streaming-stt-deepgram

name	streaming-stt-deepgram
description	Real-time streaming speech-to-text via Deepgram WebSocket API
category	voice

Deepgram Streaming STT

Real-time streaming speech-to-text using Deepgram's Nova-2 model via WebSocket.

Setup

Set DEEPGRAM_API_KEY in your environment or agent secrets store.

Features

• Sub-300 ms latency interim transcripts
• Native speaker diarization (word-level speaker labels)
• Automatic punctuation insertion
• Word-level timestamps and per-word confidence scores
• Auto-reconnect on connection drops (exponential back-off 100 ms → 5 s)
• Keyword boosting via providerOptions.keywords

Configuration

In agent.config.json:

{
  "voice": {
    "stt": "deepgram"
  }
}

Provider-specific options via providerOptions:

{
  "voice": {
    "stt": "deepgram",
    "providerOptions": {
      "model": "nova-2",
      "diarize": true,
      "keywords": ["AgentOS:2", "Deepgram:1.5"],
      "endpointing": 300
    }
  }
}

Events

Event	Payload	Description
transcript	TranscriptEvent	Every hypothesis (interim + final)
interim_transcript	TranscriptEvent	Non-final hypothesis
final_transcript	TranscriptEvent	Stable, final hypothesis
speech_start	—	First non-empty word in an utterance
speech_end	—	Deepgram speech_final flag raised
error	Error	Unrecoverable provider error
close	—	Session fully terminated