From Isolated Words to Unconstrained Documents: Bringing Handwriting Recognition to the Meeting Room1

The earliest handwriting recognition approaches date back to the eighties, when the first attempts of automatically recognizing handwritten words were proposed, e.g., in Mori et al. (1984), Burr (1983), or Bozinovic and Srihari (1989). However, it is only in the mid nineties that the domain takes off thanks to two main factors (Vinciarelli, 2002): on one hand, the diffusion of cheap image acquisition and storage technologies that made it possible to perform experiments on large databases of handwritten material. On the other hand, the extensive use of handwriting recognition tasks (in particular the automatic transcription of handwritten digits) in the machine learning community (Le Cun et al., 1990; Cortes and Vapnik, 1995). While not being aimed at the improvement of handwriting recognition technologies digit recognition was adopted because it was a challenging task for pattern recognition techniques machine learning works still contributed significantly in terms of methodology. Initially, the focus of handwriting research was on two application domains, namely the recognition of town names in handwritten postal addresses and the transcription of bank-check amounts written in letters (Plamondon and Srihari, 2000). The two tasks above dominated handwriting recognition research for at