Speech Emotion Recognition Beyond Benchmarks: Robustness, Efficiency, and Evaluation for Real-World Deployment

Robust Speech Emotion Recognition (SER) for edge applications requires evidence beyond within-corpus accuracy. We propose a three-tier protocol separating speaker, corpus, and deployment-proxy shifts rather than a new classifier. It compares Wav2Vec 2.0 Base, WavLM Base, Whisper Base, Whisper Tiny, and ReDimNet-B2 on a five-corpus composite training set. Models are selected on speaker-independent validation with unseen source-corpus speakers, then tested on unseen IEMOCAP. Whisper Tiny is also tested on 108 smartphone recordings by 9 non-professional speakers. Speaker-independent UAR ranges from 0.752 to 0.786 but falls to 0.475 to 0.531 on IEMOCAP. Whisper Tiny and ReDimNet-B2 remain close to larger Base models despite smaller encoders, while the smartphone test reveals class-dependent failures for happy and angry speech. The protocol structures stress testing for compact SER models and shows parameter count alone is an insufficient proxy for robustness or deployability.