For the lexing (lexical analysis) stage of the parser. I know that the alternative to using a lexer (for which the tokens are specified by regular languages, thus lexed by FSMs) is to use so-called scannerless parsing (this is possible since every regular-language is context-free, thus the tokens can also be specified by the original context-free grammar). But using this alternative is usually much more painful since you usually want the grammar to have a nice property that allows linear-time parsing and a unique parse tree. LR(k), LL(k) etc. are such properties. Such properties are much more difficult to ensure for scannerless parsing.
I usually write it as a small (256 byte) lookup table where each entry tells you how many characters to skip next. If you don't use a lookup table, its 4 single-line `if` statements (and if you do it's a one liner, plus however many lines the table takes).
https://www.youtube.com/watch?v=HxaD_trXwRE
It supports unicode.
There are a lot of reasons for this, but one of the basic ones is that the lexer does not need to interact in a complex way with the compiler's state. It is a relatively simple pipeline where characters go in one end and tokens come out the other.
I wrote the parser for my script language, jerboa ( https://github.com/FeepingCreature/jerboa ) over a few days; it's (I thought) entirely standard recursive descent single-pass. This is how I've always written parsers. Where would you even use FSMs for parsers?