Microsoft AI made a significant move in the competitive landscape of speech-recognition technology this Thursday with the launch of MAI-Transcribe-2. The company claims the new model is faster, more accurate, and more cost-effective than any competing offerings from industry titans like OpenAI, Google, or the specialty firm ElevenLabs. Perhaps most notably, Microsoft has set the pricing for this new capability at a mere 10 cents per hour of audio, a radical shift that signals an aggressive push to commoditize a critical piece of enterprise AI infrastructure.
This price point represents a staggering 72% reduction from the $0.36 per hour rate Microsoft charged for its first speech-recognition model just five months ago. To understand the impact on corporate balance sheets, consider an enterprise, such as a major bank or telecommunications provider, that processes 100,000 hours of call-center audio annually. Under the previous pricing structure, this volume would incur a cost of $36,000. With the introduction of MAI-Transcribe-2, that same workload would cost $10,000. At this price, transcription transitions from being a complex, debated line item in an IT budget to a standard, nearly invisible utility.
The release of MAI-Transcribe-2 is the latest step in a broader, ambitious strategy that would have seemed implausible to industry observers two years ago: Microsoft is systematically building its own frontier-class models for specific modalities and integrating them into its own product ecosystem, effectively replacing the OpenAI technology that once powered those features. Transcription has served as the vanguard for this initiative, and the new model stands as its most robust proof point. Furthermore, it offers a window into how the world’s most valuable software company plans to assert its independence and compete in the AI market without relying exclusively on the partner it has bolstered with a $13 billion investment.
Feature Sets and Enterprise Utility
The utility of MAI-Transcribe-2 for business users extends well beyond simple transcription. The model now supports 60 languages, a substantial increase from the 43 supported by the MAI-Transcribe-1.5 release in June and the 25 languages available in the original April launch. The model is accessible through Microsoft Foundry, the company’s dedicated model marketplace for developers, and via the MAI Playground testing environment. Microsoft emphasizes that the model was specifically engineered to handle the "messy" audio typical of real-world business environments—such as background noise, low-quality recordings, and overlapping speech—rather than relying on clean, studio-recorded audio.
Beyond sheer language capacity, Microsoft has bundled a suite of enterprise-grade features into the base product. Speaker diarization, which identifies and separates different speakers in a multi-person recording, is included to transform raw text into usable meeting minutes. The model also provides word-level timestamps, enabling developers to build sophisticated search, editing, and video-alignment tools. To address the problem of technical jargon—such as drug names in clinical settings, product codes, or specific employee names—Microsoft has included keyword biasing. Additionally, automatic language identification eliminates the need for users to specify the input language before processing begins.
Two specific features highlight the model’s intended use cases: a configurable output style and robust code-switching capabilities. The "verbatim" mode captures every filler word, stutter, and false start, a necessity for legal and compliance departments, while the "clean" mode strips these fillers to produce polished captions and meeting notes. The code-switching feature is designed to handle conversations that shift between languages mid-sentence—a common occurrence in bilingual markets like the U.S. Hispanic or Indian regions. Historically, specialty vendors have demanded significant premiums for these advanced capabilities, but Microsoft has chosen to include them all for the base price of a dime.
Benchmarking Performance and Technical Reality
Microsoft’s performance claims rest on three primary metrics, each providing a different perspective on the model’s capabilities. First, the company claims that MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages, achieving an average word error rate (WER) of 5.2%. FLEURS, a benchmark established by Google researchers in 2022, is widely considered the standard for multilingual speech recognition because it uses identical content across different languages. While a 5.2% WER is impressive, it is worth noting that Microsoft’s average has technically risen from the 3.7% reported for its previous model. Analysts suggest this is likely due to the inclusion of more challenging, low-resource languages in the expanded 60-language set, rather than a regression in model quality.
The second claim positions the model second on the Artificial Analysis word-error-rate leaderboard, identifying it as occupying the "accuracy-latency Pareto frontier." This designation implies that no competing model can beat MAI-Transcribe-2 on accuracy without sacrificing speed, nor can it provide greater speed without a corresponding decline in accuracy. Artificial Analysis benchmarks reflect real-world performance by testing models through public APIs, using a mix of simulated conversations, corporate earnings calls, and parliamentary speeches.
Finally, Microsoft highlights raw speed, claiming the model is 10 times faster than OpenAI’s GPT-Transcribe, seven times faster than ElevenLabs’ Scribe v2, and five times faster than Google’s Gemini 3.5 Transcribe. In the context of batch transcription, this speed translates directly into throughput efficiency. By requiring fewer GPU-hours to process a given volume of audio, Microsoft is able to maintain profitability at the lower 10-cent price point.
A Rapid Development Cadence
The speed of Microsoft’s release cycle is perhaps the most notable aspect of its current strategy. In just five months, the company has launched three iterations: the original model in April, MAI-Transcribe-1.5 in June, and the new MAI-Transcribe-2 today. This rhythm—expanding language support by approximately 40% per release while adding premium features—suggests that Microsoft has successfully stabilized its architecture and is now aggressively scaling through data and optimization.
This operational efficiency is largely attributed to the leadership of Mustafa Suleyman, CEO of Microsoft AI. Suleyman has previously described the effort as the work of a "small, focused 10-person team" operating outside the standard bureaucratic constraints of a large corporation, supported by a larger infrastructure team. This "flattened" structure is an experiment in whether such an approach can yield commercial products rather than just academic research papers.
The Strategic Pivot from OpenAI
The existence of the MAI model line raises an inevitable question: why is Microsoft building in-house alternatives to the models produced by its multi-billion dollar partner, OpenAI? The answer is rooted in a combination of strategic independence and margin optimization. As Microsoft restructured its partnership with OpenAI—eventually moving to a model that allows for independent development—it gained the freedom to pursue its own AGI objectives.
Furthermore, every request routed through an OpenAI or Anthropic model incurs a cost, whereas routing requests through proprietary, in-house models significantly improves Microsoft’s margins. With vast amounts of audio generated by Teams, Nuance’s clinical documentation services, and Azure speech services, Microsoft possesses the perfect "in-house" testing ground. By migrating these massive internal workloads onto MAI-Transcribe-2, the company stops writing checks to its partners and starts capitalizing on its own infrastructure.
For technical decision-makers, however, the current launch leaves some questions unanswered. The duration of the $0.10 "launch offer" remains unspecified, and the model’s performance in real-time, streaming scenarios—a requirement for live voice agents—has not been detailed. Additionally, while the average WER is low, potential users should look for per-language performance data and specific details regarding how audio data is handled, particularly concerning residency and retention for regulated industries.
As it stands, the release of MAI-Transcribe-2 is a template for how Microsoft intends to operate moving forward. By building specialized models that compete on price and performance, and by distributing them through the Azure ecosystem, Microsoft is effectively building a "factory" for AI capabilities. For the company, the message is clear: it no longer wants to be a renter of frontier intelligence; it intends to be the primary provider, setting the pace for the entire industry.
