Microsoft's MAI-Transcribe-2-Streaming returns first results in ~100 ms across 60 languages
Microsoft AI released MAI-Transcribe-2-Streaming, a real-time transcription model covering 60 languages with first partial results in just over 100 milliseconds. It also launched two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash, aimed at voice agents.
Microsoft AI has released MAI-Transcribe-2-Streaming, a real-time speech-to-text model that supports 60 languages and delivers its first partial results in just over 100 milliseconds. Microsoft also launched two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. All three are aimed at voice agents.
MAI-Transcribe-2-Streaming
- Languages: 60
- Latency: first partial results in just over 100 ms
- Price: $0.54 per hour of audio at an introductory rate that runs through the end of the year. Post-promotion pricing has not been disclosed.
- Accuracy: Microsoft claims the model ranks first for accuracy on Artificial Analysis. No word error rate or other numeric score was provided, and the ranking has not been independently confirmed here.
Microsoft says the low latency lets voice agents begin responding while a user is still mid-sentence.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash
MAI-Voice-2.1 is built to speak 23 languages in the same voice, with a native accent in each one, according to Microsoft.
The Flash variant targets speed and cost:
- Latency: 150 ms, according to Microsoft
- Price: $15 per million characters, down from $22. The source does not state which model the $22 figure refers to, but it is presumably the standard MAI-Voice-2.1 rate.
Both voice models can clone a voice from a few seconds of reference audio. Microsoft says built-in safeguards are meant to prevent misuse but has not detailed how they work. In a Microsoft test, about half of 4,000 participants judged the synthetic voices to belong to a real person.
Availability
The models are available through Microsoft Foundry and the MAI Playground, among other platforms. The two voice models are also listed on OpenRouter. The source does not say whether the transcription model is on OpenRouter.
Not disclosed
Parameter counts, training data cutoffs, context limits, and per-model benchmark figures were not published in the announcement coverage. The standard MAI-Voice-2.1 price is also not explicitly stated.
What this means
The release targets the latency budget of conversational voice agents. A first-partial time near 100 ms for transcription and 150 ms for speech synthesis leaves more room for the language model in between. Cascaded pipelines (speech-to-text, LLM, text-to-speech) have usually struggled with that budget compared with end-to-end speech models.
The pricing is aggressive but time-limited. The $0.54 per hour transcription rate is introductory, so teams should wait for post-promotion pricing before committing production volume. The Flash tier's $15 per million characters undercuts the $22 comparison point Microsoft cites.
Listing the voice models on OpenRouter suggests Microsoft wants distribution beyond its own cloud, which would put the MAI line in direct competition with specialist voice vendors such as ElevenLabs. The "ranks first" accuracy claim rests on a single third-party leaderboard and comes without published error rates, so developers should test on their own audio, especially for accented speech, noisy channels, and lower-resource languages among the 60 supported. The voice-cloning capability from seconds of audio, paired with a test where half of listeners took the output for a human, makes the unspecified safeguards the detail to watch.
Related Articles
Microsoft launches seven models at Ignite; claims MAI Code 2 matches 96% of Claude Opus 5.5 at 5% of price
Microsoft unveiled seven AI models at Ignite on Nov. 17, aimed at coding, cybersecurity and other enterprise tasks. The company claims MAI Code 2 delivered 96% of Claude Opus 5.5's performance across four coding benchmarks at 5% of the price, using a promotional rate.
Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license
Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.
Microsoft launches Autopilot agent, Home interface and Copilotforce with Salesforce at Ignite 2026
At Ignite 2026, Microsoft introduced Autopilot, an always-on Copilot agent with its own Entra identity, available now to Frontier program subscribers. It also unveiled a new Home interface, Copilot Workspaces, Copilot Code and Copilotforce, a Salesforce integration with 38 prebuilt skills.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Comments
Loading...