2.0's 0.49 seconds. Muse Voice Transcribe is a real-time audio perception model built by Meta Superintelligence Labs to run inside Meta's own products, where a quarter second is the difference between an assistant that feels present and one that feels slow. Grok Voice Transcribe 2.0 is a transcription API built to be called by anyone, at a price that makes batch work viable. Both numbers are vendor-reported on launch day, and only one of the two has since been independently placed on a public board.

What the leaderboard actually says now

The AA-WER Streaming board is a weighted composite — AA-AgentTalk at 50%, VoxPopuli at 25%, Earnings22 at 25%, roughly eight hours of mostly agent-shaped English audio — and both models are measured on it, which is what makes this the rare head-to-head where the numbers are at least commensurable. Side by side, vendor-reported figures included:

• Final transcript accuracy — Grok Voice Transcribe 2.0 at 2.7% WER vs Muse Voice Transcribe at 3.1%

• First partial transcript accuracy — 3.4% vs 3.6%

• Time to first partial — 0.49 seconds vs 0.13 seconds

• Time after end of speech to final transcript — 0.49 seconds vs 0.16 seconds

• Speaker diarization error rate — included, no figure published vs an average of 17.5% across Meta's evaluation

Read the accuracy rows and the latency rows separately, because they point in opposite directions and both are true. On accuracy the two models are within four tenths of a point of each other for final transcripts and two tenths for first partials — a gap small enough that it will reverse on the next release from either company, and small enough that no reasonable person should choose between them on it. On latency they are not close. Meta's model returns a partial in about an eighth of a second and finalizes in about a sixth of a second, against roughly half a second in both directions for SpaceXAI's.

That is the entire comparison in one line: Grok Voice Transcribe 2.0 spent latency to buy accuracy, and Muse Voice Transcribe did the opposite. SpaceXAI took its first-partial error rate from 18.3% in version 1.0 down to 3.4% in 2.0, and the price of that was going from 0.25 seconds to 0.49 seconds on the same measure. Meta's model was engineered the other way, with 80-millisecond audio chunks and a reinforcement-learned adaptive delay that decides how long to wait before committing to text — a mechanism whose whole purpose is to hold accuracy while keeping the commit fast. Neither choice is a mistake. They are answers to different questions.

Which question you are asking

If the transcript feeds a voice agent that has to respond inside a conversational pause, 0.16 seconds and 0.49 seconds are different products. Human turn-taking tolerates a gap of a few hundred milliseconds before the interaction starts to feel wrong, and the latency budget is shared — transcription, then reasoning, then speech synthesis. A model that hands back a final transcript in a sixth of a second leaves room for the rest of the pipeline; one that takes half a second consumes most of the budget before the model has thought about anything. That is what Meta built for, and it is why the model runs inside Meta AI on the Mac and inside Muse Code rather than being offered primarily as an endpoint for other people's pipelines.

If the transcript is a document — a call recording, a meeting, a video library, a compliance archive — then half a second is nothing, the accuracy gap is still nothing, and the only number that matters is what an hour of audio costs. Which is where these two diverge sharply, and in a direction that surprises people who only look at the streaming rate.

Half the price in batch, a tenth more in streaming

Meta lists Muse Voice Transcribe at $0.18 per hour of audio, or $3.00 per 1,000 minutes. Grok Voice Transcribe 2.0 lists at $0.10 per hour for batch and $0.20 per hour for streaming. So:

• Streaming — Grok Voice Transcribe 2.0 at $0.20 per hour vs Muse Voice Transcribe at $0.18, so Meta is about 10% cheaper

• Batch or recorded — $0.10 per hour vs $0.18, so SpaceXAI is 44% cheaper

• Included at list price — diarization, word timestamps with confidence, key-term biasing and inverse text normalization on the SpaceXAI side vs streaming transcription, diarization and endpoint detection as one model on Meta's

• Free tier or self-host — neither, on both counts

A team that only ever streams should be indifferent between them on price and should choose on latency, which means Meta. A team with any meaningful volume of recorded audio should look at the batch row, because $0.08 an hour compounds: 5,000 hours a month is $500 on one and $900 on the other, and the accuracy difference between them is smaller than the rounding on either number.

