Azure OpenAI

Azure OpenAI serves OpenAI’s text-to-speech models from a resource in your own Azure subscription. The same credential also serves speech recognition. Azure OpenAI tuning is passed in synthesizer.options.

{
  "verb": "say",
  "text": "Thanks for calling. Your order ships today.",
  "synthesizer": {
    "vendor": "azureopenai",
    "language": "en-US",
    "voice": "alloy",
    "options": {
      "speed": 1.1,
      "instructions": "Speak in a warm, upbeat tone."
    }
  }
}

You can also set these in the portal under Extra Options on the speech credential, in which case they apply to every request that uses that credential. Settings supplied on the verb take precedence.

Credential

In the Azure portal, open your Azure OpenAI resource and copy a key and the endpoint from Keys and Endpoint. Add them as an Azure OpenAI speech credential in the portal or API: the key as the API key, and the endpoint, such as https://myres.openai.azure.com, as the endpoint. jambonz sends the key in the api-key header. One credential is used for both text-to-speech and speech-to-text. Microsoft Entra ID tokens are not supported.

Deployments

Azure OpenAI addresses a model by the name of the deployment you created for it, not by the model name. Enter the TTS deployment name on the speech credential; it defaults to gpt-4o-mini-tts, so a deployment with that name needs no setting. Set options.model_id to use another deployment for one verb.

Model	Notes
`gpt-4o-mini-tts`	Default. Follows `instructions`.
`tts`	Lower latency.
`tts-hd`	Higher quality.

Voices and languages

The portal lists alloy, echo, fable, onyx, nova and shimmer; alloy is the default. Any other voice your deployment supports can be entered by name. Every voice is multilingual and speaks the language the text is written in, so synthesizer.language is not sent to Azure.

Options

model_id

string

The TTS deployment name, overriding the credential’s deployment.

speed

number

Speaking rate, 0.25 – 4.0 times normal speed.

instructions

string

How to speak, such as tone, pace or accent. Followed by gpt-4o-mini-tts deployments.

Every option is optional. An option you don’t set is not sent, so Azure’s own default applies.

Endpoint override

The endpoint on the credential is required, since every Azure OpenAI resource has its own host. To point one verb or credential elsewhere, such as a proxy, set api_uri in synthesizer.options or the credential’s Extra Options; it takes precedence over the credential’s endpoint, and the credential’s api_uri in Extra Options applies to STT too. jambonz drops any path from /openai on, so a full Azure URL works. An http:// prefix turns off TLS.

Streaming

Azure OpenAI has no token-streaming websocket, so it does not support TTS streaming: use it for one-shot say, not for say with stream: true or the agent verb.