Semi-Supervised Training of Language Model on Spanish Conversational Telephone Speech Data

This work addresses one of the common issues arising when building a speech recognition system within a low-resourced scenario adapting the language model on unlabeled audio data. The proposed methodology makes use of such data by means of semisupervised learning. Whilst it has been proven that adding system-generated labeled data for acoustic modeling yields good results, the benefits of adding system-generated sentence hypotheses to the language model are vaguer in the literature. This investigation focuses on the latter by exploring different criteria for picking valuable, well-transcribed sentences. These criteria range from confidence measures at word and sentence level to sentence duration metrics and grammatical structure frequencies. The processing pipeline starts with training a seed speech recognizer using only twenty hours of Fisher Spanish phone call conversations corpus. The proposed procedure attempts to augment this initial system by supplementing it with transcriptions generated automatically from unlabeled data with the use of the seed system. After generating these transcriptions, it is estimated how likely they are, and only the ones with high scores are added to the training data. Experimental results show improvements gained by the use of an augmented language model. Although these improvements are still lesser than those obtained from a system with only acoustic model augmentation, we consider the proposed system (with its low cost in terms of computational resources and the ability for task adaptation) an attractive technique worthy of further exploration. c © 2016 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the Organizing Committee of SLTU 2016.