Voice Live, Canada East: playground offers only "Bring my Foundry Model", and where is speech recognition processed?

Hi Priyan,

Your API results are consistent with the current Microsoft Learn regional availability documentation. However, the regional availability of a feature and confirmation that every processing stage stays in Canada are two separate questions.

1. Foundry playground model availability

The current Supported regions for Azure Speech documentation lists gpt-4.1-mini as Regional for canadaeast, and HD voices are also listed as Regional for that resource region.

Since your API session works with gpt-4.1-mini, the playground behavior appears to be a portal model-selection or regional metadata inconsistency, rather than proof that the model is unsupported in Canada East.

The error mentioning foundry-model suggests that the playground is attempting to start a session with its generic Foundry-model selection instead of a supported prebuilt model. I couldn't confirm a public incident specifically covering this behavior. For now, the API is a reasonable workaround. I would report the playground issue to Azure Support with the resource ID, region, API version, timestamp, and WebSocket error.

2. Canadian data residency

The important distinction is that Voice Live has separate processing scopes for language models, speech recognition, and HD voice synthesis.

a. Speech recognition using azure-speech

Canada East supports real-time speech transcription in the general Azure Speech regional availability table. However, the current Voice Live processing-scope table does not explicitly list a processing scope for the azure-speech transcription option. It separately identifies the scopes for HD voices and MAI Transcribe.

Consequently, I would not claim that Voice Live's azure-speech transcription is guaranteed to remain in Canada based solely on the resource region. Microsoft needs to confirm the processing location for this specific Voice Live configuration, including both en-US and fr-CA.

b. Semantic VAD, noise suppression, and echo cancellation

The Voice Live API documentation confirms that these operations are performed as part of the service's audio-processing pipeline before speech recognition and language generation. However, the public documentation does not independently specify a regional processing scope for each of these operations.

For a strict Canadian public-sector requirement, ask Microsoft to confirm whether azure_semantic_vad, noise suppression, and server-side echo cancellation are all processed exclusively in Canada East.

c. Content filtering and abuse monitoring

Voice Live applies content filtering according to the filtering configuration of the supported model or connected deployment. The gpt-4.1-mini entry is Regional for Canada East, which supports regional processing for that model's inference.

That alone does not establish that every content-filtering and abuse-monitoring operation, including any associated review or retention, is guaranteed to remain in Canada. The Voice Live privacy documentation states that the service processes data to provide its functionality and monitor for violations of applicable Product Terms. It also distinguishes the API's own retention behavior from the data handling of services it uses.

Ask Microsoft to confirm the processing and storage location for both content filtering and abuse monitoring under your exact configuration, including any retention of flagged content.

d. Cross-region routing or failover

The current regional table lists both gpt-4.1-mini and HD voices as Regional for Canada East. That is useful evidence in favor of regional processing for those components. However, it is not sufficient to certify the complete Voice Live session as Canada-only, particularly because the documentation does not explicitly establish the scope of every speech-processing and safety component above.

The public documentation also explains that Voice Live can route HD voice synthesis to an HD-capable region when a voice is unavailable locally. You should therefore obtain explicit confirmation of the routing behavior for en-US-Andrew:DragonHDLatestNeural, including what happens during capacity constraints or service failover.

Recommended next step

Because this is a Canadian public-sector project with a strict residency requirement, I would open an Azure Support case and request written confirmation for the exact configuration:

Resource region: canadaeast

Model: gpt-4.1-mini

Transcription: azure-speech, using en-US and fr-CA

Voice: en-US-Andrew:DragonHDLatestNeural

Audio processing: azure_semantic_vad, noise suppression, and server-side echo cancellation

Safety processing: content filtering and abuse monitoring, including any related data retention

Ask Microsoft to identify the processing geography for each component and confirm whether any component or failover path can process session data outside Canada.

The relevant Microsoft Learn references are Supported regions for Azure Speech, How to use the Voice Live API, and Data, privacy, and security for Azure AI Voice Live API.

My conclusion is that your prebuilt-model API usage is consistent with the published regional availability information, but the documentation currently available does not establish a complete Canada-only guarantee for every stage of this Voice Live pipeline. I would not treat the residency requirement as verified until Microsoft confirms the unspecified scopes in writing.