On-device Voice Stress Detection
On-device voice-stress detection that fuses two Snapdragon NPU models: WavLM (how you speak) and Whisper (what you say), into one confident reading. 100% local: no cloud, no recording, no LLM. Sustained stress triggers a private, calming nudge.
Core trick
Fuses WavLM audio-stress scores with a Whisper-transcript text classifier through a small perceptron, running on the Hexagon NPU.
Description
$28
README excerpt
# Slashh — On-device Voice Stress Detection **Slashh** notices when your voice is under stress and offers a calming nudge — **entirely on your phone**, in real time, with **no network, no recording, and no LLM**. It runs **two neural models on the Snapdragon Hexagon NPU** and fuses them into one confident reading: - 🎙️ **WavLM** — an audio stress model that listens to *how* you speak. - 📝 **Whisper-Tiny** — transcribes *what* you say; a trained text classifier scores the words. - 🔗 A small **fusion perceptron** combines the two into a single stress probability. Built for the **[Qualcomm × Meta ExecuTorch Hackathon](https://lablab.ai/ai-hackathons/qualcomm-x-meta-executorch-hackathon)** (San Francisco, June 27–28 2026) on a **Samsung Galaxy S25 Ultra** (Snapdragon 8 Elite). Everything in the inference hot path runs locally; the app declares **no `INTERNET` permission** and works in airplane mode. > **Want the full story, deep-dive architecture, NPU runtime details, model training, and the > research log of everything we tried?** → **[`docs/`](docs/index.md)**  --- ## Table of contents 1. [What it does](#what-it-does) 2. [How it works (current architecture)](#how-it-works-current-architecture) 3. [Team](#team) 4. [Prerequisites](#prerequisites) 5. [Part A — Build & install the app](#part-a--build--install-the-app) 6. [Part B — Bring the two models onto the NPU](#part-b--bring-the-two-models-onto-the-npu) 7. [Verify it's worki