The emergence of acoustic metamaterials (AMs) provides distinct response to acoustic waves unattainable in nature. However, conventional AMs design has relied on extensive experimentation and a trial-and-error approach requiring significant computational resource. Here, we introduce an inverse-design method based on machine learning using Gauss–Bayesian (GB) models and implement it on low-frequency broadband sound attenuating devices. By utilizing the superiority of GB model on multi-parameter design and optimization efficiency, we successively customize ventilated microperforated sound-absorbing structure, sound absorbers with coherently coupled weak resonances (CCWRs) and ventilated meta-silencers with acoustic consecutive Fano resonances (ACFRs). Under the judiciously adaptive acquisition function and Matérn kernel, the designed GB frameworks with small samples are available to seek the optimal solution established on above physical effects by the cycle of training and prediction with high accuracy and efficiency. Good agreements between numerical and experimental results further verify the effectiveness of the proposed method. With the advantage of low cost, high efficiency and flexibility, our GB-based machine learning method paves the way for tremendous opportunities in future design of various metamaterial-based applications, as well as far-reaching implications for acoustics-related fields.