Microsoft’s MAI Transcribe 2 is 10x faster than OpenAI’s
Speech to text just got cheaper. Microsoft’s MAI Transcribe 2 promises faster, more accurate transcription across 60 languages for just $0.10 an hour.
Transcription is getting faster and cheaper. Microsoft has launched MAI-Transcribe-2, its latest speech-recognition model, as competition heats up among AI companies building tools that can turn hours of audio into usable text.
According to Artificial Analysis, Microsoft’s MAI-Transcribe-2 outpaces Gemini 3.5 Transcribe by 5x, GPT-Transcribe by 10x and Scribe v2 by 7x in speed.
Faster transcription can help businesses get useful information sooner. Call centres can find customer records quickly, media teams can get transcripts faster, hospitals can finish documentation sooner, and accessibility tools can create captions more quickly.
More languages, more control
MAI-Transcribe-2 can transcribe audio in 60 languages, up from 43 in the previous model. Microsoft says it is designed for real-world recordings, including background noise, poor-quality audio and conversations where multiple people speak at the same time.
It also comes with features aimed at business use. Speaker diarisation identifies who said what, while word-level timestamps attach timing information to individual words. That can make editing captions or finding specific sections of a recording easier.
Keyword biasing lets users steer the model towards specialised terminology, which could be useful in fields such as healthcare, legal services and customer support. Automatic language identification can also detect the language being spoken without requiring users to select it manually.
The combination is important because enterprise transcription is rarely just about converting speech into plain text. Businesses often need to know who spoke, when they spoke and whether the system correctly understood industry-specific terms.
The price could be the bigger story
Microsoft is offering MAI-Transcribe-2 at $0.10 per hour of audio as a limited-time price through the end of 2026. That is significantly lower than the $0.36 per hour charged for the first model in the line, released just five months earlier.
For companies processing large volumes of audio, the savings can add up quickly. A business transcribing 100,000 hours of call-centre audio annually would pay about $36,000 at the earlier rate. At $0.10 per hour, the same workload would cost $10,000.
Accuracy is still central to the pitch. Microsoft says MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages, with an average word error rate of 5.2%. WER measures how often a transcription system gets words wrong, so a lower score generally indicates better accuracy.
The company also says it ranks second on Artificial Analysis' word-error-rate leaderboard and sits on the accuracy-latency Pareto frontier, meaning it aims to balance transcription quality with processing speed.
MAI-Transcribe-2 is available through Microsoft Foundry and MAI Playground.
For Indian startups, SaaS companies and large service businesses, the next test will be: how well does it handle local accents, mixed-language conversations and noisy recordings?
If Microsoft can deliver on those conditions, the model's combination of speed, accuracy and lower pricing could make high-volume AI transcription easier for businesses to adopt.