Efficient online learning with individual learning-rates for phoneme sequence recognition

We describe a fast and efficient online algorithm for phoneme sequence speech recognition. Our method is using a discriminative training to update the model parameters one utterance at a time. The algorithm is based on recent advances in confidence-weighted learning and it maintains one learning rate per feature. The algorithm is evaluated using the TIMIT database and was found to achieve the lowest phoneme error rate compared to other discriminative and generative models. Additionally, our algorithm converges in less iterations over the training set compared with other online methods.