Title card reading 'Scribe v3 vs MAI Transcribe 2' with the subtitle 'accuracy or throughput — pick your bottleneck', beside two bar pairs labelled 'Accuracy (AA-WER)' at 1.95% and 2.04% and 'Speed (x real time)' at 76.8x and a much longer 338.5x, with a footer reading 'All figures per Artificial Analysis, AA-WER v2.2, non-streaming'.
Guides & Insights

Scribe v3 vs MAI Transcribe 2: 0.09 Points and a 4.4x Speed Gap

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Put ElevenLabs Scribe v3 and MAI-Transcribe-2 on one harness and they finish 0.09 percentage points apart on word error rate — 1.95% against 2.04% on Artificial Analysis' non-streaming AA-WER Index. Put them on one throughput bench and one of them finishes the same hour of audio in roughly a quarter of the wall-clock time. Those two facts point in opposite directions, and which one you follow depends on a question most transcription comparisons never ask: whether your bottleneck is the words or the clock. Scribe v3 is the more accurate model by a margin you will struggle to measure in production. MAI-Transcribe-2 is 4.4× faster and half the price. This page is about which of those you should actually pay for.

Start with the corpus that looks like your audio

The 0.09-point composite gap is an average, and the average hides a reversal. On Artificial Analysis' Earnings-22 corpus — spontaneous, accented, multi-speaker telephone audio, the closest thing on the board to a real sales call — MAI-Transcribe-2 scores 2.99% and Scribe v3 scores 3.34%. Microsoft wins that corpus by 0.35 points, four times the size of the composite gap, in the opposite direction. On the core read-speech set Scribe v3 wins 1.47% to 1.83%. On VoxPopuli, parliamentary speech, it is 1.52% to 1.50% — a dead heat inside the noise.

So "Scribe v3 is more accurate" is a statement about an average, and the average is only meaningful if your audio is an average of audiobooks, parliament and conference calls. Contact-centre and voice-agent workloads are not; they are Earnings-22-shaped, and on that corpus the ranking flips. Podcast and media-archive workloads are read-speech-shaped, and on those Scribe v3's advantage is real and larger than the composite suggests.

Two-column scoreboard card titled 'Scribe v3 vs MAI-Transcribe-2: the scoreboard', with six identical rows per column — AA-WER composite, AA-WER on earnings calls, median speed factor, price per 1,000 minutes, documented languages and verbatim mode — showing ElevenLabs Scribe v3 at 1.95%, 3.34%, 76.8x, $3.33, 90+ and none, and Microsoft MAI-Transcribe-2 at 2.04%, 2.99%, 338.5x, $1.67 introductory, 60 and yes, over a footer separating Artificial Analysis figures from vendor list prices.

Both sets of numbers are Artificial Analysis' measurements on a shared harness, which is the only reason the comparison is worth making at all. Neither vendor's own launch material is comparable to the other's: ElevenLabs reports a 27% relative FLEURS gain over Scribe v2 on 102 languages, Microsoft reports a 5.2% FLEURS average across 60 languages, and those two numbers are computed on different language sets with different averaging. The third-party board is the only place the two models meet on equal terms.

The 4.4× is the number that changes architecture

Throughput is where the two models stop looking like substitutes. MAI-Transcribe-2 runs at a median speed factor of 338.5 on Artificial Analysis' bench; Scribe v3 runs at 76.8. Both are far faster than real time — Scribe v3 clears an hour of audio in under a minute — so for a nightly batch job that backfills a podcast archive, the difference is invisible. For anything interactive it is not.

Screenshot of the Microsoft Learn page for MAI-Transcribe in Azure Speech, with body text describing MAI-Transcribe as a next-generation speech-to-text model built in-house by the Microsoft AI team that transcribes across 60 languages and serves video captioning, meetings, clinical notes, call-centre documentation, accessibility tools, content creation and voice agents, above the Foundry portal, REST API and SDK language tabs.

Consider a voice agent that needs a transcript before it can decide what to say next. At 76.8×, an hour of audio takes 47 seconds of compute. That is fine when the audio is a finished recording and useless when the audio is arriving live, which is why ElevenLabs sells a separate realtime model for that job rather than pointing at the batch one. Microsoft's answer is a separate streaming SKU too. Neither vendor's batch model is the right tool for live audio; the throughput number matters for a different case — high-volume asynchronous pipelines where you are processing thousands of hours a day and the queue, not the caller, is waiting.

At 10,000 hours a month, 76.8× needs about 130 hours of compute and 338.5× needs about 30. That is the difference between a cluster you rent by the hour and one you barely notice. It is also the difference between a backfill that finishes overnight and one that finishes over a weekend.

