Skip to main content

Microsoft Azure Provider Guide

Azure is the one provider that needs two configuration paths in Plug-N-Meet. Use the correct section for each service:

PartProvider typeServices
Part 1 — native Azure Cognitive Servicesazuretranscription, translation, speech-synthesis
Part 2 — Azure OpenAI / Azure AI Foundryopenaiai_agent, ai_text_chat, meeting_summarizing

Why Two Parts​

Part 1 — native azure. Azure Speech and Azure Translator are purpose-built speech and translation services, and the azure provider talks to them directly: Speech-to-Text over the Speech SDK WebSocket, the Translator REST API, and Neural TTS. These three services are exactly what native azure implements. It has no LLM runtime — ai_text_chat, meeting_summarizing, and ai_agent are not available on provider: "azure".

Part 2 — Azure OpenAI through provider: "openai". Azure OpenAI and Azure AI Foundry expose an OpenAI-compatible surface at your resource's /openai/v1 path. By defining that endpoint as an openai account with auth_header: "api-key", the LLM services run on Azure models with no other code path.

Both parts live in the same insights block, so an Azure-only deployment uses two provider sections and six service entries.

Part 1: Native Azure Cognitive Services​

Provider Accounts​

Two accounts, because Speech and Translator are separate Azure resources with separate keys:

insights:
providers:
azure:
# Regional Speech resource: STT and Neural TTS
- id: "eastus_standard"
credentials:
api_key: "YOUR_AZURE_SPEECH_KEY"
region: "eastus"

# Global Translator resource: chat and caption translation
- id: "transGlobal"
credentials:
api_key: "YOUR_AZURE_TRANSLATOR_KEY"
region: "global"
options:
endpoint: "https://api.cognitive.microsofttranslator.com/translate?api-version=3.0"
KeyTypeRequiredDescription
idstringYesUnique account name, for example "eastus_standard" or "transGlobal".
credentials.api_keystringYesResource key.
credentials.regionstringYesAzure region. Use the Speech resource region for transcription and speech-synthesis, and "global" for a global Translator resource.
options.endpointstringTranslator onlyFull Translator request URL including /translate?api-version=3.0. The translation service fails without it.
warning

The Translator endpoint must be the complete operation URL: https://api.cognitive.microsofttranslator.com/translate?api-version=3.0. Set it on the provider account, not on the service entry. A bare host is not enough.

transcription​

Streams allowed speakers' microphone audio to Azure Speech-to-Text and publishes partial and final caption chunks to the room.

transcription:
provider: "azure"
id: "eastus_standard"
options:
max_selected_trans_langs: 2
pricing:
default:
price_per_hour: 0.36
KeyTypeDefaultDescription
options.max_selected_trans_langsinteger2Maximum simultaneous caption languages per room.

Azure Speech Translation performs live speech-to-text and target-language caption translation in a single pass, so rooms that request translated captions get them from the same audio stream rather than a second service call.

translation​

Uses the Azure Translator REST API for chat messages and bulk text translation. Rate, length, and language limits are Azure's.

translation:
provider: "azure"
id: "transGlobal"
options:
max_selected_trans_langs: 5
pricing:
default:
price_per_million_characters: 10.00
KeyTypeDefaultDescription
options.max_selected_trans_langsinteger5Maximum target languages per room. Separate from the caption limit.

This service must point at the Translator account. Pointing it at the regional Speech account fails, because Speech keys are not accepted by the Translator API.

speech-synthesis​

Speaks translated captions with Azure Neural voices. Voices are mapped per language, which is the main advantage of native Azure TTS.

speech-synthesis:
provider: "azure"
id: "eastus_standard"
options:
voice-en: "en-US-JennyNeural"
voice-es: "es-ES-ElviraNeural"
voice-fr: "fr-FR-DeniseNeural"
voice-de: "de-DE-AmalaNeural"
voice-bn: "bn-IN-TanishaaNeural"
pricing:
default:
price_per_million_characters: 16.00
KeyTypeDescription
options.voice-<lang>stringNeural voice used when synthesizing that target language, for example voice-en, voice-es, voice-fr, voice-de, voice-bn.

Add one voice-<lang> key per language you want spoken. A language without a voice is not spoken aloud, and an invalid voice ID fails only that language, not the session.

Pricing​

ServiceKeyMeters
transcriptionpricing.default.price_per_hourSPEECH_TRANSCRIPTION_USAGE from transcribed duration
translationpricing.default.price_per_million_charactersCHAT_TRANSLATION_USAGE from translated character count
speech-synthesispricing.default.price_per_million_charactersSYNTHESIZED_SPEECH_USAGE from synthesized characters

Translation cost is metered by characters, not tokens. Omit the rate and translation artifacts show zero cost.

Part 2: Azure OpenAI via the openai Provider​

Add your Azure OpenAI resource as a second provider section. The base URL is the resource's OpenAI-compatible v1 path:

insights:
providers:
openai:
- id: "azure-openai"
credentials:
api_key: "YOUR_AZURE_OPENAI_KEY"
options:
endpoint: "https://<your-resource>.openai.azure.com/openai/v1"
auth_header: "api-key"
KeyValueWhy
credentials.api_keyAzure OpenAI keySent as the request credential.
options.endpointhttps://<your-resource>.openai.azure.com/openai/v1Azure's OpenAI-compatible surface. https is converted to wss automatically for WebSocket services.
options.auth_header"api-key"Azure expects api-key: <key>. The default Authorization: Bearer is rejected.

