MAI-Transcribe-2 vs Muse Voice Transcribe: The Same Week, Two Opposite Bets on Audio

The week of September 1, 2026 produced two speech models from two frontier labs, and MAI-Transcribe-2 and Muse Voice Transcribe are best understood as opposite answers to the question of where audio intelligence should live. Muse Voice Transcribe, released by Meta Superintelligence Labs on September 1, is a real-time audio perception model: it transcribes, separates more than twenty speakers, and detects when a speaker has finished, all in one streaming pass over 80-millisecond chunks of live audio, and it tops the streaming speech-to-text board it was measured on. MAI-Transcribe-2, released by Microsoft two days later, on September 3, is the opposite bet: a batch engine that does not chase live latency at all, and instead pours everything into transcribing pre-recorded files at the best word-error rate on the independent non-streaming board, roughly 411× real time, for about $1.67 per 1,000 minutes. Comparing them head-to-head as "transcription models" hides the actual decision, which is about the moment in the audio's life you need words: while it is happening, or after it is over.

Two products pointed at different moments

Muse Voice Transcribe is built for the moment the audio is still happening. It is an autoregressive multimodal model in Meta's Muse family that folds three jobs into one pass — automatic speech recognition, speaker diarization for more than twenty speakers, and endpointing, the detection that a speaker has stopped talking so the system can respond. Meta reports that its final transcript lands about 0.16 seconds after a speaker stops, with a 3.1% word-error rate on those final transcripts and 3.6% on first partials — figures Meta published on launch day, citing the Artificial Analysis streaming leaderboard, and that had not been independently reproduced as of this writing. It handles audio longer than an hour in a single stream, switches languages mid-sentence, and was trained on 70+ languages with 25 extensively verified at launch. Its price is $3.00 per 1,000 minutes of audio, about $0.18 per hour, through the Meta Model API. Meta has confirmed the weights will not be opened.

MAI-Transcribe-2 is built for the moment the audio is already a file. Microsoft's model measures 2.0% AA-WER on Artificial Analysis's non-streaming leaderboard — second place, and the best number among the managed batch APIs — and runs at roughly 411× real time, which is a throughput figure, not a latency promise. It transcribes in 60 languages with automatic language identification, bundles speaker diarization, word-level timestamps, keyword biasing, and verbatim-versus-clean styles, and handles code-switching such as Hinglish and Spanglish. It is available through Microsoft Foundry in public preview and the MAI Playground at $0.10 per audio-hour as a limited-time launch offer through the end of the year, with no streaming product announced. Microsoft's launch post positions it explicitly for long-form and batch audio — the workloads where a streaming model's low latency is worthless but its per-minute cost is not.

Why the two accuracy numbers do not fight

The natural instinct is to compare 2.0% to 3.1% and declare a winner, and it would be wrong twice over. First, the numbers measure different tasks: MAI-Transcribe-2's 2.0% is non-streaming accuracy on pre-recorded audio, where the model can look at the whole file; Muse Voice Transcribe's 3.1% is streaming accuracy on final transcripts, produced under the constraint of transcribing as the audio arrives. Streaming is the harder problem, which is why the streaming leaderboard and the non-streaming leaderboard are separate boards with separate scores. Second, the labels are different weights: MAI-Transcribe-2's figure is an Artificial Analysis measurement of the model, while Muse Voice Transcribe's is Meta's launch-day report of what Artificial Analysis measured — vendor-reported and unreproduced. The honest statement is that both models are credible claims at the top of their respective boards, and neither number tells you anything about the other lane.

Diarization is where the real comparison lives, because it is the one feature both products carry and the one where their ambitions visibly differ. Muse Voice Transcribe separates more than twenty speakers as a core streaming capability, with Meta reporting an average speaker diarization error rate of 17.5% against a competitor range it cites of 21.1% to 28.6% — again Meta-reported and unverified. MAI-Transcribe-2 bundles diarization too, but Microsoft publishes no speaker ceiling and no diarization error rate at all. For a two-party call, both products are fine and the difference is immaterial. For a twelve-person focus group or a panel recording, Muse Voice Transcribe is the only one of the two whose multi-speaker ceiling is even stated, and MAI-Transcribe-2's unmeasured diarization is the single biggest reason to test it before trusting it with your harder audio.

The price comparison that makes sense

On price the two are closer than the batch-versus-streaming framing suggests, because $1.67 per 1,000 minutes (MAI-Transcribe-2, early-bird) and $3.00 per 1,000 minutes (Muse Voice Transcribe) are both at the cheap end of managed speech. The comparison only makes sense within a lane. For pre-recorded batch transcription, MAI-Transcribe-2 is less than half the price of the streaming-native model while carrying the better non-streaming WER — but you would only route recorded files to Muse Voice Transcribe if you specifically wanted its streaming pipeline, which is not the efficient way to process an archive. For live meeting audio, Muse Voice Transcribe is the only product in this article that works at all, and its $3.00 rate undercuts the other real-time transcription APIs it launched against — Gemini 3.5 Transcribe Live at about $9 per 1,000 minutes, GPT Live Transcribe at $17 — while adding more-than-twenty-speaker diarization those products do not match. The honest price headline is that Meta set the streaming price low and Microsoft set the batch price lower, and both numbers are promotional-era pricing that will drift once each vendor names its standard rate.

Which one for which audio

If your audio is live — meetings being transcribed in real time, voice agents, live captioning, anything where words are needed while they are being spoken — Muse Voice Transcribe is the pick, and it is not close. It is the only streaming model here, it separates more than twenty speakers in the same pass, and it was priced to win that lane on day one. Its weaknesses are the ones to carry into a trial: the accuracy and diarization figures are Meta-reported, and a week-old streaming model has not yet shown how it behaves on the messy audio that lives outside launch benchmarks.

If your audio is already files — archives, call recordings, media libraries, anything processed in bulk — MAI-Transcribe-2 is the better number on every axis that matters: lower measured WER, roughly an order of magnitude more throughput, and a per-1,000-minute price below the streaming model's. Its risks are the mirror image of Muse's: four days old, no streaming SKU, diarization quality unmeasured, and an early-bird rate with an undisclosed standard price behind it.

The Microsoft launch page that introduces MAI-Transcribe-2 lists features Microsoft bundles that the streaming rival does not offer in the same pass, and it is the document to check when you are deciding which claims are product surface area and which are still benchmark promises.

And if the transcript is only the first half of your pipeline, the choice between these two matters less than the choice you make above them. Live meeting audio transcribed by Muse Voice Transcribe still needs summarization, action-item extraction, and follow-up drafting before it is useful; archived calls transcribed by MAI-Transcribe-2 still need to be searched, redacted, or routed to the right downstream system. That is the layer where a multi-model platform earns its keep — OrcaRouter's one API across 200+ models, with provider list prices passed through at 0% markup, lets that downstream work call a different model per workload over a single integration, so whichever speech model wins your pilot, the stack above the transcript does not have to be rebuilt to switch. Neither Muse Voice Transcribe nor MAI-Transcribe-2 is on that roster today; both live on their vendors' own APIs. The bet this week really settled is narrower and more useful: real-time and batch transcription have both just become cheap enough that the differentiator is no longer the words — it is what you do with them.