nix-config/home/linux/gui/base/vinput/README.md

Fcitx5 Vinput

On ai and shoukei, install a Sherpa-ONNX model in Vinput under Resources → Models, then select it in Control → ASR Providers. Hold and release right Alt to dictate. Right Shift toggles Chinese/English; left Shift remains a normal modifier. Open Vinput from the application launcher for settings.

Models are stored in ~/.local/share/vinput/models/sherpa-onnx/. Review their source and license before installing.

Backends

• shoukei uses CPU. Leave the model's provider at its default.
• ai uses OpenVINO when the model's vinput-model.json sets "provider": "openvino"; restart vinput-daemon after changing it. The user needs the render group and a new login after deployment.

NPU execution status: unverified. On ai, the OpenVINO runtime and NPU plugin load, the daemon opens /dev/accel/accel0, and the bundled WAV sample transcribes. However, the NPU busy-time counter did not increase, and no NPU profiling data has been collected. These observations do not confirm that model computation ran on the NPU; CPU fallback or partial fallback remains possible.

Diagnostics

journalctl --user -b -u vinput-daemon

Check the Sherpa/OpenVINO logs for the selected device. A definitive check needs NPU profiling enabled for inference and non-zero per-layer NPU timings; a loaded plugin or device busy-time counter alone is not conclusive.