Price: a factor of two, with an expiry date

On list price the gap is cleaner than the accuracy gap. MAI-Transcribe-2 is $0.10 per audio hour, or $1.67 per 1,000 minutes. Scribe v3 is $0.20 per audio hour after ElevenLabs' 17 November cut, or $3.33 per 1,000 minutes. Exactly double, at both ends.

The Microsoft rate carries a qualifier that belongs in any budget built on it: $0.10 an hour is an introductory offer running through the end of 2026, and Microsoft has not published the standard rate that follows. A 2027 plan pinned to $1.67 per 1,000 minutes is a plan with an unpriced line in it. The ElevenLabs rate is the standing list price after a cut, not a promotion, which makes it the more defensible number for a multi-year model even though it is the higher one today.

Two other costs sit outside the headline rate and both favour ElevenLabs. Keyterm prompting on Scribe v3 bills at $0.05 an hour on top of the base rate; transcript editing at $0.07. Microsoft bundles keyword biasing into the transcription call. And ElevenLabs' per-hour rate covers diarization, entity detection and word timestamps in the same pass, while Microsoft meters through Azure Speech resources where the surrounding services have their own line items. The two-times multiple is real, but it is a floor on the gap, not a ceiling.

Features where they genuinely diverge

• Verbatim mode — MAI-Transcribe-2 has it, as a switch between verbatim and clean output styles. Scribe v3 does not; ElevenLabs documents "no verbatim mode" for the model. For compliance recording, QA scoring and anything where a false start is evidence, that is a hard requirement and it decides the choice on its own.

• Language breadth — Scribe v3 is documented at 90+ languages; MAI-Transcribe-2 at 60, with a per-language support table published and automatic language identification on by default. ElevenLabs lists language names but its accuracy breakdown is narrower than its coverage claim.

• Diarization — Scribe v3 up to 32 speakers, with the vendor claiming 12% fewer diarization errors than v2 and 17% fewer at three or more speakers. MAI-Transcribe-2 offers diarization as a parameter with speaker-labelled segments, and Microsoft publishes no diarization error rate at all. Unmeasured is not the same as bad, but it is unmeasured.

• Domain vocabulary — Scribe v3 takes up to 1,000 keyterms per batch request. MAI-Transcribe-2 takes a phrase list of arbitrary length as recognition hints, with the documentation explicit that terms are hints and not forced output.

• Deployment shape — MAI-Transcribe-2 runs through Azure Speech's enhanced-mode endpoint in named regions, with REST and SDK paths and a Voice Live integration for input transcription. Scribe v3 is a model ID change on an existing ElevenLabs integration, with US processing by default and EU, India and Singapore residency options on enterprise plans.

• Compliance — ElevenLabs requires a signed Business Associate Agreement through sales before HIPAA work, and its docs do not state that Scribe v3 itself is HIPAA-eligible, only that healthcare customers must sign a BAA first. Microsoft's Azure Speech stack carries the usual regional and enterprise compliance surface. Read both carefully before putting patient audio through either.

Screenshot of the ElevenLabs Speech to Text documentation model cards, showing Scribe v3 with 27% fewer word errors than Scribe v2 on FLEURS, diarization to 32 speakers, secondary language codes for mixed-language audio and keyterm prompting up to 1,000 terms, and Scribe v3 Live with a 10,000-term keyterm library, turn prediction and realtime timestamps.

Running both without a second contract

The realistic answer for most teams is not a winner but a split: one model for the corpus where it wins, routed by workload. That is the case OrcaRouter's routing layer exists for further up the pipeline — transcripts get summarised, classified and actioned, and those are the calls that multiply. One API across 200+ models, provider list prices passed through at 0% markup, automatic failover, and a routing DSL that can send a meeting transcript to one summariser and a compliance transcript to another without a second integration. Neither Scribe v3 nor MAI-Transcribe-2 is on our roster — both are served by their vendors — so the split still means two speech contracts. What collapses is everything downstream of the transcript.

If you do want a single vendor, the tiebreak is cheap to run: take a hundred hours of your own audio, score both on it, and look at the Earnings-22-shaped subset rather than the average. The composite gap of 0.09 points is smaller than the difference either model shows between corpora, which means the published leaderboard is a prior, not an answer.

Who should pick which

Pick MAI-Transcribe-2 if your audio is conversational, your volume is high, your pipeline is asynchronous, and you need verbatim output — and budget for the rate changing in January. Pick Scribe v3 if you need breadth of languages, large speaker counts, a published diarization claim, or a price you can put in a three-year model without an asterisk. Pick both, split by corpus, if you have the operational appetite — the accuracy difference between them is small enough that the decision is really about throughput, verbatim support and whose pricing you trust to still exist next year.