🎙️ Voice Stress Detector
Say a sentence the way you feel it. The model listens to how you speak, not the words.
Your voice
Tap to record
Speak for 2–10 seconds, then tap again. Try “I can't believe this is happening” angrily, then calmly.
or
Vocal stress index
How it works & limitations
Your audio is converted to mono in the browser and sent to the model on Hugging Face Spaces. There, silence is trimmed and volume normalised. A frozen WavLM speech model turns the audio into an embedding, and a linear classifier trained on RAVDESS + CREMA-D (115 actors) outputs probabilities for 6 emotions: neutral, happy, sad, angry, fearful and disgust.
The stress index weights those probabilities by how strongly each emotion reflects tense, negative, high-arousal speech (angry and fearful count most, calm and neutral count zero).
- 75% accuracy is measured on actors the model never heard. A model trained on one dataset scored only 38% on another, so your microphone and room, effectively a new dataset, will lower accuracy.
- Trained on acted North-American English, so other accents, languages and noisy mics reduce accuracy.
- Stress is inferred, not measured. This is not a medical or HR tool.
- The free Space sleeps when idle. The first request may take about a minute while it wakes up.