Duyu-small

Turkish + English speech recognition that runs entirely in your browser: 19.3 M active parameters, ONNX int8 (56 MB) or 4-bit (34 MB), onnxruntime-web on your CPU. Your audio never leaves the device. It writes whatever is spoken, Turkish, English or both in one sentence, cased and punctuated.

Built for live, short speech next to the microphone; long files are cut into 25 s chunks and decoded one after another. Model, engine and benchmarks: KaanAydinli/Duyu-small (Apache-2.0).