Microsoft AI released MAI-Transcribe-2 on Thursday, a dedicated speech recognition model available through Microsoft Foundry and MAI Playground. The model is priced at 10 cents per audio hour for a limited period, undercutting the per-hour rates commonly charged for OpenAI's GPT-Transcribe, Google Gemini 3.5 Transcribe, and ElevenLabs Scribe v2. The launch extends Microsoft's push into internal AI development and gives enterprises a lower-cost option for high-volume transcription workloads such as call-center analytics, meeting recordings, and regulatory compliance.

Key facts at a glance

• Developer: Microsoft AI
• Availability: Public preview through Microsoft Foundry and MAI Playground
• Introductory price: $0.10 per audio hour until the end of 2026
• Languages: 60 languages with automatic language detection and code-switching
• Benchmark result: 2.0% word error rate in non-streaming independent tests
• Speed: Median speed factor of 410.7x real time in those tests
• Enterprise features: Speaker diarization, timestamps, clean and verbatim formatting, and keyword biasing
• File support: WAV, MP3, and FLAC up to 300 MB or two hours per file

What makes MAI-Transcribe-2 different

Unlike general-purpose multimodal AI models that can process text, images, and audio in one system, MAI-Transcribe-2 is built for one task: speech-to-text conversion. That narrower focus can deliver faster batch processing and better control over transcription workflows. The model's dedicated design also lets Microsoft tune it for domain-specific needs. Such specialization matters for industries that depend on precise audio records rather than generic AI outputs.

Microsoft has packed the release with several enterprise-oriented capabilities. Speaker diarization labels the person speaking each segment, while word-level timestamps make it easy to search or jump to a specific part of an audio file. The output can be adjusted to preserve disfluencies and filler words for legal or audit scenarios, or to generate readable meeting notes by returning a clean transcript. Built-in automatic language detection is joined by code-switching support for mixed-language conversations such as Hinglish and Spanglish. Domain adaptation is also included through keyword biasing, which lets organizations teach the model industry-specific acronyms, jargon, and proper names in advance.

Benchmark performance and independent testing

The new model arrives with strong benchmark claims. Microsoft says MAI-Transcribe-2 leads the public FLEURS multilingual benchmark across 60 languages with an average word error rate of 5.2%. In tests conducted by Artificial Analysis, the model ranked second overall for accuracy behind Alibaba's streaming Fun-Realtime-ASR preview. More importantly for enterprise teams, the non-streaming test results show a 2.0% word error rate and a median speed factor of 410.7 times real time. That means a one-hour audio file could theoretically be transcribed in less than nine seconds, assuming a comparable hardware environment.

Microsoft says the model ran roughly 10 times faster than OpenAI's GPT-Transcribe, 7.5 times faster than ElevenLabs Scribe v2, and 4.6 times faster than Google Gemini 3.5 Transcribe in the independent tests. Speed comparisons can shift as vendors update their software and test conditions, but the magnitude of the differences is notable for companies that process many thousands of hours each month.

Artificial Analysis normalizes pricing based on the cost of processing 1,000 minutes of audio. The figures are rolling median results and may change as new tests are completed. While composite averages can help compare models, actual accuracy varies by language, audio quality, and topic.

Language coverage is strong but uneven

MAI-Transcribe-2 shows elite performance in several major languages. It outclasses competitors in Chinese with a 4.5% word error rate and in French with a 2.8% word error rate. However, it is not uniformly the best transcription option. Google's Gemini 3.1 Pro beats it on Armenian, posting a 6.1% error rate versus the Microsoft model's 12.6%, and on Danish with a 5.3% error rate versus 5.4%. ElevenLabs Scribe v2 also takes the lead on Afrikaans, recording a 10.6% error rate compared with 13.1% for MAI-Transcribe-2.

These language-level differences matter for global organizations. A model that works superbly for English, Chinese, and French may be less suitable for a localized legal or customer service operation. Transcription tools need to be tested using realistic audio files, including terms and accents relevant to the company's industry. The 60-language promise is broader than many rivals offer, but broad support does not mean equal quality across every language.

