• speech is a two-dimensional process with axes in frequency and time
  
  

"cut the yards and paint the houses"

• correlation exists across frequency bins: if bin 'i' has high energy, it is likely that bin 'i+1' and 'i-1' will also have high energy - this can be learned but requires significant amounts of data
  
• correlation exists across time instances - consecutive observations are correlated
  
• we want a system that takes correlations in both dimensions into account