Azure Speech adds real-time multichannel speech-to-text in public preview

Microsoft has launched Real-Time Multichannel Speech-to-Text in public preview for Azure Speech, part of Microsoft Foundry. The feature lets Azure Speech process stereo audio while preserving channel separation throughout the real-time transcription workflow. This is useful for organizations that already record call participants on separate channels.

In contact center and conversational AI work, customer and agent audio is often captured on separate channels. That separation is valuable, especially when speakers overlap, interrupt, or react while the other person is still talking. Previously, keeping channels separate meant one of two compromises. Teams could split the audio and run parallel transcription pipelines, preserving quality but adding operational overhead and processing cost. Or they could merge channels into a single stream, simplifying implementation but losing the benefit of separate channels for overlapping speech.

The new capability avoids this tradeoff. Azure Speech can ingest stereo audio directly and return results with channel attribution preserved, without requiring separate recognition pipelines per channel.

For existing users, the integration is straightforward. It works with the existing Speech SDK starting with version 1.51.0, and enabling it is a single configuration property. A standard Python recognition loop uses speech_config.set_property(speechsdk.PropertyId.Speech_EnableMultiChannelProcessing, "true"), then reads results via evt.result.channel. The rest of the workflow—event handling, continuous recognition, application architecture—stays the same. There is no separate service to deploy and no need to create parallel recognition sessions for each channel.

The main architectural implication is that teams that previously maintained multiple transcription workflows and merged results afterward can consolidate into a single pipeline. Channel information remains available in real time for downstream work like summarization or insight generation, where knowing who said what matters.

There are caveats. This is a public preview, and Microsoft notes that individual transcription modes may differ in how they interact with speaker diarization. The feature is valuable specifically in scenarios where separate channels already exist—it preserves conversational structure, it does not create it. The benefits for overlapping speech may vary, as actual quality gains over mixed-audio transcription will depend on recording quality and acoustic conditions.

Documentation is available on Microsoft Learn, and Microsoft is directing feedback through Azure support, Stack Overflow, and the Azure AI Speech SDK samples repository on GitHub.