To ensure the

user_speech_committed

event fires only once per complete user utterance with the final transcript, you can follow these steps: • State Management: The system uses a

_user_turn_committed

flag to prevent duplicate processing. Ensure this flag is properly managed in your implementation. • End-of-Turn Detection: Use a multi-layered approach that coordinates VAD (Silero), STT (Deepgram), and turn detection. This includes: ◦ Setting appropriate

min_endpointing_delay

and

max_endpointing_delay

values to control the timing of turn commitment. ◦ Using turn detection models to predict end-of-turn probability. • Final Transcript Handling: Ensure that the system processes the

FINAL_TRANSCRIPT

event from Deepgram STT. This event should trigger the final transcript commitment logic, which includes: ◦ Flushing the STT buffer if the last final transcript was received more than 0.5 seconds ago. ◦ Waiting for the final transcript to be available before committing the turn. • Configuration: Optimize your configuration for reliable performance: ◦ Set VAD parameters such as

min_speech_duration

and

min_silence_duration

to appropriate values. ◦ Ensure endpointing delays are tuned to your use case. • State Reset: After committing a user turn, clear the transcript and reset the

_user_turn_committed

flag to prepare for the next utterance. This approach ensures that the

user_speech_committed

event fires exactly once per complete user utterance, capturing the final transcript without duplicates. Sources: livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/stt.py | Speech & Audio Processing | livekit/livekit_composite | DeepWiki | Speech-to-text (STT) integrations | LiveKit Docs