Unit-norm tight frame-based sparse representation with application to speech inpainting