Towards a new image-based spectrogram segmentation speech coder optimised for intelligibility