Smart-Turn-Topo

The end-of-turn judge of Topo on an iPhone: Daily's Smart Turn v3.2 as a Core ML ML Program. It is not a new model or a finetune.

The weights are pipecat's fp32 model, read out of smart-turn-v3.2-gpu.onnx at f766f81d3cfdf7737ac64aad813d91bbfd56bf93 into a PyTorch rebuild of SmartTurnV3Model (whisper-tiny encoder, attention pooling, classifier, sigmoid inside), traced and converted with coremltools 9.0 to an FP16 ML Program for iOS 17. It is not the int8 smart-turn-v3.2-cpu.onnx pipecat also publishes, which scores up to 0.40 away from the fp32 model on the same features.

What is here

smart-turn-v3.2-fp16.mlpackage/, compiled on the device with MLModel.compileModel(at:):

file	sha256
Manifest.json	4af5579bfb2fa82aa7914662d1708ac75fc9ae1824639578be11cbf4b8882a54
Data/com.apple.CoreML/model.mlmodel	d178340c5e7e8ad6458a117c31415298cf2824bfbb61a5dc0a6cd83d291592a7
Data/com.apple.CoreML/weights/weight.bin	ae5c3eb8be60b25320fb2fbfa0914c13f02b9d71b3fedfaad59d691a8a914446

• Input input_features, float32 [1, 80, 800]: Whisper log-mel features of the last 8 s of 16 kHz audio, zero-padded at the front (pipecat's whisper_features.py).
• Output logits, float32 [1, 1]: the probability the turn is complete (the sigmoid is in the model).

Against pipecat's fp32 ONNX on 21 clips: within 0.05, and no clip on the other side of 0.85. The recipe and the measurements are smart-turn-coreml in samdu/experiments (convert.py at c8d4467). A rerun of the conversion does not reproduce Manifest.json byte for byte, so the files here are the package and the script is how it was made.

Licence

BSD 2-Clause, as Smart Turn is: Copyright (c) 2024–2025, Daily. The licence text is in LICENSE.