How Our AI Detects Emotional Patterns in Your Voice

When you record a journal entry, a lot happens in the background. Your audio is transcribed, parsed, and analyzed across multiple dimensions simultaneously. Here's a transparent look at how it all works.

Step 1: Transcription

We use a fine-tuned Whisper model to convert your speech to text with high accuracy, even in noisy environments. We preserve filler words ("um", "like", "you know") because they're emotionally meaningful — high filler frequency often correlates with uncertainty or stress.

Step 2: Sentiment & Emotion Classification

The transcript is passed through a multi-label classifier trained on 2.4 million annotated emotional text samples. It detects not just positive/negative sentiment, but 27 discrete emotional states including anticipation, contempt, nostalgia, and overwhelm.

Step 3: Temporal Pattern Detection

Single entries are useful, but the real power is in trends. Our system maintains a rolling emotional timeline per user. It uses LSTM-based sequence modeling to identify recurring patterns — like mood dips every Monday or elevated stress before major events.

Privacy First

All processing happens on-device for the initial transcription step. Emotional analysis is done server-side on anonymized text only — your audio is never stored after transcription.

"The goal isn't to label you — it's to help you see yourself more clearly."