Serveurperso/Smart-Turn-v3-GGUF

Smart Turn v3 GGUF

GGUF weights of Smart Turn v3.2 (Pipecat) for s2s.cpp, a C++17/GGML voice assistant. The model reads the last 8 seconds of the audio stream and tells whether the speaker has finished the turn, from the waveform alone, no transcript needed: a pause to think does not hand over the floor.

Files

file	size
smart-turn-v3.2-F32.gguf	32 MB

Whisper Tiny encoder (4 layers, width 384) on 80 mel bins, attention pooling and a small classifier, 8M parameters. The model runs once per speech to silence boundary, so F32 is the only variant.

Layout

convert.py builds the file from the FP32 GPU export smart-turn-v3.2-gpu.onnx of pipecat-ai/smart-turn-v3, every dimension a turn.* metadata key. The layout is the one src/smart-turn.cpp reads through its C ABI, it does not load in llama.cpp.

Quick start

git clone https://github.com/ServeurpersoCom/s2s.cpp.git
cd s2s.cpp
git submodule update --init
./buildcuda.sh
./models.sh      # this file and the rest of the pipeline -> models/
./server.sh      # then open http://localhost:8088

s2s.cpp runs Smart Turn on the CPU, on a sliding window of the stream rather than the current turn alone, as the upstream reference does.

Parity

Against onnxruntime on the upstream ONNX, over the example recording: completion probabilities at cosine similarity 1.000000000 (max abs 2.1e-6), decisions 17/17.

License

Upstream model: Smart Turn v3 by Pipecat (Daily), BSD 2-Clause

GGUF tooling: s2s.cpp, MIT