Skip to main content
The Deepgram plugin provides high-quality speech recognition (STT) and text-to-speech (TTS) capabilities.

​
Installation

uv add vision-agents-plugins-deepgram

​
Authentication

Set your API key in the environment: 
export DEEPGRAM_API_KEY=your_deepgram_api_key

​
Components

​
STT - Speech-to-Text

High-quality speech recognition using Deepgram’s Flux model: 
from vision_agents.plugins import deepgram

stt = deepgram.STT(
    model="flux-general-en",
    eager_turn_detection=True
)

# Use in an agent
agent = Agent(
    stt=stt,
    # ... other config
)
​
model
string
default:"flux-general-en"
Deepgram model to use for transcription. See Flux models
​
eager_turn_detection
bool
default:"True"
Enable eager end-of-turn detection for faster response times
​
api_key
string
Optional API key. Defaults to DEEPGRAM_API_KEY environment variable

​
TTS - Text-to-Speech

Low-latency text-to-speech using Deepgram’s Aura model: 
from vision_agents.plugins import deepgram

tts = deepgram.TTS(
    model="aura-2-thalia-en",
    sample_rate=16000
)

# Use in an agent
agent = Agent(
    tts=tts,
    # ... other config
)
​
model
string
default:"aura-2-thalia-en"
Deepgram Aura voice model. See Available Voices
​
sample_rate
int
default:"16000"
Audio sample rate in Hz
​
api_key
string
Optional API key. Defaults to DEEPGRAM_API_KEY environment variable

​
Available Voices

Deepgram offers various Aura voice models:
VoiceDescriptionLanguage
aura-2-thalia-enDefault female voiceEnglish
aura-2-orion-enMale voiceEnglish
aura-2-asteria-enFemale voiceEnglish
aura-2-perseus-enMale voiceEnglish
See TTS Models for all options.

​
Usage Example

Combine STT and TTS in a voice agent: 
from vision_agents.core import Agent, User
from vision_agents.plugins import deepgram, openai, getstream

agent = Agent(
    edge=getstream.Edge(),
    agent_user=User(name="Voice Assistant"),
    instructions="You are a helpful voice assistant.",
    llm=openai.LLM("gpt-4.1"),
    stt=deepgram.STT(
        model="flux-general-en",
        eager_turn_detection=True
    ),
    tts=deepgram.TTS(
        model="aura-2-thalia-en",
        sample_rate=16000
    )
)

​
Configuration Tips

​
For Fastest Response

stt = deepgram.STT(
    model="flux-general-en",
    eager_turn_detection=True  # Detect turn end quickly
)

tts = deepgram.TTS(
    model="aura-2-thalia-en",
    sample_rate=16000  # Lower sample rate for speed
)

​
For Best Quality

stt = deepgram.STT(
    model="flux-general-en",
    eager_turn_detection=False  # Wait for complete utterance
)

tts = deepgram.TTS(
    model="aura-2-thalia-en",
    sample_rate=24000  # Higher sample rate
)

​
Features

​
Speech-to-Text (STT)

  • Built-in turn detection
  • Real-time streaming transcription
  • High accuracy with Flux models
  • Automatic punctuation and formatting

​
Text-to-Speech (TTS)

  • Low-latency WebSocket streaming
  • Natural-sounding voices
  • Multiple voice options
  • Configurable sample rates

​
Environment Variables

DEEPGRAM_API_KEY=your_deepgram_api_key_here

​
Technical Details

​
STT Implementation

  • Uses Deepgram’s real-time streaming API
  • WebSocket connection for low latency
  • Automatic reconnection handling
  • Built-in silence detection for turn taking

​
TTS Implementation

  • WebSocket streaming for fast audio delivery
  • Chunked audio output for immediate playback
  • Configurable sample rates (8000-48000 Hz)
  • Automatic audio format conversion

​
References

Build docs developers (and LLMs) love

Get started for freeTalk to us