We present a method for global estimation of joint velocities in rigid manipulators. A class of velocity observers with smooth dynamics are introduced whose structure depend on freely selectable functions and gains giving the flexibility of achieving multiple design objectives at the same time. Unlike most methods, no a priori knowledge of an upper bound for velocity magnitude is used. An adaptive version of the observer is also presented to handle a class of structured uncertainties in manipulator model. Simulation example illustrates low noise sensitivity of the globally convergent observer in comparison with semi-globally convergent observers.