Microsoft AI releases new transcription and text-to-speech models for voice agents
decoder ·
Microsoft AI has introduced new transcription and text-to-speech models designed to enhance voice agent capabilities. The company released MAI-Transcribe-2-Streaming, a real-time transcription model that supports 60 languages and delivers initial results in approximately 100 milliseconds. This low latency allows voice agents to respond while users are still speaking, with introductory pricing set at $0.54 per hour of audio. Additionally, Microsoft launched two new text-to-speech models, including the MAI-Voice-2.1 variant which can speak 23 languages with native accents using a single voice identity. The Flash version of this model achieves a latency of 150 milliseconds and offers a reduced cost of $15 per million characters compared to previous rates. Both voice models feature voice cloning capabilities from short audio samples and include safeguards against misuse. These tools are accessible via Microsoft Foundry, the MAI Playground, and OpenRouter, with testing indicating high human-like realism.
decoder