DEEP LEARNING
EmotiVoice AI — Speech Emotion Recognition
LSTM-based system that classifies 8 emotions from recorded speech, served from Hugging Face Spaces behind a Next.js analytics dashboard.
Public demo
8 — emotion classes
The problem
Existing speech emotion tools are black boxes — they return a label with no confidence breakdown or per-emotion trend, making them unusable for UX research or mental-health monitoring applications.
The approach
A 7-layer LSTM architecture (2.4M params) trained on merged speech corpora, exported to HuggingFace Spaces as a REST API. The Next.js frontend uses the Web Audio API to record 16kHz audio, streams it to the model, and renders a Grafana-style analytics dashboard with per-emotion time series.
The results
Classifies 8 emotion categories (happiness, sadness, anger, fear, disgust, surprise, neutral, pleasant surprise) with per-emotion confidence and a time-series dashboard. The repository README reports 94.2% weighted accuracy on the RAVDESS test set and an earlier version of this page said 85–92%; neither figure is backed by a committed evaluation file, so neither is used as a headline here.
Stack: Next.js 15, TypeScript, LSTM, HuggingFace Spaces, Web Audio API, Recharts