EmoSense uses a deep learning ensemble — CNN-CBAM, BiLSTM, and Transformer models — to classify speech into 4 emotions.
Speech Emotion Recognition (SER) is the task of automatically identifying human emotional states from audio signals. EmoSense combines multiple deep learning architectures trained on a merged, balanced dataset.
The system processes raw audio through two parallel feature extraction pipelines — Mel spectrograms and MFCCs — feeding into an ensemble of 5 specialized models whose predictions are combined via weighted soft voting.
Detected Emotions
Joyful, excited speech patterns
Low energy, slow cadence
High intensity, sharp peaks
Balanced, flat intonation
Upload an audio file or record live to see real emotion detection from the deployed model
Supports WAV, MP3, M4A, FLAC
Click to start recording
Detected Emotion
Confidence scores
Was this prediction correct?
Two parallel feature pipelines feed into a 5-model ensemble for maximum accuracy
Final prediction = average of all 5 models' probability outputs
Every prediction's audio is saved. When you confirm or correct a result, that label is stored for periodic retraining — improving the model over time using real user data.
Interested in the project, want to collaborate, or have questions? Reach out below.