The headline numbers
The comparison uses Artificial Analysis' non-streaming AA-WER Index. The index is a duration-weighted average across about eight hours of audio from three datasets. The weights are AA-AgentTalk at 50%, VoxPopuli at 25% and Earnings22 at 25%. Lower is better.
| Metric | Scribe v3 | MAI-Transcribe-2 |
|---|---|---|
| AA-WER (OrcaRouter's precise snapshot) | 1.95% | 2.04% |
| AA-WER (leaderboard, rounded) | 1.9% | 2.0% |
| Median speed factor | 76.8 | 338.5 |
| Artificial Analysis price per 1,000 minutes | $3.33 | $1.67 |
The accuracy and speed figures are Artificial Analysis' leaderboard values. The two-decimal WER figures come from OrcaRouter's snapshot, because the live board shows rounded numbers and can change. Artificial Analysis notes that its speed medians use 10-minute audio over a seven-day window. It also says speed can vary for other durations, especially very short clips.
Why the average hides a reversal
OrcaRouter says the composite gap reverses on conversational audio. Its claimed corpus-level figures:
- Earnings-22 (accented, multi-speaker phone audio): MAI-Transcribe-2 at 2.99% against Scribe v3 at 3.34%.
- Core read-speech set: Scribe v3 at 1.47% against 1.83%.
- VoxPopuli (parliamentary speech): 1.52% against 1.50%, effectively a tie.
I could not confirm these per-dataset numbers on the Artificial Analysis pages I could inspect, so treat them as OrcaRouter's claims. The method behind them is sound either way. A composite blends audiobook-style, parliamentary and call-style audio. A contact centre or voice-agent workload may not look like that blend. The practical advice is to score both models on your own audio and look at the subset that resembles your calls.
What the speed gap means
Using the speed factors as simple ratios, an hour of audio takes about 47 seconds at 76.8× and about 11 seconds at 338.5×. At 10,000 audio hours a month, that is roughly 130 hours of processing against 30. These are illustrative calculations. Real queue times depend on file length, concurrency and service limits.
The gap matters most for high-volume asynchronous pipelines. For a one-off archive backfill it barely registers.
Neither batch figure says anything about live latency. MAI-Transcribe-2 accepts a completed audio file and returns a transcript when processing finishes. Microsoft sells a separate streaming model for live audio. ElevenLabs likewise documents a separate realtime product, Scribe v3 Live.
Pricing
Scribe v3 is cheaper to forecast. ElevenLabs' November 17 changelog set batch transcription at $0.20 per audio hour on self-serve and Business plans, with realtime at $0.22. Enterprise pricing is unchanged.
MAI-Transcribe-2's $0.10 per hour is the cheaper figure, but it is a promotion. Microsoft's launch post says it is priced at $0.10 per hour through December 31, 2026. Microsoft has not published what happens after that, according to eesel.ai's review.
Some signals suggest the post-promotion price could be higher:
- eesel.ai notes that MAI-Transcribe-1.5 is still listed at $0.36 per hour.
- Developers Digest read Azure's retail price API as showing $0.54 per hour for MAI-Transcribe-2 in East US. It also says the API doesn't state whether that figure reflects the promotion. Treat it as an unresolved discrepancy, not a confirmed post-promotion rate.
Any 2027 budget built on $0.10 is therefore a guess. The 2× price ratio could shrink or vanish.
OrcaRouter also says Scribe v3 adds $0.05 per hour for keyterm prompting and $0.07 for transcript editing, while Microsoft bundles keyword biasing. I could not verify those add-on fees, so check them against ElevenLabs' current pricing before relying on them.
Features that split the choice
- Verbatim output. Microsoft documents a transcription style that captures speech exactly as spoken, including filler words and false starts, for compliance, QA and analysis. A clean style is also available. ElevenLabs documents no verbatim mode for Scribe v3. If false starts count as evidence, that settles it.
- Languages. MAI-Transcribe-2 supports 60 languages. ElevenLabs documents 90+ for Scribe v3.
- Diarization. Scribe v3 supports up to 32 speakers. MAI-Transcribe-2 offers diarization as a request parameter that returns speaker-labelled segments. I found no published diarization error rate for Microsoft. OrcaRouter's claimed ElevenLabs error-reduction percentages are also not confirmed in the documentation I reviewed.
- Vocabulary biasing. Scribe v3 accepts up to 1,000 keyterms. Microsoft's phrase list biases recognition, but its documentation says terms are hints and not forced output.
- Integration. On Azure, you select the model through enhanced mode, with
enhancedMode.enabledset to true andenhancedMode.modelset to MAI-Transcribe-2, using a Speech resource in a supported region. ElevenLabs users switch by settingmodel_idtoscribe_v3.
The caveat the comparison underplays
Microsoft's own documentation flags MAI-Transcribe as a public preview. The streaming model's page says previews come without a service-level agreement and aren't recommended for production workloads. I saw that wording on the streaming page. The batch page I inspected did not carry it in the captured text, and the Microsoft launch post I found describes the batch model as being in public preview. For an enterprise IT team, a missing SLA matters as much as a 0.09-point WER difference. Confirm the current status on Microsoft Learn before committing production traffic.
Compliance deserves the same care. OrcaRouter says ElevenLabs requires a signed Business Associate Agreement before HIPAA work. Verify that with both vendors before routing patient audio through either service.
A sensible way to decide
- Collect a representative sample of your audio with human reference transcripts. OrcaRouter suggests about 100 hours, which is its advice and not an established standard.
- Split the sample by condition, such as call audio, clean speech and multi-speaker meetings.
- Score word error rate separately for each split, not just overall.
- Measure processing time at your real file lengths and concurrency.
- Check output-style fit: verbatim or clean, timestamps, and speaker labels.
- Compute the full billed cost, including add-ons and post-promotion pricing.
Bottom line
On the evidence available, MAI-Transcribe-2 is much faster and, for now, half the price. Scribe v3 is more accurate on the composite by a gap you will struggle to measure in production. It also offers wider language coverage, larger speaker counts and a price that is already settled. If you need verbatim output and run high-volume asynchronous jobs, Microsoft's model is attractive. Budget for its pricing changing and for its preview status. OrcaRouter's analysis is useful but comes from a vendor with a commercial stake in multi-model routing. Neither of these two models is on its roster, so test the claims on your own audio.
References
- Scribe v3 vs MAI-Transcribe-2: accuracy against throughput - OrcaRouter OrcaRouter · 2026-11-20T03:28:02.342000+00:00
- MAI-Transcribe-2 review: is Microsoft's speech-to-text model worth it? eesel.ai
- MAI-Transcribe-2: Highest quality transcription, at the fastest speed and lowest cost techcommunity.microsoft.com