Microsoft’s new MAI voice models stream transcripts in just over 100 ms

Building on MAI-Transcribe-2, its $0.10-per-hour transcription model, Microsoft has launched a streaming version alongside two new text-to-speech models for voice agents. MAI-Transcribe-2-Streaming starts returning transcript fragments in just over 100 ms; MAI-Voice-2.1-Flash generates speech with 150 ms end-to-end latency. Source

Read the full article at the source.