What is MAI-Transcribe-2-Streaming?

MAI-Transcribe-2-Streaming is Microsoft AI's first in-house streaming speech-to-text model, released October 1, 2026. It transcribes in real time across 60 languages with automatic, continuous language detection, returns its first partial hypothesis roughly 100 ms after receiving audio, and debuted ranked #1 for accuracy on both final and partial transcripts on the Artificial Analysis streaming leaderboard. Microsoft prices it at an introductory $0.54 per hour of audio through the end of 2026 and pairs it with the MAI-Voice-2.1 text-to-speech family for voice agents.

Is MAI-Transcribe-2-Streaming more accurate than Grok Voice Transcribe 2.0?

On the Artificial Analysis streaming leaderboard, yes as of October 2, 2026: Microsoft says MAI-Transcribe-2-Streaming ranks #1 for accuracy on both final and partial transcripts, which displaces Grok Voice Transcribe 2.0, the model that held #1 among 32 streaming models when it launched September 18, 2026. The margin is not published, and leaderboard order changes often. Both vendors report telephony and short-phrase gains; test on your own audio before switching.

Which streaming speech-to-text API is cheapest in 2026?

Per hour of streamed audio, list prices as of October 2, 2026: Grok Voice Transcribe 2.0 $0.20, Deepgram Nova-3 $0.29 at its promotional rate ($0.46 regular), ElevenLabs Scribe v2 Realtime $0.39, AssemblyAI Universal-3.5 Pro $0.45, and MAI-Transcribe-2-Streaming $0.54 introductory through December 31, 2026. Grok is 2.7x cheaper than Microsoft; Microsoft is selling latency and accuracy, not price.

Which should I pick for a voice agent?

If first-word latency decides your agent's feel, MAI-Transcribe-2-Streaming's ~100 ms partials and Pareto-frontier accuracy-versus-latency make it the current pick, and Microsoft's own MAI-Voice-2.1-Flash closes the loop at 150 ms end-to-end. If per-hour cost dominates, Grok Voice Transcribe 2.0 at $0.20/hr with diarization included is hard to beat. Deepgram Nova-3 remains the safe incumbent with on-prem options; ElevenLabs Scribe v2 wins when you need 90+ languages and 48-speaker diarization in one vendor alongside TTS.

Quick Answer

MAI-Transcribe-2-Streaming vs Grok Transcribe 2 vs Deepgram

Published: October 2, 2026

The short answer

Microsoft’s MAI-Transcribe-2-Streaming, released October 1, 2026, is the new accuracy leader for real-time speech-to-text — #1 on Artificial Analysis for both final and partial transcripts, with first partials in about 100 ms — but at $0.54 per hour it costs 2.7x Grok Voice Transcribe 2.0’s $0.20. Pick Microsoft for latency-sensitive agents, Grok for cost, Deepgram for a mature production stack, Scribe v2 for language breadth. Facts verified October 2, 2026.

Side by side

	MAI-Transcribe-2-Streaming	Grok Voice Transcribe 2.0	Deepgram Nova-3	ElevenLabs Scribe v2 Realtime
Maker	Microsoft AI	SpaceXAI (formerly xAI)	Deepgram	ElevenLabs
Released	Oct 1, 2026	Sep 18, 2026	2025, updated 2026	2026
Streaming price / hour	$0.54 (intro to Dec 31, 2026)	$0.20	$0.29 promo / $0.46 regular	$0.39 ($0.28 annual Business)
Batch price / hour	Streaming-only model	$0.10	$0.26	$0.22 (Scribe v2 batch)
Languages	60, continuous auto-detect	Dozens, mid-recording switching	Multilingual	90+
First partial	~100 ms	Not published	Sub-300 ms (vendor)	~150 ms
Diarization	Not stated at launch	Included	Included	Included, up to 48 speakers
AA streaming rank	#1 accuracy (final + partial)	#1 at its Sep 18 launch; now displaced	Mid-pack	~2.3% AA-WER (non-streaming index)
Open weights	No	No	No	No
Paired TTS	MAI-Voice-2.1 / 2.1-Flash	Grok Voice	Aura-2	Eleven v4 / v4 Turbo
Best for	Lowest-latency agents, live captions	Cheapest production audio	Incumbent stacks, on-prem	Multilingual, one vendor for STT+TTS

Prices are pay-as-you-go list rates read from vendor pages on October 2, 2026; volume and annual tiers are lower.

