Skip to main content
Pipecat Smart Turn uses a Whisper-based audio model (~8 MB) to predict turn completion directly from speech audio. It detects prosodic cues — falling intonation, speech rate changes — that indicate a caller has finished speaking. Provider identifier: pipecat_smart_turn_eos

​
Source Location

api/assistant-api/internal/end_of_speech/internal/pipecat/
├── pipecat_end_of_speech.go   # Main implementation
├── turn_detector.go           # ONNX inference wrapper
├── models/
│   └── smart-turn-v3.2-cpu.onnx   # Downloaded at Docker build (~8 MB)

​
How It Works

  1. UserAudioPacket audio is accumulated in a rolling float32 buffer (max ~5 seconds at 16 kHz)
  2. When a final SpeechToTextPacket arrives, the model runs inference on the buffered audio
  3. The model outputs a turn-completion probability (0.0 – 1.0)
  4. If probability >= threshold → set timer to quick_timeout (caller is likely done)
  5. If probability < threshold → set timer to silence_timeout (caller is likely still speaking)
  6. Interim transcripts reset the timer to fallback_timeout
  7. When the timer fires, EndOfSpeechPacket is emitted
Audio → buffer (rolling ~5s)
STT final → model inference → P(complete) = 0.73
  0.73 >= 0.5 (threshold) → timer = 200ms (quick)
  ...silence...
  Timer fires → EndOfSpeechPacket

​
Parameters

Option KeyDefaultRangeDescription
microphone.eos.threshold0.50.1–0.9Turn completion probability threshold
microphone.eos.quick_timeout250 ms50–1000 msSilence buffer when model says “done”
microphone.eos.silence_timeout2000 ms500–5000 msSilence duration when model says “still speaking”
microphone.eos.timeout500 ms500–4000 msFallback timeout for interim transcripts and inference failures

​
Model Setup

​
Docker

The model is downloaded from Hugging Face during the Docker build and patched for ONNX Runtime compatibility. No manual action required. 
# Set automatically in the Dockerfile
PIPECAT_TURN_MODEL_PATH=./models/pipecat_turn/smart-turn-v3.2-cpu.onnx

​
From Source

Download the model manually: 
mkdir -p api/assistant-api/internal/end_of_speech/internal/pipecat/models

curl -fsSL -o api/assistant-api/internal/end_of_speech/internal/pipecat/models/smart-turn-v3.2-cpu.onnx \
    "https://huggingface.co/pipecat-ai/smart-turn-v3/resolve/main/smart-turn-v3.2-cpu.onnx"
If you encounter ONNX opset errors, patch the model: 
pip install onnx
python3 -c "
import onnx
m = onnx.load('api/assistant-api/internal/end_of_speech/internal/pipecat/models/smart-turn-v3.2-cpu.onnx')
m.ir_version = 9
used = {n.domain for n in m.graph.node}
keep = [o for o in m.opset_import if o.domain == '' or o.domain in used]
del m.opset_import[:]
m.opset_import.extend(keep)
onnx.save(m, 'api/assistant-api/internal/end_of_speech/internal/pipecat/models/smart-turn-v3.2-cpu.onnx')
"
To override the model path: 
export PIPECAT_TURN_MODEL_PATH=/path/to/smart-turn-v3.2-cpu.onnx
Requires ONNX Runtime (libonnxruntime) — same dependency as Silero/FireRed VAD.