1 · Provide a voice sample
Speak naturally for 3–10 seconds. Longer clips give steadier estimates.
Privacy by architecture
Nothing leaves your device
Audio is decoded in memory, converted to spectral features, and scored by a ~1 MB neural network running in WebAssembly inside this tab. There is no server, no API call, no telemetry on your voice.
Model runs locally
The ONNX model is downloaded once, then served from your browser cache. You can verify this in DevTools → Network: after first load, audio requests show (disk cache) and zero uploads.
Nothing is stored
Recordings live only in JavaScript memory for the duration of the session. Close the tab and they are gone. No cookies, no accounts, no history.
Verifiable
Svara is open source. Read the exact feature code, model architecture, training procedure, and this privacy claim in the repository — and audit the network traffic yourself.
Ethics & limitations
- Not diagnostic. Speech-emotion estimates are not validated clinical instruments and must not drive medical, hiring, insurance, or legal decisions.
- Acted training data. The model is trained on CREMA-D, where actors portray emotions. Real emotional expression is subtler, mixed, and culturally shaped — accuracy in the wild is lower than lab numbers.
- Bias. CREMA-D is 91 American English speakers. Performance may differ across accents, languages, ages, and vocal conditions not represented in training.
- Content, not identity. The model reads prosody and spectral shape; it is not a speaker-identification or lie-detection system, and using it as one would be a misuse.
- Consent. Only analyze voices you have permission to analyze.
Model card (summary)
| Task | 6-class speech emotion recognition (anger, disgust, fear, happy, neutral, sad) |
|---|---|
| Dataset | CREMA-D — 7,442 clips, 91 actors, speaker-disjoint splits |
| Architecture | Compact CNN on 64-band log-mel, 2 s windows, ~0.3 M params |
| Deployment | Quantized ONNX (int8), onnxruntime-web (WASM), fully client-side |
| Intended use | Research, education, self-reflection |
| Out of scope | Clinical diagnosis, surveillance, deception detection |