October 3, 2026 ·

microsofttranscriptionspeech-to-texttext-to-speechvoice-agentspricingintroductory-pricingaudioprocurementcost-modelling

Microsoft bought the transcription accuracy crown at 3x the price floor — and the price expires in 89 days

TL;DR: On 1 October 2026 Microsoft AI shipped MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, claiming the no. 1 accuracy rank on Artificial Analysis for both final and partial transcripts — reportedly a 2.5% word error rate, first hypotheses in just over 100ms, 60 languages with automatic detection. It costs $0.54 per audio-hour, which is 3x Meta’s $0.18 and roughly 34x Alibaba’s China-hosted rate. The accuracy that buys you, against the cheapest streaming competitor, is about six fewer wrong words per thousand. And $0.54 is an introductory rate “through the end of the year” — 89 days from publication — with no successor rate published. The two text-to-speech models launched the same day are the more actionable news: MAI-Voice-2.1 at $22 per million characters and Flash at $15 — not the cheapest Western voice today (Gemini 3.8 Flash-Lite TTS and ElevenLabs’ discounted v4 Turbo both undercut it), but on a rate card with no published increase while Google’s doubles on 1 January 2027. All three are public preview with no SLA.

Correction (3 October 2026): this article originally priced ElevenLabs Eleven v3 at ”~$100 per million characters” and called MAI-Voice-2.1-Flash “the cheapest credible Western voice on the board.” Both were wrong. ElevenLabs’ API list rate is $80 per million characters for Eleven v4 and $40 for v4 Turbo, discounted to $22 and $11 until 12 October 2026, and Gemini 3.8 Flash-Lite TTS is cheaper than MAI-Voice-2.1-Flash per hour of generated audio. The text-to-speech section and FAQ have been rewritten; see MAI-Voice 2.1 vs ElevenLabs and the corrections log.

The accuracy crown is now the most expensive seat in the category

For eighteen months the direction of travel in speech-to-text was one-way: better and cheaper, together. Whisper set a quality bar and then specialists undercut it. Google’s Gemini 3.5 Transcribe arrived in August 2026 at about $0.30 per audio-hour, meeting the Deepgram and AssemblyAI floor rather than breaking it. On 1 September Meta broke it, shipping Muse Voice Transcribe at $0.18 per audio-hour and taking the top slot on Artificial Analysis’ streaming accuracy benchmark at 3.1% word error rate. Then on 22 September Alibaba’s Qwen-Audio-3.1-ASR priced out near $0.016 per audio-hour — an order of magnitude below everyone.

Microsoft’s launch breaks that pattern, and it is the first launch in the category this year to do so. MAI-Transcribe-2-Streaming takes the accuracy crown — and charges triple the price floor for it.

Streaming speech-to-text	Per audio-hour	Reported streaming WER
Qwen-Audio-3.1-ASR (Beijing node)	~$0.016	not in this benchmark
Muse Voice Transcribe (Meta)	$0.18	3.1%
Deepgram / AssemblyAI	$0.23–0.28	—
Gemini 3.5 Transcribe Live	~$0.30	4.0%
ElevenLabs Scribe v2 Real-time	—	3.6%
MAI-Transcribe-2-Streaming	$0.54 (intro, to 31 Dec 2026)	2.5%

Microsoft’s own post claims the rank rather than the number; the 2.5% figure and the 0.13-second final-transcript latency come from secondary reporting, and independent write-ups put the benchmark field at 38 models. Take the rank as sourced and the decimal as approximate.

Six words per thousand, priced at 200%

The honest way to read a 0.6-point WER gap is to convert it into something a buyer can picture. At 2.5%, a thousand transcribed words contain 25 errors. At 3.1%, they contain 31. Microsoft’s model is better by six words per thousand, and costs three times as much.

Scale that. A mid-size contact centre running 10,000 audio-hours a month pays $5,400 on Microsoft against $1,800 on Meta — $43,200 a year for those six words. A podcast network doing 200 hours a month pays $108 against $36, which is a rounding error and not worth optimising either way.

So the purchase splits cleanly by whether an individual error is expensive:

Worth it: medical and legal transcription, where one missed negation is a liability event; live captioning and broadcast, where partial-transcript accuracy is literally what the viewer reads, and Microsoft claims the lead on partials specifically; real-time voice agents where a misheard account number ends the call.
Not worth it: meeting notes, support-call summarisation, podcast transcripts, search indexing — anything where a language model downstream reads the transcript and silently repairs most of the errors anyway.

Most buyers are in the second group, which is why the accuracy crown is the least interesting thing about this launch.

