SS
Back to Projects

Speech Emotion Recognition

Built a real-time speech emotion recognition pipeline combining audio preprocessing, MFCC feature extraction, and hybrid CNN-RNN deep learning models for affective computing.

TensorFlowPyTorchLibrosaMFCCCNNRNNReal-time Inference

SYSTEM ARCHITECTURE

SYSTEM ARCHITECTUREAudio preprocessing + neural emotion classificationAudio InputSpeech waveformPreprocessingCleaning + augmentationMFCC / FeaturesTime-frequency representationCNNSpatial feature extractionBiLSTMTemporal modellingEmotion PredictionEmotion classesPublished research focuses on data augmentation and lightweight CNN-based speech emotion recognition.

Published research focuses on data augmentation and lightweight CNN-based speech emotion recognition.

Problem

Human emotional expression in speech varies significantly across speakers, recording conditions, and languages. The challenge was to build a robust pipeline that generalizes across these variations while maintaining real-time inference performance.

Architecture

Modular pipeline: audio acquisition → preprocessing (normalization, silence removal) → MFCC feature extraction → hybrid CNN-RNN classifier → real-time emotion prediction. Separated signal processing from model inference for maintainability.

Challenges

  • Speaker variability and recording quality differences
  • Class imbalance across emotion categories
  • Background noise robustness
  • Variable audio duration handling
  • Real-time inference latency constraints
  • Generalization across speakers and datasets

Benchmarks

Evaluated on RAVDESS and CREMA-D datasets. Achieved competitive accuracy across 7-8 emotion classes with sub-100ms inference latency for real-time applications.

Lessons Learned

  • High-quality preprocessing matters more than model complexity for speech
  • MFCCs provide strong perceptual features with low dimensionality
  • Hybrid CNN-RNN captures both local patterns and temporal dependencies
  • Consistent train/inference pipelines are critical for deployment