New perspectives on likelihood-based inference for latent and observed Gaussian mixture models