The price has an expiry date and no successor

Microsoft’s wording is specific: $0.54 per hour of audio “through the end of the year.” That is 89 days from publication, and Microsoft has published no rate for 1 January 2027 — no multiplier, no commitment, no “list price” for the introductory rate to revert to.

This is now a pattern across the category rather than one vendor’s quirk, and it is the single most under-priced risk in AI audio procurement. Google documents its Gemini 3.8 Flash TTS and Flash-Lite TTS rates as introductory through 31 December 2026, doubling on 1 January 2027 — at least Google published the second number. Microsoft has published the first number only. If Microsoft follows Google’s pattern the real rate is $1.08 per audio-hour, which would be six times Meta’s current price. If it holds $0.54, the model is merely expensive rather than uncompetitive.

The practical consequence: any 2027 budget built on $0.54 is built on a number with a published expiry and no published replacement. Price the pilot at $0.54 and the production year at $1.08, and if the gap between those two scenarios changes your decision, you do not have enough information to sign yet.

The voice models are the part worth acting on

Buried under the transcription headline, Microsoft shipped two text-to-speech models — and the structural point about them survives even though the “cheapest Western voice” framing does not.

Text-to-speech	Per 1M characters	Note
Qwen-Audio-3.1-TTS	~$3.30	China-hosted node; no provenance marking
MAI-Voice-2.1-Flash	$15	150ms end-to-end; 23 languages, 26 locales
MAI-Voice-2.1	$22	23 languages, 26 locales
Gemini 3.8 Flash-Lite TTS	~$22	doubles to ~$44 on 1 Jan 2027
Gemini 3.8 Flash TTS	~$33	doubles to ~$66 on 1 Jan 2027
ElevenLabs Eleven v4 Turbo	$11 (launch discount to 12 Oct; $40 list)	no. 1 Speech Arena rank; EEA/UK cloning
ElevenLabs Eleven v4	$22 (launch discount to 12 Oct; $80 list)	voice library, licensing, documented provenance

Read the right-hand column before the left. Two of the six rates above carry published end dates and Microsoft’s two carry none — and that, not the headline number, is the story.

Priced in hours of generated audio, which is the only unit all three vendors convert to cleanly, MAI-Voice-2.1-Flash costs $0.81 per hour — exactly level with Gemini 3.8 Flash TTS and above Gemini 3.8 Flash-Lite TTS at $0.54. Until 12 October, ElevenLabs’ v4 Turbo launch discount sits below both at $0.59. So Microsoft is not the cheap option today on any of the three comparisons. It becomes the cheapest Western rate on 1 January 2027, when Google’s introductory rates double to $1.08 and $1.62 per hour and ElevenLabs’ discount is long gone. An unstated price is not a fixed price, and Microsoft could reprice voice the same week it reprices transcription — but one vendor has published the date its rates rise and the other has not, and that asymmetry is what you are actually buying. Full dated arithmetic and a decision table: MAI-Voice 2.1 vs ElevenLabs.

What this does not change: ElevenLabs remains the pick for narration, audiobooks and production voiceover, because the purchase there is the voice library, the commercial licensing, the cloning tooling and the compliance story — not the per-character rate. Microsoft is competing for the API-volume buyer generating speech at scale, and against Google specifically.

The constraint that overrides all of the above

All three models are public preview, with no service-level agreement, and Microsoft Foundry’s own listings describe them as not recommended for production. Microsoft’s launch post names neither an SLA nor a GA date.

For a batch workload that is tolerable — if the model changes, re-run the job. For streaming transcription sitting in a live customer-call path, it is not: a preview service can change behaviour, change price or withdraw with no contractual remedy, and you find out in production. Availability is wide (Microsoft Foundry, MAI Playground, Vercel, OpenRouter, Azure Voice Live, with LiveKit coming), which makes it easy to pilot and easy to over-commit to.

Pilot it. Do not route revenue through it until the terms exist.

What this means if you are buying audio AI this quarter

Transcription: default to Meta’s $0.18 unless an individual error is expensive. Microsoft is the pick for medical, legal, broadcast captioning and live voice agents — and only once it leaves preview.
Text-to-speech at API volume: the cheapest provenance-marked rate today is Gemini 3.8 Flash-Lite TTS at $0.54 per hour of audio, with ElevenLabs’ v4 Turbo discount at $0.59 until 12 October. MAI-Voice-2.1-Flash at $0.81 wins from 1 January 2027, when Google’s rates double. Verify provenance marking first.
Text-to-speech as a product: ElevenLabs still wins on library, licensing, EEA/UK cloning and documented provenance. The price gap is the cost of the compliance story.
Meeting tools: nothing changes. Otter, Granola and Fireflies compete on integration, not accuracy, and you cannot choose their upstream model anyway.
Every introductory rate in this category now expires within 90 days. Build the 2027 scenario at double the quoted price and see whether the decision survives.

