In gas metal arc welding, like other welding techniques, the quality of welded joint may be described in terms of weld bead geometry and the presence of welding defects. In turn, the characteristics of welding signals, such as voltage, current and sound, may be used to predict and improve the quality of welded joint. In this work, two sets of adaptive neuro-fuzzy inference system have been used to predict and improve the weld quality characteristics. The required data for modeling were obtained from 57 experiments based on D-optimal design of experiments. The first set is developed to predict the possible welding defects (discontinuity, lack of fusion and overlap) and shape factor of the weld bead. These "predicting adaptive neuro-fuzzy inference system models" have been developed using 13 statistical parameters of the sound, voltage and current signals. The objective of the second set of models, called "improving adaptive neuro-fuzzy inference system models," is to adjust the input welding parameters in such a way that the weld defects are minimized. These models simulate the experiences of professional human welders as the learning databases. Verification tests reveal that the proposed predicting adaptive neuro-fuzzy inference system models can accurately estimate the main weld quality indices in actual gas metal arc welding process. Moreover, experimental results for improving the adaptive neuro-fuzzy inference system models confirm that the defects of faulty weldment can be eliminated after applying the process parameters settings given by these models. The proposed adaptive neuro-fuzzy inference system models may pave the way in assisting the human welder to predict and enhance the weld quality characteristics.