I did plenty of graduate work with speech recognition; and detecting cognitive states based on speech is not much better than a coin-flip. Depression, drunkenness, deceit, happiness, etc. are hard to detect for an arbitrary person with accuracy. Speech signals for latent states in human behavior are highly idiosyncratic. The answer I always give when asked if something can be detected by speech (most requested - lying): "Yes, but only if you have a lot of time and want it to be accurate for at most a couple people. And it's going to be expensive."