codewithmoin/smart-turn-hinglish

Smart Turn Detection (whisper-tiny, 8s)

Audio-only endpointing for voice agents: has the speaker finished their turn, or just paused? Tuned for English + Hindi/Hinglish. ~7.8M params, runs from a single 31 MB ONNX file at ~38 ms/clip on CPU (mel baked into the graph, so it needs only onnxruntime + numpy).

Note: on the free ZeroGPU tier the first request warms a shared GPU ("waiting for GPU"), then runs. Inference itself is CPU-only ONNX, ~38 ms.

Record or upload speech, get P(turn complete). The threshold slider trades interruptions against latency (higher = more patient).

Code, training, ablations and the full report: see the GitHub repo.