Deep learning · CNN + LSTM

Speech Emotion Recognition

A hybrid CNN + LSTM model for real-time emotion detection from voice. Extracted MFCC audio features from voice datasets and tuned the model for strong classification performance across emotion categories.

PyTorchCNNLSTMAudioMFCC

Role: ML engineer
Timeline: 2023

View on GitHub ↗

Speech emotion classifier hero — waveform visualisation with predicted emotion 'Happy' at 82% confidence — Classifier UI — live prediction

Signal chain diagram from raw audio through MFCC, CNN, LSTM to softmax with 94.2% train and 88.6% validation accuracy — Signal chain — 94.2% / 88.6%

Problem

Detecting emotion from raw audio in real time — traditional classifiers on hand-crafted features plateaued in accuracy.

Approach

01Extracted MFCC features from open voice datasets (RAVDESS-style).
02Trained a hybrid CNN + LSTM: CNN for local spectral patterns, LSTM for temporal dynamics.
03Tuned dropout, learning rate, and window size; used class-balanced sampling.
04Wrapped inference in a lightweight Python service for real-time streaming input.

Stack

PyTorch
Librosa
NumPy
CNN
LSTM

Outcomes

→Strong classification performance across primary emotion categories.
→Sub-second inference on CPU for short audio clips.

View on GitHub ↗