| title: Pipecat VAD + Smart Turn v3.2 Simulator | |
| emoji: 🎙️ | |
| colorFrom: indigo | |
| colorTo: purple | |
| sdk: gradio | |
| sdk_version: 6.19.0 | |
| app_file: app.py | |
| pinned: false | |
| # Pipecat pipeline simulator — VAD ➜ Smart Turn v3.2 | |
| Replays an audio clip through the **exact** logic pipecat runs in a live call: | |
| ``` | |
| audio chunks → Silero VAD → (after stop_secs of silence) → Smart Turn v3.2 | |
| ``` | |
| The Silero VAD state machine and the Smart Turn audio-buffering / segment | |
| extraction are faithfully ported from | |
| [pipecat](https://github.com/pipecat-ai/pipecat). Every time VAD detects the | |
| user stopped speaking, the buffered segment is handed to Smart Turn and the | |
| end-of-turn decision is reported — with the exact audio that was fed to the | |
| model playable per turn. | |
| All VAD and Smart Turn parameters are editable in the UI: | |
| - **VAD:** `confidence`, `start_secs`, `stop_secs`, `min_volume` | |
| - **Smart Turn:** `pre_speech_ms`, `max_duration_secs`, `stop_secs` fallback, decision `threshold` | |
| - **Simulation:** chunk size, trailing silence | |
| Inference (Silero VAD + Smart Turn ONNX models) runs on the Space CPU via | |
| `onnxruntime`. | |