For ai_agent, also set realtime_auth_header: "api-key". The agent opens its own WebSocket and does not inherit the account header.

Deployment names, not model names​

Every model field in the service options is an Azure deployment name, not the upstream OpenAI model identifier:

  • options.realtime_model / options.live_model
  • options.chat_model
  • options.summarize_model
  • options.transcription_model
  • options.delegation_model

pricing keys follow the same rule. If your deployment is named gpt-realtime-prod, write pricing: gpt-realtime-prod: ... (quoted), or use pricing.default to avoid repeating deployment names.

Not every OpenAI model is available in every region. Deploy the models you need in the Foundry portal first, then reference those deployment names.

ai_text_chat​

ai_text_chat:
provider: "openai"
id: "azure-openai"
options:
chat_model: "gpt-chat-prod"
summarize_model: "gpt-mini-prod"
context_window: 5
pricing:
default:
input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 10.00

meeting_summarizing​

meeting_summarizing:
provider: "openai"
id: "azure-openai"
options:
transcription_model: "whisper-prod"
summarize_model: "gpt-mini-prod"
pricing:
default:
input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 10.00

Large recordings are chunked into segments of 20 MB or less, each transcribed with the audio transcription deployment, then summarized.

ai_agent​

Both agent architectures run on Azure deployments. Realtime mode:

ai_agent:
provider: "openai"
id: "azure-openai"
options:
realtime_model: "gpt-realtime-prod"
summarize_model: "gpt-mini-prod"
default_voice: "marin"
turn_detection: "semantic_vad"
realtime_auth_header: "api-key"
pricing:
realtime_model:
input_price_per_million_tokens: 4.00
audio_input_price_per_million_tokens: 32.00
image_input_price_per_million_tokens: 4.00
output_price_per_million_tokens: 16.00
audio_output_price_per_million_tokens: 64.00
summarize_model:
input_price_per_million_tokens: 0.25
output_price_per_million_tokens: 2.00

If your resource exposes a Live model, set live_model and delegation_model (both deployment names). delegation_model is enforced at startup, and Mode 2 bills speech by pricing.live_model.price_per_minute instead of per token. See the OpenAI & Self-Hosted guide.

Complete All-Azure Production Configuration​

All six services on Microsoft Azure: Speech and Translator for the speech stack, Azure OpenAI for the LLM stack.

insights:
enabled: true
providers:
azure:
- id: "eastus_standard"
credentials:
api_key: "YOUR_AZURE_SPEECH_KEY"
region: "eastus"
- id: "transGlobal"
credentials:
api_key: "YOUR_AZURE_TRANSLATOR_KEY"
region: "global"
options:
endpoint: "https://api.cognitive.microsofttranslator.com/translate?api-version=3.0"

openai:
- id: "azure-openai"
credentials:
api_key: "YOUR_AZURE_OPENAI_KEY"
options:
endpoint: "https://<your-resource>.openai.azure.com/openai/v1"
auth_header: "api-key"

services:
transcription:
provider: "azure"
id: "eastus_standard"
options:
max_selected_trans_langs: 2
pricing:
default:
price_per_hour: 0.36

translation:
provider: "azure"
id: "transGlobal"
options:
max_selected_trans_langs: 5
pricing:
default:
price_per_million_characters: 10.00

speech-synthesis:
provider: "azure"
id: "eastus_standard"
options:
voice-en: "en-US-JennyNeural"
voice-es: "es-ES-ElviraNeural"
voice-fr: "fr-FR-DeniseNeural"
voice-de: "de-DE-AmalaNeural"
voice-bn: "bn-IN-TanishaaNeural"
pricing:
default:
price_per_million_characters: 16.00

ai_text_chat:
provider: "openai"
id: "azure-openai"
options:
chat_model: "gpt-chat-prod"
summarize_model: "gpt-mini-prod"
context_window: 5
pricing:
default:
input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 10.00

meeting_summarizing:
provider: "openai"
id: "azure-openai"
options:
transcription_model: "whisper-prod"
summarize_model: "gpt-mini-prod"
pricing:
default:
input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 10.00

ai_agent:
provider: "openai"
id: "azure-openai"
options:
realtime_model: "gpt-realtime-prod"
summarize_model: "gpt-mini-prod"
default_voice: "marin"
turn_detection: "semantic_vad"
realtime_auth_header: "api-key"
pricing:
realtime_model:
input_price_per_million_tokens: 4.00
cached_input_price_per_million_tokens: 0.40
audio_input_price_per_million_tokens: 32.00
audio_cached_input_price_per_million_tokens: 0.40
image_input_price_per_million_tokens: 4.00
output_price_per_million_tokens: 16.00
audio_output_price_per_million_tokens: 64.00
summarize_model:
input_price_per_million_tokens: 0.25
output_price_per_million_tokens: 2.00

Everything stays inside your Azure tenant and region: audio never leaves the speech resources, and LLM traffic goes to your own Foundry deployment.

Common Pitfalls​

  • provider: "azure" for the agent, chat, or summaries. Native azure does not implement those services. Use the azure-openai account with provider: "openai".
  • Missing Translator endpoint path. The endpoint must be the full /translate?api-version=3.0 URL on the provider account.
  • Default auth header against Azure OpenAI. Set auth_header: "api-key" on the account and realtime_auth_header: "api-key" on the agent, or every call returns 401.
  • OpenAI model names as deployment names. Use the deployment name you created in Foundry, and quote any pricing key that contains a hyphenated custom name you did not create as default.
  • Wrong region on the Speech account. TTS and STT both fail when region does not match the resource.