Add Inworld STT provider to livekit-plugins-inworld

[cshape]

Summary

• Add streaming speech-to-text support to the existing livekit-plugins-inworld package via Inworld's WebSocket STT API (wss://api.inworld.ai/stt/v1/transcribe:streamBidirectional)
• Follows established plugin patterns (Soniox, Deepgram, Google) including reconnection loop with tasks_group cleanup, stream tracking, START/END_OF_SPEECH events, and request_id
• Pass-through model selection — any model string accepted (default: inworld/inworld-stt-1)
• Eager end-of-turn defaults (endOfTurnConfidenceThreshold=0.3, minEndOfTurnSilenceWhenConfident=200ms) for low-latency voice agent use
• Voice profile detection enabled by default — age, gender, emotion, accent data surfaced via SpeechData.metadata["voice_profile"]
• update_options() for runtime config changes
• Adds metadata: dict[str, Any] | None field to SpeechData in livekit-agents for plugin-specific data (backwards-compatible, defaults to None)

Changes

• livekit-agents/livekit/agents/stt/stt.py — Add optional metadata field to SpeechData for plugin-specific data
• livekit-plugins/livekit-plugins-inworld/livekit/plugins/inworld/stt.py — New STT + SpeechStream implementation
• livekit-plugins/livekit-plugins-inworld/livekit/plugins/inworld/__init__.py — Add STT/SpeechStream exports
• livekit-plugins/livekit-plugins-inworld/pyproject.toml — Update description/keywords for STT+TTS
• livekit-plugins/livekit-plugins-inworld/README.md — Add STT usage docs
• examples/voice_agents/inworld_agent.py — Voice agent using Inworld STT+TTS
• .gitignore — Add .env

Usage

from livekit.plugins import inworld

session = AgentSession(
    stt=inworld.STT(),                          # default: inworld/inworld-stt-1
    tts=inworld.TTS(voice="Clive"),
    llm="openai/gpt-4.1-mini",
)

# Access voice profile metadata
@session.on("user_input_transcribed")
def on_input(ev):
    print(ev.transcript)

# Override model or tune end-of-turn behavior
stt = inworld.STT(
    model="assemblyai/universal-streaming-multilingual",
    end_of_turn_confidence_threshold=0.5,
    min_end_of_turn_silence_when_confident=400,
)

Test plan

• Verified imports with SDK 1.5.2
• Tested streaming transcription end-to-end with 26s audio file (interim + final transcripts)
• Tested full agent workflow via console mode (Inworld STT + TTS + OpenAI LLM)
• Verified update_options() works for runtime model changes
• Verified no secrets in committed files
• ruff check and ruff format pass
• mypy type checks pass (resolved all 11 prior errors)
• CI checks (framework change to SpeechData may need separate PR if CI runs cross-package)

[tinalenguyen]

left a few comments, i also think the example would be the most useful and accessible in the plugin README