Detect speech turn completion for conversation flow control

Use This When

• Building conversational AI systems that need to know when speaker finishes utterance
• Implementing turn-taking logic for multi-party dialogue systems
• Creating voice assistants that should respond only after user completes thought
• Gating LLM processing to complete utterances to improve response relevance

What It Does

• Predicts the probability (0.0-1.0) that a speech turn is complete using a trained endpoint model
• Uses Smart Turn v3 (ONNX + Whisper features) with an 8-second audio window (pads/truncates to last 8s)
• Resamples audio to 16kHz and normalizes to [-1, 1] range
• Returns completion probability so downstream logic can apply its own thresholding / smoothing
• Processes each frame independently for real-time conversation flow decisions
• Auto-selects an ONNX file from the model repo by default (override via onnx_filename if needed)
• Supports selecting execution device via device ('auto', 'cuda', or 'cpu')

Works Best With

• detect-voice-activity → this component → query-llm for conversational AI pipelines
• Speech recognition → this component → determine when to send transcript to LLM
• Integration with chatbot components to trigger response generation at natural pauses
• Voice UI systems that need to distinguish partial from complete user commands

Caveats

• Model trained on specific speaker patterns; accuracy varies with accents and speaking styles
• Sample rate mismatch requires resampling which adds latency; configure appropriately
• Natural pauses may appear as turn endings depending on how you threshold/smooth the probability
• Normalization assumes calibrated audio levels; extremely quiet speech may fail classification