The Deepgram Native flavor uses the Deepgram's new Voice Agent API that provides direct Speech-to-Speech interpretation of the user's conversation. This setup allows for real-time interpretation and response generation, streamlining interactions without intermediate steps.
The project uses the native WebSocket connection to push the decapsulated RTP from the user to the Deepgram engine and get the response back. Then, it grabs the response, packs it back by adding the RTP header, and streams it back to the user.
It does not have any transcoding capabilities, thus communication is limited to g711 PCMU and PCMA codecs.
The following parameters can be tuned for this engine:
| Section | Parameter | Environment | Mandatory | Description | Default |
|---|---|---|---|---|---|
deepgram_native |
key |
DEEPGRAM_API_KEY |
yes | Deepgram API key | not provided |
deepgram_native |
speech_model |
DEEPGRAM_NATIVE_SPEECH_MODEL |
no | Deepgram's speech detection model | nova-2-conversationalai |
deepgram_native |
voice |
DEEPGRAM_NATIVE_VOICE |
no | Deepgram's voice used for speaking back the response | aura-asteria-en |
deepgram_native |
welcome_message |
DEEPGRAM_NATIVE_WELCOME_MSG |
no | A welcome message to be played back to the user when the call starts | `` |
deepgram_native |
instructions |
DEEPGRAM_INSTRUCTIONS |
no | Configures the LLM instructions | `` |
deepgram_native |
llm_url |
DEEPGRAM_LLM_URL |
no | Configures the LLM URL | `` |
deepgram_native |
llm_auth |
DEEPGRAM_LLM_KEY |
no | Configures the LLM API key | `` |
deepgram_native |
llm_model |
DEEPGRAM_LLM_MODEL |
no | Configures the LLM model | `` |
deepgram_native |
disable |
DEEPGRAM_NATIVE_DISABLE |
no | Disables the flavor | false |