In the field of Soft Robotics, viscoelasticity has been proved beneficial for human assistance applications. The human skeletal muscle system, as well as many soft materials commonly used in soft robotic applications, have viscoelastic properties. Viscoelasticity can be modelled using a set of equations known as the Linear Viscoelastic Models (LVMs). This modelling approach has two main limitations: high mathematical complexity and high computational cost. Here, these limitations are addressed in two ways. Firstly, the Piecewise Linearisation method is used to reduce the mathematical complexity of LVMs. Secondly, a modelling approach based on feedforward artificial neural networks (ANNs) is used to reduce the computational cost. The aim of both modelling approaches is to describe the non-linear, strain-dependent, and time-dependent stress response of seven thermoplastic elastomers. On the one hand, the implementation of the Piecewise Linearisation method yielded the PL-SLS model and the PL-Wiechert model. Both models were successful in predicting the viscoelastic behaviour of the materials, outperforming similar modelling tools documented in the literature. On the other hand, four different architectures of ANN models are developed, categorized in rate-dependent and rate-independent. Results highlight the rate-dependent architecture as the most suitable. The ANN models achieved a similar prediction performance as the PL models. The ANN model for the natural rubber material is further validated in a real-time simulation environment, in Simulink. This soft material is found to be the best candidate to imitate the mechanical properties of the human tendon. On the one hand, the performance prediction of the ANN models is adequate for a sine wave strain input, when the strain rate is constant. On the other hand, the response of the ANN model is unstable under variable strain rates. This highlights an important limitation of the training set used for developing the ANN models, which only contains data for three different strain rates. Finally, the three modelling tools developed in this research are a direct improvement to current modelling approaches. Nonetheless, a richer training set is required to improve the ANN models real-time response.