Microsoft has released three models for building real-time voice agents. MAI-Transcribe-2-Streaming transcribes 60 languages in real time and produces its first text about 100 ms after receiving speech, MAI-Voice-2.1 speaks 23 languages in the same voice, and MAI-Voice-2.1-Flash generates 45 seconds of audio with about 150 ms of latency. They are built so an AI can start understanding and acting before a person finishes speaking, linking listening, deciding, and speaking without a break.
Microsoft AI announced the three models at 01:15 KST on October 2, with an 18-second intro video.
Source: X @MicrosoftAI
MAI-Transcribe-2-Streaming: transcription that starts before the speech ends
MAI-Transcribe-2-Streaming transcribes 60 languages in real time and keeps detecting the language automatically even if it changes mid-speech. It doesn't wait for the speaker to finish: about 100 ms after receiving audio it emits a first estimated text (a partial), refines it as more context arrives, and finalizes stable transcripts right away. A voice agent can start reasoning or calling tools in the middle of a sentence, and live captions appear as the person speaks. In Microsoft's internal evaluation of live transcription and captioning, words appear twice as fast as the closest competing model.
In Artificial Analysis's streaming speech recognition evaluation, it also ranked first on both final-transcript and first-partial-transcript accuracy.
Source: X @ArtificialAnlys
Lower WER (word error rate) means higher accuracy, and latency is the time from the end of speech until text appears. The figures are from the Artificial Analysis streaming leaderboard.
| Final transcript | WER | Latency |
|---|---|---|
| MAI-Transcribe-2-Streaming (1st of 38 models) | 2.5% | 0.13 s |
| Grok Voice Transcribe 2.0 | 2.7% | 0.49 s |
| Muse Voice Transcribe | 3.1% | 0.16 s |
| ElevenLabs Scribe v2 Realtime | 3.6% | 0.14 s |
| First partial transcript | WER | Latency |
|---|---|---|
| MAI-Transcribe-2-Streaming | 2.5% | 0.12 s |
| Grok Voice Transcribe 2.0 | 3.4% | 0.49 s |
| Muse Voice Transcribe | 3.6% | 0.13 s |
| ElevenLabs Scribe v2 Realtime | 3.6% | 0.13 s |
The introductory price is $0.54 per hour of audio ($9 per 1,000 minutes) through the end of this year.
MAI-Voice-2.1: 23 languages in one voice
MAI-Voice-2.1 supports 23 languages and 26 locales. When one voice switches between English, Chinese, and German, it stays the same person's voice and speaks each language with a natural local accent. You can use it in a tutoring app where the teacher's voice stays the same when the language changes mid-lesson, a multilingual assistant that answers in the same voice in whatever language it is spoken to, or a service that keeps its brand voice in every language. The price is $22 per million characters.
MAI-Voice-2.1-Flash: 45 seconds of audio in 150 ms
Flash supports the same languages and multilingual speakers and is tuned for high-volume, low-latency work. It generates 45 seconds of audio, start to finish, with 150 ms of latency, with inference 55% faster than comparison models at about 60% lower cost. The price is $15 per million characters.
Both voice models clone a voice from a reference clip a few seconds long across all supported languages and have consent safeguards built in. In Azure, voice prompts use a 5-to-60-second reference clip as a limited-access feature, and you need approval before you can use it.
Voice agents that listen, decide, and speak
A voice agent has to run the loop of listening, understanding, deciding, and speaking fast enough that a person experiences it as conversation. Pairing Transcribe-2-Streaming with Voice-2.1-Flash buys time at both ends, listening and speaking, and that gives the agent room to reason, use tools, and check its answer.
- A customer support agent that transcribes as the caller speaks and starts the needed work before the call ends
- A multilingual assistant that recognizes the language spoken and replies in the same voice in that language, out of 23
- Learning and media services that assign separate speakers in tutoring, role-play, and narration
Pricing and where to use them
| Model | What it does | Price |
|---|---|---|
| MAI-Transcribe-2-Streaming | Live transcription | $0.54 per hour (introductory price through year-end) |
| MAI-Voice-2.1 | Multilingual speech generation | $22 per million characters |
| MAI-Voice-2.1-Flash | Fast speech generation | $15 per million characters |
| Where | Models and how to access them |
|---|---|
| Microsoft Foundry | All three models. Transcription docs, voice docs |
| Azure Voice Live | All three models |
| Vercel AI Gateway | All three models. Example: microsoft/mai-transcribe-2-streaming |
| OpenRouter | Voice-2.1 and Flash: microsoft/mai-voice-2.1, microsoft/mai-voice-2.1-flash |
| LiveKit | All three models coming soon |
| MAI Playground | Try a voice agent running all three models together in the Chatter demo |
Related reading
- ElevenLabs Releases Eleven v4: An AI Voice Model That Speaks Like a Person, in 90+ Languages Including Korean
- Revolut Adopts ElevenLabs Agents and Cuts AI Voice Support Resolution Time by 8x
- The "Grok Wins" Benchmark Musk Shared, Checked Against the Numbers