Frequently asked questions

What does MAI-Transcribe-2-Streaming actually cost, and for how long?

Microsoft's launch post prices it at $0.54 per hour of audio 'through the end of the year' — so the rate is contractually good until 31 December 2026 and no further. That works out to $9.00 per 1,000 minutes, or $0.009 per minute. Microsoft has published no successor rate, no multiplier and no commitment about what happens on 1 January 2027. Google set the precedent for how this usually resolves: its Gemini 3.8 TTS introductory rates are documented on the same pricing page as doubling on 1 January 2027. If Microsoft follows the same pattern, the real rate is $1.08 per audio-hour, which would be six times Meta's current price. Nobody outside Microsoft knows, and that is the point — if you are modelling a 2027 budget on $0.54, you are modelling on a number with a published expiry date and no published replacement.

Is the accuracy gain over cheaper models worth 3x the price?

For most buyers, no. Microsoft claims the no. 1 accuracy rank on Artificial Analysis for both final and partial transcripts, and secondary reporting puts the word error rate at 2.5% against Meta Muse Voice Transcribe's 3.1%. Convert that to something you can feel: 2.5% is 25 wrong words per 1,000, 3.1% is 31. You are paying 200% more for roughly six words per thousand. At 10,000 audio-hours a month — a mid-size contact centre — that is $5,400 versus $1,800, a difference of $43,200 a year. The cases where it does pay are narrow and real: medical and legal transcription where a single missed negation is a liability event, and live captioning where partial-transcript accuracy is what the viewer actually reads. For meeting notes, support-call summarisation or podcast transcripts, the cheaper model's errors are corrected by the downstream language model anyway.

Can I use these models in production today?

Microsoft's own launch post names no service-level agreement and no general-availability date, and secondary reporting of the Microsoft Foundry listings describes all three models as public preview, without an SLA, and not recommended for production. Treat that as the binding constraint rather than the pricing. A preview model can change behaviour, change price or disappear with no contractual remedy, and for a streaming transcription service sitting in a live call path that is a materially different risk from a batch job you can re-run. Build against it, pilot on it, do not route customer calls through it until Microsoft commits to terms.

What about the two text-to-speech models — are those the better story?

They are the part of this launch most likely to change a purchase, though not for the reason first reported. MAI-Voice-2.1 lists at $22 per million characters and MAI-Voice-2.1-Flash at $15, with Flash claiming 150ms end-to-end latency, across 23 languages and 26 locales. Flash is not the cheapest credible non-Chinese voice model: priced per hour of generated audio it costs $0.81, level with Gemini 3.8 Flash TTS and above Gemini 3.8 Flash-Lite TTS at $0.54, and ElevenLabs' Eleven v4 Turbo is below both at $0.59 until its launch discount expires on 12 October 2026. ElevenLabs' API list rates are $80 per million characters for Eleven v4 and $40 for v4 Turbo — not the $100 figure widely quoted for the older Eleven v3. The structural point does hold: Google's TTS rates are documented as introductory and double on 1 January 2027, to $1.08 and $1.62 per hour, while Microsoft's voice prices carry no published expiry — so from January, MAI-Voice-2.1-Flash is the cheapest Western rate on the board. That is not a guarantee, because an unstated price is not a fixed price, but the two vendors' rate cards are heading in opposite directions on a date Google has already published.

Does this change the recommendation for Otter, Granola or Fireflies?

No, and the reason is worth understanding because it will keep being true. Meeting tools compete on calendar integration, meeting history, search and CRM sync, not on raw transcription accuracy, and none of them expose which upstream model they use as a purchasing choice. Cheaper and more accurate speech-to-text reaches you as gradually better transcripts inside the product you already pay for, not as a reason to switch products. The upstream price collapse — from roughly $0.30 per audio-hour in August 2026 to $0.18 in September to about $0.016 on Alibaba's China-hosted node — matters to the meeting-tool vendors' margins, and eventually to their prices. It does not currently matter to your shortlist.

If I publish synthetic voice into the EU, is there a compliance catch?

Check provenance marking before you switch, because it is not uniform across this category and Microsoft's launch post foregrounds accuracy and latency rather than marking. Google marks Lyria audio output with SynthID at the model, which is the strongest position among the major vendors, and ElevenLabs ships documented provenance for published output. Alibaba's Qwen-Audio-3.1-TTS, which is cheaper than everything else here by an order of magnitude, ships no provenance marking at all. Under the EU AI Act's Article 50 transparency duties, synthetic audio you publish needs to be disclosed as such, and a machine-readable mark applied at generation is the cheapest way to satisfy that durably. Confirm what each vendor marks before you build a publishing pipeline on it.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.