What is the purpose of speech recognition? To decide from which person a voice comes? Or to recognize speech, equally from which person? What must be detected? Certain persons, certain keywords, all sorts of words?

The problem is not only to recognize certain words, but to recognize them in all sorts of situaions. Means, your detection algorithm must be able to detect words, equally how fast they are spoken, equally how distinct they are spoken, equally how loud they are spoken, equally whether the person has a cold, is angry, is tired, is happy,...

So, by the help of an algorithm, which gains some kind of frequency spectrum in the range of frequencies, where people's voices differ (something between 300Hz and 3kHz) and which gains some information how this spectrum changes with time, you might be able to detect a word from time to time, assumed you speak as monotoneous as a robot. But to find an algorithm which is not vulnerable, equally how the words are spoken, seems to me to be well beyond the scope of a microcontroller. Keep in mind, that even PC based systems with their mighty pentiums are not free of making errors.

Kai