vision-agents-plugins-assemblyai 0.4.4

AssemblyAI streaming STT integration for Vision Agents

AssemblyAI Plugin

Streaming Speech-to-Text (STT) plugin for Vision Agents using AssemblyAI's Universal-3 Pro model.

Features

• Real-time streaming transcription via async WebSocket
• Built-in punctuation-based turn detection with configurable silence thresholds
• Streaming diarization — identify speakers in real time
• Native SpeechStarted event support
• Custom prompt and keyterms boosting support
• Sub-300ms time to complete transcript latency
• Built-in reconnection with exponential backoff

Installation

uv add "vision-agents[assemblyai]"
# or directly
uv add vision-agents-plugins-assemblyai

Usage

from vision_agents.plugins import assemblyai

stt = assemblyai.STT(
    speech_model="u3-rt-pro",  # Default model
    sample_rate=16000,
)

With streaming diarization

Enable speaker_labels to identify speakers in a mixed audio stream. Each transcript event will carry a distinct participant per speaker and the raw label in response.other["speaker_label"].

stt = assemblyai.STT(
    speaker_labels=True,
    max_speakers=2,  # optional hint (1-10)
)

With keyterms boosting

stt = assemblyai.STT(
    keyterms_prompt=["AssemblyAI", "Vision Agents"],
)

With custom turn silence thresholds

stt = assemblyai.STT(
    min_turn_silence=100,   # ms before speculative EOT check
    max_turn_silence=1200,  # ms before forcing turn end
)

Configuration

Parameter	Description	Default
api_key	AssemblyAI API key (falls back to ASSEMBLYAI_API_KEY env var)	None
speech_model	Model identifier	"u3-rt-pro"
sample_rate	Audio sample rate in Hz	16000
min_turn_silence	Silence (ms) before speculative end-of-turn check	API default
max_turn_silence	Maximum silence (ms) before forcing turn end	API default
prompt	Custom transcription prompt (cannot be combined with keyterms_prompt)	None
keyterms_prompt	List of terms to boost recognition for (cannot be combined with prompt)	None
speaker_labels	Enable streaming diarization for multi-speaker identification	False
max_speakers	Hint for expected number of speakers, 1-10 (requires speaker_labels=True)	None
max_reconnect_attempts	Maximum reconnect attempts on transient failures	3
reconnect_backoff_initial_s	Initial backoff delay in seconds	0.5
reconnect_backoff_max_s	Maximum backoff delay in seconds	4.0

Environment Variables

Set ASSEMBLYAI_API_KEY in your environment or pass api_key to the constructor.

Dependencies

• aiohttp>=3.9.0
• vision-agents

Docs

• https://www.assemblyai.com/docs/streaming/universal-3-pro
• https://www.assemblyai.com/docs/streaming/streaming-diarization-and-multichannel