MAI-Transcribe-2-Streaming: live speech to text
Scoped recheck: 24 October 2026. Microsoft Learn's overview now lists Southeast Asia, not South India. Price, 60 languages, Malayalam and public preview were re-confirmed. Original publication date unchanged.
Microsoft's MAI-Transcribe-2-Streaming turns live audio into text while a person speaks. It supports Malayalam among 60 languages. This is a paid developer service in public preview, not a free dictation app.
Official publisher artwork hosted by its publisher. Official source. Independent guide, not affiliated or sponsored. Artwork and trademarks belong to their owners.
What does it do?
The launch also introduced MAI-Voice-2 and MAI-Voice-2-Custom, separate voice-generation models rather than features included in this transcription price.
The model sends quick, unfinished text as it hears speech. These partial results can change. It then sends a stable final transcript for each segment. A caption app can show words immediately and replace them with the final text.
Microsoft says first partial results arrive just over 100 milliseconds after audio arrives. Its accuracy and speed rankings are vendor-reported benchmark claims, not a test of your microphone, Kerala accent or network.
Who is it for in India?
Developers building live captions, voice assistants, call tools and meeting apps may find it useful. The official language table includes Malayalam, Hindi, Tamil, Kannada, Telugu and other Indian languages. That confirms support, not perfect recognition of mixed-language speech or names.
On 24 October 2026 the Microsoft Learn overview listed Sweden Central, Central US, East US 2 and Southeast Asia, and said the model can be accessed globally. South India was not listed. Region lists can change, so use the current Foundry deployment screen before planning data residency.
How to try it
- Read the Microsoft Learn overview and preview terms.
- Create a Microsoft Foundry resource in a supported region. Deploy the named model; check the available region and billing details.
- Choose the Azure Speech SDK or the OpenAI Realtime-compatible WebSocket route. A WebSocket keeps a live connection open as you send audio.
- Use a short recording you have permission to process. The Realtime guide requires mono PCM16 audio at 16,000 or 24,000 samples per second.
- Show partial text as temporary. Use finalized text for saved captions or actions.
- Test Malayalam and English names, numbers, background noise and code-switching against a human transcript.
- Set usage limits, keep credentials on your server, and review privacy rules before sending customer or patient audio.
Price and limits
The model card lists an introductory US$0.54 per audio hour through 31 December 2026. At the sourced approximate exchange rate of Rs 96.43 per US dollar, this is roughly Rs 52.07 per hour. This is a conversion estimate, not an Indian invoice; tax, other services and later prices can differ.
Microsoft Learn says each session can last up to one hour. It calls the feature public preview, without a service-level agreement, and does not recommend it for production workloads.
What changed since the original post?
We confirmed Malayalam in the language list and added the preview and one-hour limits. The launch article lists LiveKit as coming soon, so it is not presented as an available route. No model was deployed or audio tested for this guide.
Frequently asked questions
Is it free?
No free allowance was confirmed. The published introductory audio rate is paid.
Can ordinary users install it?
It is an API model for developers, not a standalone consumer app.
Can I trust a partial transcript?
It may change. Review the final transcript, especially before acting on names, money or instructions.
Does it translate?
This model transcribes. A separate translation step is needed if you want another language.
Sources: Microsoft launch · Model card · Preview overview · Realtime guide and languages · Approximate currency conversion. Checked 7 October 2026.
Scoped recheck: Learn overview and regions, Realtime guide, Microsoft launch. Checked 24 October 2026.