Electromyogram (EMG)-torque modeling is of value to many different application areas, including ergonomics, clinical biomechanics and prosthesis control. One important aspect of EMG-torque modeling is the ability to account for the joint angle influence. This manuscript describes an experimental study which relates the biceps/triceps surface EMG of 12 subjects to elbow torque at seven joint angles (spanning 45-135°) during constant-posture, quasi-constant-torque contractions. Advanced EMG amplitude (EMGσ) estimation processors (i.e., whitened, multiple-channel) were investigated and three non-linear EMGσ-torque models were evaluated. When EMG-torque models were formed separately for each of the seven distinct joint angles, a minimum "gold standard" error of 4.23±2.2% MVCF90 resulted (i.e., error relative to maximum voluntary contraction at 90° flexion). This model structure, however, did not directly facilitate interpolation across angles. The best model which did so (i.e., parameterized the angle dependence), achieved an error of 4.17±1.7% MVCF90. Results demonstrated that advanced EMGσ processors lead to improved joint torque estimation. We also contrasted models that did vs. did not account for antagonist muscle co-contraction. Models that accounted for co-contraction estimated individual flexion muscle torques that were ∼29% higher and individual extension muscle torques that were ∼68% higher.