New DCRF-BiLSTM model hits 100% accuracy on speech emotion datasets
Achieves perfect scores on TESS and EmoDB, 93.76% across five benchmarks.
A team of researchers from the University of Louisiana at Lafayette and other institutions has introduced a novel hybrid deep learning architecture called DCRF-BiLSTM for speech emotion recognition (SER). The model combines deep convolutional recurrent features with bidirectional long short-term memory networks to classify seven emotional states: neutral, happy, sad, angry, fear, disgust, and surprise.
The DCRF-BiLSTM was evaluated on five benchmark datasets—RAVDESS, TESS, SAVEE, EmoDB, and Crema-D—individually and in combination. On single datasets, it achieved near-perfect or perfect scores: 97.83% on RAVDESS, 97.02% on SAVEE, 95.10% on Crema-D, and 100% on both TESS and EmoDB. On the combined R+T+S subset, accuracy reached 98.82%, outperforming previous state-of-the-art results.
Notably, this is the first study to evaluate a single SER model across all five datasets simultaneously, achieving a robust 93.76% overall accuracy. The results confirm the model's generalizability across languages, recording conditions, and speaker demographics—a critical step toward deploying SER in real-world human-computer interaction systems.
- DCRF-BiLSTM achieves 100% accuracy on TESS and EmoDB datasets, and 97.83% on RAVDESS.
- First SER model evaluated simultaneously on all five major benchmarks (RAVDESS, TESS, SAVEE, EmoDB, Crema-D), scoring 93.76% overall.
- Outperforms prior work on the combined R+T+S dataset with 98.82% accuracy.
Why It Matters
Enables more reliable emotion-aware AI, improving human-computer interaction and applications in mental health, call centers, and assistive tech.