Handwriting has been known to be a very strong identifying characteristic of an individual and can be considered a behavioural biometric trait. This has made hand writer identification an important area of research. In this paper, a novel offline writer identification system is proposed using ensemble of multi-scale local ternary pattern histogram features. Features are extracted at multiple scales and the resulting feature histograms are subjected to dimensionality reduction via kernel discriminant analysis using spectral regression (SRKDA). Feature vectors extracted at every scale are used to generate models for all writers which are then used to identify a query document. The final decision on the identity of the unknown query document is obtained using majority voting from the generated models. The proposed system has been assessed on two challenging databases (Arabic and English) and the results show that it outperforms the current state of the art systems.