We do not (need to) assume anything on the margins. Generally, in a mixture model setting we start with a single density $f(\cdot;\theta)$ and then we assume that we have random variables $(X_1, ..., X_N, Z_1, ..., Z_N)$ such that $(X_i, Z_i)$ are independent and for a parameter $\Theta = (\tau_1, ..., \tau_K, \theta_1, ..., \theta_k)$ we have that $X = (X_1, ..., X_N)$ and $Z = (Z_1, ..., Z_N)$ have a common density $p(x,z)$ and that
$$p(x|z) = \prod_{i=1}^N p(x_i|z_i) = \prod_{i=1}^N f(x_i;\theta_{z_i})$$
and
$$p(z) = \prod_{i=1}^N p(z_i) = \prod_{i=1}^N \tau_{z_i}$$
Let us assume that $N=1$ then we simply compute
$$p(x) = \int_{\mathcal{Z}} p(x,z) dz = \int_{\mathcal{Z}} p(x|z)p(z) dz = \sum_{k=1}^K f(x;\theta_{k}) \tau_k$$
i.e. the marginal of each and every data variable $X_i$ is the same and it is this mixture expression that you have written above and not a single Gaussian!