Microsoft has introduced three new models for real-time voice interaction: the MAI-Transcribe-2-Streaming speech-to-text model, along with the MAI-Voice-2.1 and MAI-Voice-2.1-Flash text-to-speech models. These tools are designed to enable developers to build voice assistants and conversational agents capable of processing audio streams and responding with natural speech in real time.
The MAI-Transcribe-2-Streaming model supports 60 languages and features automatic language detection. According to Artificial Analysis, it currently ranks first in streaming transcription benchmarks, achieving a word error rate of approximately 2.5% and delivering final text about 0.13 seconds after speech concludes. The model is priced at $0.54 per hour of audio, with a promotion valid until the end of 2026.
The MAI-Voice-2.1 model supports 23 languages and 26 regional variants, priced at $22 per million characters. For low-latency needs, the MAI-Voice-2.1-Flash version offers a 55% increase in inference speed and is priced at $15 per million characters. Both models support cross-lingual voice cloning, provided the user has explicit authorization and consent from the voice owner.
The models are currently available via Microsoft Foundry, MAI Playground, and Vercel. Additionally, the voice models are integrated into OpenRouter and accessible through Azure Voice Live. Microsoft has designated these releases as public previews, noting that they currently lack service-level agreements and are not recommended for production workloads at this stage.