Pricing math and what it means for buyers

The limited-time rate of $0.10 per audio hour is roughly 72% lower than the $0.36 per-hour price Microsoft lists for MAI-Transcribe-1 and MAI-Transcribe-1.5. For a contact center transcribing 100,000 hours of calls per month, the cost under that rate would be about $10,000 a month. At the older $0.36 price, the same workload could cost $36,000. Even at a smaller scale, the gap is meaningful. Rival offerings at roughly $0.22 to $0.30 per hour look more expensive than Microsoft's introductory rate, though their pricing is not temporary.

Cloud providers often use promotional pricing to attract early adopters. The current $0.10 rate is scheduled to expire at the end of 2026, and Microsoft has not disclosed what the model will cost afterward. That uncertainty complicates long-range budgeting for finance teams. If the post-promotional price increases to a level near or above competitors, the initial cost advantage may disappear. Procurement teams should model both the promotional year and a likely price increase scenario before committing to a transcription vendor.

Production trade-offs and enterprise risks

Even with compelling speed and price, enterprise engineering teams need to weigh the trade-offs. Microsoft Learn currently classifies MAI-Transcribe as a public preview without a formal service-level agreement. Public preview status can make a service unsuitable for mission-critical use because an SLA does not guarantee uptime or response times. The documentation also does not recommend using MAI-Transcribe for production workloads at this stage. The supported file types are WAV, MP3, and FLAC, with a maximum size of 300 MB or a maximum duration of two hours. Those limits are enough for many meetings but may be restrictive for long-form audio, such as multi-hour depositions or extended call-center session files.

The lack of production readiness heightens the need for a structured pilot. Organizations should run the model against representative recordings and measure output quality, handling of accents, punctuation, speaker changes, and noisy audio. Success in offline benchmarks does not automatically translate to clean transcripts of a live webinar or a crowded hospital room.

Microsoft's in-house AI expansion

MAI-Transcribe-2 is another signal that Microsoft is increasing its number of internally developed AI models. The company has made major investments in chatbots, foundation models, and assistant software. Speech recognition itself is a strategic area because it feeds so many user-facing workflows. Accurate transcription is central to intelligent meeting summaries, agent coaching, automated captioning, and voice assistants.

Microsoft owns Nuance, a company known for healthcare and contact-center conversational AI, and Teams has built-in transcription capabilities. Microsoft has not stated whether MAI-Transcribe-2 will replace the speech technology used in Teams, Nuance, or other products. It is possible the model will eventually be embedded across those services, but the company has not given a public roadmap. For now, it is available through Microsoft Foundry and MAI Playground, inviting developers to test it in their own pipelines.

The launch also reflects a broader trend in the AI market. Vendors are no longer shipping only general-purpose chatbots; they are releasing task-specific models for speech, image understanding, and code generation. Dedicated speech models can be optimized for latency and token usage more easily than massive multimodal systems. Microsoft is betting that price and speed, not just raw accuracy, will attract transcription-heavy organizations. That bet may prove particularly useful in call-center analytics, media localization, legal discovery, compliance monitoring, and market research, all of which require large volumes of accurate transcripts on a tight budget.

How to evaluate MAI-Transcribe-2

Early results suggest MAI-Transcribe-2 is worth evaluating for any team that needs multilingual batch transcription. The best way to validate it is to prepare a test set of recent audio files that reflect real production conditions. The set should include multiple languages, regional accents, background noise, overlapping speech, and domain vocabulary. Teams can compare the output of MAI-Transcribe-2 with their current provider, measuring word error rate by language, processing speed, and cost per hour. They should also check formatting behavior for timestamps and speaker labels, which can have a significant impact on downstream workflows. Where exact wording is essential, the verbatim mode may help; where readability is more important, the clean mode can reduce the effort spent editing raw transcripts.

Because MAI-Transcribe-2 is in public preview and its introductory price ends in 2026, the sensible path is a controlled proof of concept rather than immediate large-scale migration. The model has demonstrated that Microsoft can compete on accuracy, pace, and price in the crowded speech recognition market, but the details of reliability, data governance, and future pricing will decide whether it becomes a durable enterprise choice. For now, it is a fast-moving option worth watching and testing against the workloads that matter.

Source: eWeek News