Microsoft's Voice You Can't Tell From a Human's

I have been thinking about speech recognition lately. I am considering voice agents, where low latency matters a lot for usability. Microsoft AI recently published its speech recognition model, called MAI-Transcribe-2-Streaming. It recognizes 60 languages and returns first results in just over 100 milliseconds. The company claims it leads on Artificial Analysis. Microsoft also introduced new speech synthesis models: MAI-Voice-2.1 and a faster variant called Flash. These support 23 languages with one voice each, and each voice has a natural accent for its language. The models can clone any voice in a few seconds from a short audio sample. Almost half of the 4,000 people surveyed thought the synthetic voice was human. Flash costs just $15 per million characters, compared to $22 for the standard version.

At such low latency, around 100 to 150 milliseconds, you can make a voice assistant either seem like it is listening or seem like it is interrupting. That is exactly what Microsoft wants to achieve with this product: letting agents respond while the user is still speaking. Nobody could tell the synthetic voice apart from a human one in half the cases. For cold calls or an answering service, that is good enough. The price gap between Flash and the standard version adds up once your agent handles thousands of conversations. You can plug this into an existing setup through OpenRouter or the MAI Playground, but that means another paid vendor key. I would test it on a small batch of conversations first, before putting it into production.