This is always possible. I will first explain the local construction, and then sketch a proof of how to globalise.
Local problem. Let $A \to B$ be a finite map of finite type $k$-algebras. Does there exist a finite map $A' \to B'$ of smooth $k$-algebras admitting a surjection to $A \to B$?
Construction. Suppose $B = A[b_1,\ldots,b_r]$; then each $b_i$ satisfies some monic minimal polynomial $f_i \in A[x_i]$. This gives a surjection
$$A[x_1,\ldots,x_r]/(f_1,\ldots,f_r) \twoheadrightarrow B.$$
This map need not be an isomorphism in general, but the result for $A[x_1,\ldots,x_r]/(f_1,\ldots,f_r)$ implies that for $B$. Thus, we may assume $B = A[x_1,\ldots,x_r]/(f_1,\ldots,f_r)$.
Now write $f_i = \sum_j a_{ij}x_i^j$ for some $a_{ij} \in A$. Let $\phi \colon k[y_1,\ldots,y_s] \twoheadrightarrow A$ be any presentation, and adjoin extra variables corresponding to the $a_{i0}$:
\begin{align*}
A'=k[y_1,\ldots,y_s,z_1,\ldots,z_r] &\twoheadrightarrow A\\
y_i &\mapsto \phi(y_i),\\
z_i &\mapsto a_{i0}.
\end{align*}
Choose lifts $a'_{ij} \in A'$ of the $a_{ij} \in A$ such that $a'_{i0} = z_i$ and such that $g_i = \sum_j a'_{ij}x_i^j$ is again monic. Let $B'$ be the finite extension $A'[x_1,\ldots,x_r]/(g_1,\ldots,g_r)$. Then $B'$ is the polynomial algebra
$$B' \cong k[y_1,\ldots,y_s,x_1,\ldots,x_r],$$
since we may eliminate $z_i = a'_{i0} = -\sum_{j > 0} a'_{ij}x_i^j$ from the equation $g_i$. Moreover, we now actually have a pullback:
$$B \cong A \underset{A'}\otimes B',$$
by construction of the $g_i$. $\square$
Global version. Recall: if $\mathscr E$ is a locally free sheaf on a scheme $X$, then the (geometric) vector bundle $\mathbb V(\mathscr E)$ of $\mathscr E$ is $\operatorname{\underline{Spec}}_X(\operatorname{Sym}^*(\mathscr E^\vee))$, viewed as scheme over $X$ through the map $\pi \colon \mathbb V(\mathscr E) \to X$. This gives an equivalence between locally free sheaves and geometric vector bundles; see e.g. [Stacks, Tag 062N]¹. If $\mathscr E = \mathscr L$ is a line bundle, we have
$$\pi_*\pi^* \mathscr L \cong \bigoplus_{i \geq 0} \mathscr L^{1-i},$$
so the section $1 \in H^0(X,\mathscr L^0)$ gives a canonical section $x_\mathscr L \in H^0(\mathbb V(\mathscr L), \pi^*\mathscr L)$.
Construction. Let $\mathcal O(1)$ be a very ample sheaf on $Y$, and let $\mathscr B = f_* \mathcal O_X$. Then there exists a surjection $\mathcal O(-a)^r \twoheadrightarrow \mathscr B$ for some $a \gg 0$ (i.e. $\mathscr B(a)$ is globally generated for some $a$). Replacing $\mathcal O(1)$ by $\mathcal O(a)$, we may assume $a = 1$ (for convenience).
Set $\mathscr E = \mathcal O(-1)^r$. The surjection $\mathscr E \twoheadrightarrow \mathscr B$ induces a surjection $\operatorname{Sym}^*(\mathscr E) \twoheadrightarrow \mathscr B$ by the ring structure on $\mathscr B$. For each of the factors $\operatorname{Sym}^*(\mathcal O(-1)) \to \mathscr B$, the canonical section $x_i$ satisfies some polynomial equation
$$f_i = \sum_{j=0}^{n_i} a_{ij} x_i^j = 0,$$
for some $a_{ij} \in H^0(Y,\mathcal O(n_i-j))$. Again, we may replace $\mathscr B$ by $\operatorname{Sym}^*(\mathscr E)/(f_1,\ldots,f_r)$.
On the other hand, we get a locally closed immersion $Y \hookrightarrow \mathbb P^N_k$ (say given by our very ample line bundle $\mathcal O(1)$). If $\mathscr F = \mathcal O(n_1) \oplus \ldots \oplus \mathcal O(n_r)$ on $\mathbb P^N$, then the section $(a_{10},\ldots,a_{r0}) \in H^0(Y,\mathscr F|_Y)$ induces a locally closed immersion
$$Y \hookrightarrow Y'=\mathbb V(\mathscr F) \twoheadrightarrow \mathbb P^N.$$
Consider the coherent $Y'$-algebra given by $\mathscr B' = \operatorname{Sym}^*(\mathcal O(-1)^r)/(g_1,\ldots,g_r)$, where
$$g_i = \sum_{j=0}^{n_i} a'_{ij}x_i^j$$
is a monic lift² such that $a'_{i0} = z_i$ (the canonical section $z_{\mathscr O(-1)}$ of $\pi^* \mathcal O(1)$ on $\mathbb V(\mathcal O(1))$). Then $X' = \operatorname{\underline{Spec}}_{Y'}(\mathscr B')$ is a smooth $k$-scheme, finite over $Y'$, such that
$$X = X' \underset{Y'}\times Y.$$
This finishes the construction in the global case. $\square$
References.
[Stacks] A.J de Jong et al, The stacks project.
¹The Stacks project (following EGA) does not put a dual in the definition of $\mathbb V(\mathscr E)$. The reason appears to be that duality is not well-behaved in the larger generality of quasi-coherent sheaves. However, the dual version is the 'geometrically correct' version, e.g. it is covariant, and sections of $\pi \colon \mathbb V(\mathscr E) \to X$ are in bijection with $H^0(X,\mathscr E)$.
²There are potential surjectivity issues on global sections, but these can easily be fixed by introducing variables lifting the $a_{ij}$ like we did for the $a_{i0}$.