In my experience a "signature" is something that puts additional constraints on a language made of syntactic terms. I'll try to provide an example.
Consider the alphabet $\Sigma = \{f,g,a,b,(,),;\}$. From this we can build the set of all the words $\Sigma^*$. This contains any word, including gibberish like "$)g;;f(($".
Then, we can craft the set of syntactic terms $T \subseteq \Sigma^*$ through an inductive definition.
We start by defining a subset $F \subseteq \Sigma$ as $F = \{f,g,a,b\}$, and we call this the set of function symbols. Then, inductively, we state that if $t_1,\ldots,t_n$ (with $n\geq 0$ are terms and $s\in F$, the word $s(t_1;\ldots;t_n)$ is a term.
Note that $n=0$ above gives the "base case" of the induction. We then have that $f()$ is a term, $a()$ is a term, $f(a();b())$ is a term, $f(f(a();b();a();a(b())))$ is a term, and so on.
Now, we might want to put more constraints on those terms. For instance, we might want to associate to each $s \in F$ a fixed nubmer of arguments, and consider only those terms respecting that. For instance, if we decide that $f$ has arity two (f is binary), then we allow $f(a();b())$ but not $f(a())$ nor $f(a();f())$. This restriction essentially defines a subset language $L\subseteq T$.
This is sometimes done by specifying a signature. E.g. something like
$$
\begin{array}{l}
f : A \times A \to A \\
g : A \to A \\
a : 1 \to A \\
b : 1 \to A \\
\end{array}
$$
The above is a fancy way to state that $f$ is binary, $g$ is unary, and $a,b$ are both constants (= zero-arity function symbols). $1$ is used to mean no-arguments.
A possible term here is $f(g(a());b())$.
The name $A$ above is called a "sort", and has no deep meaning on its own. The above is called a single-sorted signature since it uses only one sort. By contrast
$$
\begin{array}{l}
f : A \times B \to A \\
g : A \to B \\
a : 1 \to A \\
b : 1 \to A \\
\end{array}
$$
uses two sorts. This now forbids $f(g(a());b())$ since $g(a)$ has not sort $A$ as required. Instead, $f(b();g(a()))$ is still allowed.
When terms are logical formulae, it is common to use signatures to keep our words meaningful. For instance
$$
\begin{array}{l}
\land : P \times P \to P \\
{\sf prime} : \mathbb N \to P \\
0 : 1 \to \mathbb N \\
s : \mathbb N \to \mathbb N \\
\end{array}
$$
allows $\land({\sf prime}(0());{\sf prime}(s(0())))$ (i.e. ${\sf prime}(0) \land {\sf prime}(s(0))$, in more casual notation), but disallows nonsense such as
$\land({\sf prime}(0());0())$ (i.e. ${\sf prime}(0) \land 0$) which tries to use term $0()$ as if it were a proposition (a term of sort $P$).
