Decoding speech imagery from brain signals potentially assists individuals with speech impairments. However, limited data and complex brain activity in recent studies have made accurate decoding challenging. We analyzed both the spectral (frequency) and spatial (location) aspects of brain activity to enhance decoding accuracy in small datasets. To our knowledge, previous studies have mainly used generated features without adequately considering spectro-spatial aspects. We trained machine learning with time-frequency representation (TFR) features using a public dataset from the Brain-Computer Interface (BCI) competition (BCI-DB) and our own recordings (PrimAudio-DB). The results showed prominent feature patterns of speech imagery in the frontal region and Gamma band, achieving accuracies of 98.6% in BCI-DB (exceeding benchmarks) and 81.7% in PrimAudio-DB. Moreover, our analysis revealed insights regarding speech differences (language and semantics). This study contributed to non-invasive speech imagery decoding and offered valuable insights for future speech rehabilitation and assistive technologies. © 2013 IEEE.