Conventional spectrometers are typically large, costly, and constrained by a trade-off between spectral measurement range and spectral resolution. This paper proposes a disorder-dispersion miniature spectrometer based on deep learning. In the proposed spectrometer, mass-producible frosted glass is utilized as the spectral encoding device, significantly reducing the manufacturing cost. Further, the spectrometer employs an on-chip detection scheme that does not include any optical components. The distance between the ground glass and detector is only 2 mm, remarkably reducing the device size. Additionally, considering the accuracy and speed of spectral reconstruction, this paper presents a deep learning-based spectral reconstruction method. Experimental results indicate that the proposed spectrometer has a spectral resolution of 1.4 nm and spectral detection range of 420?700 nm. Even in noisy environments, the root-mean-square error between the reconstructed and actual spectra is 1.38×10^－3, and the reconstruction time for a single spectrum is as low as 16 μs.