Microsoft releases three audio models, with Flash speech generation inference taking approximately 45ms

Article

Microsoft completes the full voice Agent pipeline: transcribe while speaking, speech generation as low as 45ms Beating AI News reports that Microsoft AI has launched 3 audio models at once:MAI-Transcribe-2-Streaming、MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The former handles real-time speech-to-text, while the latter two generate speech from text, primarily targeting low-latency voice Agents.

Among them,MAI-Transcribe-2-Streaming can continuously output text before the speaker has finished, supporting 60 languages and automatic language detection. Artificial Analysis's streaming speech transcription leaderboard shows that its final transcription and first partial transcription accuracy both rank first, with a word error rate of 2.5%. The final text is returned approximately 0.13 seconds after speech end is detected, while Grok Voice Transcribe 2.0 is 2.7%and 0.49 seconds, respectively.MAI-Voice-2.1 supports 23 languages and allows the same voice to switch between different languages. The Flash version is specifically optimized for low latency, with model inference taking approximately 45ms and priced at 15 USD per million characters; the standard version takes approximately 550ms and is priced at 22 USD. Both models support matching a voice using a short reference audio clip.

This combination addresses the waiting time that most affects the user experience of voice Agents. The system can begin transcription and inference while the user is speaking, then use Flash to quickly generate a spoken response, without waiting for the entire sentence to end before starting work.

Original link https://m.theblockbeats.info/flash/370130

Key points

Beating AI reports that Microsoft has launched MAI-Transcribe-2-Streaming、MAI-Voice-2.1 and MAI-Voice-2.1-Flash。
reports that MAI-Transcribe-2-Streaming is a real-time speech-to-text model and supports 60 languages and automatic language detection.
The article states that the Artificial Analysis leaderboard shows MAI-Transcribe-2-Streaming has a word error rate of 2.5%, with the final text returned approximately 0.13 seconds after speech end is detected.
The article states that MAI-Voice-2.1 supports 23 languages and allows the same voice to switch between different languages.
The article states that Flash model inference takes approximately 45ms and costs 15 USD per million characters; the standard version takes approximately 550ms and costs 22 USD。
The article states that both speech generation models support matching a voice through a short reference audio clip.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This model combination splits voice interaction into two stages: first, continuously converting the user's speech into text, and then generating speech from the reply text. According to the report, the system can begin processing before the user has finished speaking, reducing the time a voice Agent waits for a response.

Why it matters to readers

If the performance, pricing, and capabilities reported are confirmed, these models could help developers build voice Agents with faster responses, multilingual support, and voice matching.

Beginners can think of it as a voice assistant's “ears” and “mouth”: the transcription model listens and turns speech into text, while the speech-generation model reads the reply aloud.

Risks and unknowns

The material does not include an official Microsoft announcement or product page, so the release information has not been confirmed by a primary source.
The rankings, accuracy, latency, and pricing lack independent verification.
The available scope, launch platforms, testing conditions, and actual adoption of the models are not specified.
Matching a voice using reference audio may involve authorization, impersonation, and privacy risks, but the material does not describe any relevant safeguards.

Microsoft releases three audio models, with Flash speech generation inference taking approximately 45ms

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center