Confidence- and margin-based MMI/MPE discriminative training for off-line handwriting recognition

Abstract

We present a novel confidence-and marginbased discriminative training approach for model adaptation of a hidden Markov model (HMM) based handwriting recognition system to handle different handwriting styles and their variations.Most current approaches are maximum-likelihood (ML) trained HMM systems and try to adapt their models to different writing styles using writer adaptive training, unsupervised clustering, or additional writer specific data. Here, discriminative training based on the maximum mutual information (MMI) and minimum phone error (MPE) criteria are used to train writer independent handwriting models. For model adaptation during decoding, an unsupervised confidence-based discriminative training on a word and frame level within a two-pass decoding process is proposed.The proposed methods are evaluated for closedvocabulary isolated handwritten word recognition on the IFN/ENIT Arabic handwriting database, where the word-error-rate is decreased by 33% relative compared to a ML trained baseline system. On the largevocabulary line recognition task of the IAM English handwriting database, the word-error-rate is decreased by 25% relative.

A convergence analysis of log-linear training and its application to speech recognition

“…We used the maximum likelihood GHMM system of [19] as our baseline system. In the feature extraction of this system, only elementary preprocessing steps (deslanting and size normalization) are used, which are commonly employed in image recognition.…”

“…The baseline system achieves a WER of 32.8% on the development corpus and 39.4% on the test corpus. A more detailed description of the baseline system can be found in [19].…”

“…Especially interesting is the comparison to the results obtained by methods which are commonly used in speech recognition. Dreuw et al [19] obtain 31.6% WER with GHMMs with lattice-based MMI training and 30.0% with MPE training. Their system is further improved with an additional discriminative adaptation method (29.0% WER).…”

Interactive Transcription of Old Text Documents

“…Current state-of-art approaches use additional steps and techniques over the basic approach to increment the system performance. For instance in (Dreuw et al, 2011), discriminative training is used to improve HMM estimation. Alternatively, in (España-Boquera et al, 2011), Neural Networks (NN) are used within the HMMs to improve their performance.…”

“…Unfortunately, GIDOC technology has became outdated. For instance, GIDOC feature extraction method was used in our baseline system until we integrated the feature extraction method of Dreuw et al (2011), which improved this important step. Similarly, neural network based (Graves et al, 2009) (España-Boquera et al, 2011) (Hinton et al, 2012) (Kozielski et al, 2013) systems have recently outperformed Gaussian HMMs, which are GIDOC baseline, becoming the state-of-the-art.…”