Microsoft releases live speech to text for 61 locales

Microsoft made multilingual live transcription available in Azure Speech, part of its Foundry AI platform. One live session can recognize many languages, also when speakers change language in the same session. A new engine removes a second recognition pass, and final text usually comes in about 100 milliseconds. Microsoft gave no test results for precision.

How it works

In the preview, a second recognition pass was done after live recognition to refine each final result. Users had to wait for it. At this time, a streaming speech model does this work in the live system. Users of this tool get the new engine with no code change.

What you get

• You can give the tool names and terms to look for, a maximum of 200 phrases each.
• 61 locales (language and region pairs), up from 29 in the preview.
• Transcripts name each speaker.
• 6 regions, for example East US and North Europe.

What is not known

Microsoft gave no test results for precision, and did not compare it with other tools. The speed number is from Microsoft, and it tells that final text comes usually in that time.

Insight: Voice tools now compete on where they operate

Nvidia gave a free model for recorded audio only. Mistral tells that its Voxtral 3 operates in the EU or on your servers. ElevenLabs tests live translation in alpha, with no results. Microsoft puts live transcription in its cloud in 6 regions. This reads as a move to be first in live voice, where the area of data is the test. Find where each vendor keeps audio.