MAI-Transcribe-2-Streaming — what Microsoft shipped

Microsoft AI released MAI-Transcribe-2-Streaming on October 1, 2026, alongside two text-to-speech models (MAI-Voice-2.1 and MAI-Voice-2.1-Flash). Four claims matter:

#1 on Artificial Analysis for streaming accuracy, on both final transcripts and partial (in-progress) hypotheses, and on the Pareto frontier of AA’s accuracy-versus-latency plot — meaning no listed model is both more accurate and faster.
~100 ms to first partial. The model emits a hypothesis just over 100 ms after receiving audio, revises as context arrives, and commits a stable transcript immediately. Microsoft’s internal tests say words appear 2x faster than its closest competitor for dictation and subtitling.
60 languages with continuous automatic detection, so a caller can switch languages mid-sentence without a config change.
$0.54 per hour of audio, introductory through the end of 2026. Microsoft has not published the post-promo rate; budget higher.

The design target is explicit: a voice agent is a hear-understand-decide-speak loop, and partials let the agent start reasoning or calling tools before the speaker finishes. Microsoft’s Chatter demo in the MAI Playground pairs it with MAI-Voice-2.1-Flash (150 ms end-to-end for 45 s of audio, $15 per 1M characters). This is Microsoft’s second-generation in-house STT after MAI-Transcribe-1, and part of the MAI self-sufficiency push away from OpenAI models.

Grok Voice Transcribe 2.0 — the price floor

SpaceXAI’s Grok Voice Transcribe 2.0 (September 18, 2026) held the #1 streaming-accuracy slot for 13 days. It remains the cheapest serious option: $0.20/hr streaming, $0.10/hr batch, with diarization, word timestamps with confidence, up to 100 key terms, 8-channel audio and smart turn detection included rather than billed as add-ons. SpaceXAI’s internal sets showed telephony first-final WER at 2.7% and short-phrase multilingual WER at 6.8%. Endpoint api.x.ai/v1/stt; API-only, no open weights. Full breakdown in Grok Transcribe 2.0 vs Deepgram vs AssemblyAI vs Scribe v2.

Deepgram Nova-3 — the incumbent

Most 2025-era voice-agent stacks were built on Nova-3. Streaming is $0.0048/min (~$0.29/hr) at a promotional rate against a regular $0.0077/min (~~$0.46/hr) with no published end date — model your budget at the regular rate. Batch is $0.0043/min (~~$0.26/hr); new accounts get $200 credit. Deepgram’s advantages are maturity, keyword boosting, custom vocabularies and on-prem deployment for enterprise. Its disadvantage is that two newer models now beat it on both accuracy and price.

ElevenLabs Scribe v2 Realtime — breadth

Scribe v2 Realtime is $0.39/hr ($0.28 on annual Business plans) with ~150 ms latency, 90+ languages, diarization for up to 48 speakers and non-speech event tags; keyterm prompting (+$0.05/hr) and entity detection (+$0.07/hr) are add-ons. Its case is consolidation: ElevenLabs shipped Eleven v4 and v4 Turbo TTS on September 28, 2026, so one vendor covers both ends of a voice agent. See MAI-Voice-2.1 vs Eleven v4 vs Cartesia Sonic 3.6 for the TTS side.

Also in the market

AssemblyAI Universal-3.5 Pro — $0.45/hr realtime, $0.21/hr async, billed per second; diarization +$0.12/hr realtime; strong on entity accuracy and code-switching across 18 languages; Voice Agent API bundles STT+LLM+TTS at $4.50/hr.
Meta Muse Voice Transcribe — $0.18/hr streaming and batch (September 1, 2026), zero-data-retention option at the same price.
OpenAI GPT Transcribe — $0.0045/min ($0.27/hr); legacy Whisper API $0.006/min ($0.36/hr). Whisper large-v3 remains the self-hosting default because none of the models above have open weights.

How to choose

Agent where a 300 ms pause feels like a bug: MAI-Transcribe-2-Streaming, paired with MAI-Voice-2.1-Flash or Eleven v4 Turbo.
High-volume call transcription where cost dominates: Grok Voice Transcribe 2.0 — 2.7x cheaper than Microsoft and diarization is free.
Existing Deepgram integration that works: stay on Nova-3, but renegotiate — your vendor is now mid-pack on price and accuracy.
Need 90+ languages or one vendor for STT and TTS: ElevenLabs Scribe v2.
Audio cannot leave your servers: self-hosted Whisper large-v3; none of these four qualify.

A warning on leaderboards: Grok held #1 for 13 days. Artificial Analysis rankings reorder with each release; the durable facts are latency architecture and price, both in the table above.

Last verified: October 2, 2026. Microsoft figures from the October 1 announcement; Grok, Deepgram, ElevenLabs and AssemblyAI prices from vendor pricing pages.