Microsoft launches new transcription model, calls it world's most accurate
What's the story
Microsoft has unveiled its latest innovation, the MAI-Transcribe-2-Streaming model. The tech giant's AI CEO Mustafa Suleyman has dubbed it as the most accurate real-time audio transcription model in the world. Along with this groundbreaking model, Microsoft also launched two companion speech generation models: MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The new models are aimed at providing developers with cost-effective and responsive tools to create autonomous conversational voice agents.
Model features
MAI-Transcribe-2-Streaming
The MAI-Transcribe-2-Streaming model offers continuous, low-latency live speech-to-text across 60 languages with automatic language detection.
It has been rated as the top model for accuracy on the independent benchmark platform Artificial Analysis.
Unlike other models that wait for a sentence to end before providing text, this one gives initial text hypotheses within around 100 milliseconds of receiving audio input.
Voice engine
MAI-Voice-2.1
To complement its speech recognition capabilities, Microsoft has also launched MAI-Voice-2.1, a multilingual text-to-speech engine supporting 23 languages and 26 regional locales.
The model's unique feature is cross-lingual identity preservation where a single synthetic voice can switch between different languages without carrying over accents from one language to another.
This means businesses can have consistent virtual ambassadors across the globe or customer support agents responding in callers' native tongues seamlessly.
Flash version
MAI-Voice-2.1-Flash
The company has also launched MAI-Voice-2.1-Flash, a high-speed version of the original model.
This one is designed for mission-critical applications that require instant response times.
It offers the same 23-language voice consistency as its predecessor but is optimized for high-throughput enterprise workloads, generating 45 seconds of synthesized audio with an end-to-end latency of just 150 milliseconds.