
Slashh turns your phone into a private, always-on companion that notices when your voice is under stress and offers a calming nudge — entirely on-device, in airplane mode, with no recording and no LLM. Most stress detectors look at only one signal. Slashh is multimodal: it runs two transformer models on the Snapdragon Hexagon NPU at the same time. WavLM listens to how you speak (vocal tension and arousal), while Whisper-Tiny transcribes what you say and a tiny Dreaddit-trained classifier scores the words for stress. A small fusion perceptron combines the two into one confident probability — far more reliable than either signal alone — and only raises a gentle relief nudge when stress genuinely sustains. Everything in the inference loop runs locally via PyTorch ExecuTorch and QNN context binaries; the app declares no INTERNET permission, so audio, transcripts, and scores never leave the device. Built and verified live on a Galaxy S25 Ultra (Snapdragon 8 Elite): two neural models on the NPU, a responsive energy fallback, and real-time relief — fully private by construction.
28 Jun 2026