The paper presents a method to estimate frequencies of a multi-sinusoidal signal in finite time. Expressing the signal via delayed measurements, we transform this nonlinear problem into a linear regression, where unknown parameters depend on signal frequencies. Dynamic Regressor Extension and Mixing method is used to replace nth order regression model with scalar regression. After, the parameters are estimated separately with the standard gradient descent method. On the last step, the finite time frequency estimate is found algebraically. In contrast to the standard gradient descent method, there is not a trade-off between estimation duration and sensitivity to measurement noise.