BiLSTM-CNN Hyperparameter Optimization for Speech Emotion and Stress Recognition
Agustinus Bimo Gumelar, Eko Mulyanto Yuniarno, Derry Pramono Adi, Adri Gabriel Sooai, Indar Sugiarto, Mauridhi Hery Purnomo · 2021
The most automated speech recognition (ASR) systems are extremely complicated, integrating many approaches and requiring a high variety of tuning parameters. Deep understanding and experience of each component are required to achieve optimal performance in ASR, confining the development of ASR systems to the experts. Hyperparameters are crucial for machine learning algorithms because they directly regulate the behavior of training algorithms and have a major impact on model performance. As a result, developing an effective hyperparameter optimization technique to optimize any given machine learning method would considerably increase machine learning efficiency. This work investigates the use of Random Forest and Bayesian to automatically optimize BiLSTM-CNN systems. We built the ASR based on the BiLSTM-CNN model and customized its hyperparameters value to heed our low-hardware specification during optimization. Furthermore, we gathered 1,000 clips of speech data from various movies, classifying them according to emotion and stress classes. In pursuit of contextual-level understanding in our ASR, we transcribed our speech data and used the bigram textual feature. Our Random Forest-optimized BiLSTM-CNN model ultimately reaches 84% of accuracy result and learning runtime in under 17 seconds.