🎙️ Voice Emotion Recognition
Speak for a few seconds. The model listens to how you say it, not what you say.
🔒 Runs entirely in your browser. Your audio never leaves your device.
Your voice
Loading model (~120 MB, cached after the first visit)…
🎧
WAITING…
Emotion scores
Emotion over time
How it works
1. Audio is converted to 16 kHz mono, silence is trimmed and the volume is normalised.2.
wav2vec2, a transformer pretrained on 960 hours of speech, turns the sound into features.3. A classifier on top gives a probability for each emotion.
4. For the timeline, a 2-second window slides across the clip 1 second at a time.
The model was exported from PyTorch to ONNX and compressed to 8-bit, so it runs here with
onnxruntime-web, with no server.