The licensing position is identical in the way that matters and different in the way that does not. Both are closed weights — Muse Voice Transcribe is proprietary and available only through Meta's hosted API, and Grok Voice Transcribe 2.0 is a commercial API with no download. Neither is an option if your requirement is that audio never leaves infrastructure you control, and both leave you exposed to a vendor changing its price, its model version or its deprecation schedule under you. That is a real consideration and not a hypothetical one: Grok Voice Transcribe 1.0 is already on a deprecation path less than two weeks after its successor shipped.

Diarization, which is the feature neither of them leads with

Both models do speaker attribution in a single pass, which two years ago was a separate system glued on afterwards, and the comparison here is unusually concrete because Meta published a number and SpaceXAI did not.

Meta reports an average diarization error rate of 17.5% across its evaluation, handles 20 or more speakers without post-processing, and handles them as part of the streaming model rather than as a downstream step — the "who said what" arrives with the text rather than after it. That is genuinely hard to do in a streaming context, where a speaker turn is only identifiable once you have heard enough of it, and 17.5% is a real number with a real caveat: it is Meta's own, averaged over an evaluation set Meta chose.

SpaceXAI's model includes diarization at no extra cost and also supports up to eight independent audio channels in a single request, which is a different way of solving the same problem — if your audio arrives as separate per-participant channels, which contact-centre telephony often does, channel separation is more reliable than diarization and needs no error rate at all. If your audio arrives as one mixed stream, you are dependent on the diarization quality, and only one of these two vendors has told you what that is.

There is a second-order difference worth noting: Muse Voice Transcribe treats diarization, transcription and endpoint detection as one model producing one output, which means the components cannot fail independently. Grok Voice Transcribe 2.0 lets you take channels, key terms and diarization as separate levers. A single model is simpler to run and harder to debug.

Languages, and where the two model cards stop being comparable

Meta trained Muse Voice Transcribe on more than 70 languages and verified 25 of them extensively at launch, with native code-switching inside and between sentences and support for audio longer than an hour. Grok Voice Transcribe 2.0 handles automatic language detection with switching mid-recording and applies inverse text normalization — spoken dates, currencies, phone numbers and email addresses rendered in written form — across 25 languages.

The overlap is the 25 extensively verified languages on one side and the 25 normalization languages on the other, and the two sets are not the same 25 and neither vendor has published the list. "70+ languages trained" and "25 languages with formatting support" are not comparable claims and should not be subtracted from each other. What can be said is that Meta is making a broader multilingual claim and SpaceXAI is making a narrower but more operational one: detecting the language is table stakes, and formatting what the model heard into something a downstream system can parse is the part that breaks integrations when it is missing.

The choice, and the reason it may not be yours to make

Strip the marketing and the decision reduces to three sentences. Pick Muse Voice Transcribe if the transcript is consumed live inside a conversation, because 0.16 seconds is not a detail. Pick Grok Voice Transcribe 2.0 if you have recorded audio at volume, because half the batch price is not a detail either. If you need neither extreme, pick on whatever else constrains you — region, contract, existing integration — because the accuracy difference will not settle it.

The complication is that one of these two models is not really for sale in the usual sense. Muse Voice Transcribe exists to make Meta's own assistant feel fast, and it is available through the Meta Model API largely because Meta has decided the ecosystem value of exposure exceeds the value of exclusivity. That could change, and it could change quickly: Meta spent the same week opening Muse's connector platform to third-party developers, which is a company investing in its own distribution channel rather than in being a neutral supplier to everyone else's. Building a production dependency on a competitor's internal model is a bet that the terms will not change, and that bet has a bad historical record.

That asymmetry is the argument for not hard-coding either of them. OrcaRouter exposes 200+ models behind a single OpenAI-compatible key, with automatic failover across providers and 0% markup on provider list price — so when SpaceXAI flips the default to 2.0 and deprecates 1.0, or when Meta repositions its speech API, the change is a model string rather than a migration project. For a category where the leader has changed hands twice in three weeks, the routing layer is the part of the stack that should be stable, and the model is the part that should not be.

One thing to watch over the next month: whether Artificial Analysis re-measures Muse Voice Transcribe on the streaming board now that Grok Voice Transcribe 2.0 has displaced it. Meta's 3.1% and SpaceXAI's 2.7% were both reported at launch, but only SpaceXAI's has been independently placed. If Meta's number holds up when someone re-runs it, the accuracy gap is as small as it looks and the latency gap decides everything. If it does not, the seventeen-day reign was the whole story.