Speech Emotion Recognition Using Feature Selection with Adaptive Structure Learning

On creating small datasets for training embedded acoustic scene classification systems via time-frequency segmentation

“…In these applications, the strict requirements for energy consumption, memory usage, latency and data storage play a crucial role in the feasibility of embedding emerging technologies [9], [10]. In this regard, the research on speech recognition systems is considerably ahead of the one in ASC, as deep-learning models [8], [9], learning frameworks [11], and datasets [12] that make the embedding of these models possible have been already proposed years ago. By taking as reference the path followed by the speech recognition research, and considering current trends in embedded machine learning models, one could identify some topics that should be addressed in order to make small ASC systems viable.…”

Peer Review #1 of "Effect on speech emotion classification of a feature selection approach using a convolutional neural network (v0.1)"

“…In (Sezgin et al (2012)), several distinguishing acoustic features were used to identify emotions: spectral, qualitative, continuous, and Teager energy operator-based (TEO) features. Thus, many researchers have suggested that the feature set comprises more speech emotion information (Rayaluru et al (2019)). However, combining feature sets complicates the learning process and enhances the possibility of overfitting.…”