As the character-based end-to-end automatic speech recognition (ASR) models evolve, the choice of acoustic modeling units becomes important.Since Korean is a fairly phonetic language and has a unique writing system with its own Korean alphabet, it's worth investigating modeling units for an end-to-end Korean ASR task.In this work, we introduce lexicon-free modeling units in Korean, and explore them using a hybrid CTC/Attention-based encoder-decoder model.Five lexicon-free units are investigated: Syllable-based Korean character (with English character for a code-switching task), Korean Jamo character (with English character), subword on syllable-based character (with sub-word in English), sub-word on Jamo character (with sub-words in English), and finally byte unit, which is a universal one across language.Experiments on Zeroth-Korean (51.6 hrs) and Medical Record (2530 hrs) are done for Korean and Korean-English codeswitching ASR tasks, respectively.Sequence-to-sequence learning with sub-words based on Korean syllables (and subwords in English) performs the best for both tasks without lexicon and an extra language model integration.

Data augmentation is crucial to improving the performance of deep neural networks by helping the model avoid overfitting and improve its generalization. In automatic speech recognition, previous work proposed several approaches to augment data by performing speed perturbation or spectral transformation. Since data augmented in this manner has similar acoustic representations as the original data, it has limited advantage in improving generalization of the acoustic model. In order to avoid generating data with limited diversity, we propose a voice conversion approach using a generative model (WaveNet), which generates a new utterance by transforming an utterance to a given target voice. Our method synthesizes speech with diverse pitch patterns by minimizing the use of acoustic features. With the Wall Street Journal dataset, we verify that our method led to better generalization compared to other data augmentation techniques such as speed perturbation and WORLD-based voice conversion. In addition, when combined with the speed perturbation technique, the two methods complement each other to further improve performance of the acoustic model.

The analysis of query logs from blog search engines show that news-related queries occupy a significant portion of the logs. This raises a interesting research question on whether the blogosphere can be used to identify important news stories. In this paper, we present novel approaches to identify important news story headlines from the blogosphere for a given day. The proposed system consists of two components based on the language model framework, the query likelihood and the news headline prior. For the query likelihood, we propose several approaches to estimate the query language model and the news headline language model. We also suggest several criteria to evaluate the news headline prior that is the prior belief about the importance or newsworthiness of the news headline for a given day. Experimental results show that our system significantly outperforms a baseline system. Specifically, the proposed approach gives 2.62% and 10.19% further increases in MAP and P@5 over the best performing result of the TREC'09 Top Stories Identification Task.

Blog feed search aims to identify a blog feed with a recurring interest in a given topic. In this paper, we investigate the "pseudo-relevance feedback" for blog feed search task, where its unit of relevance judgment is not based on a blog post but a blog feed (the collection of all its constituent posts). This paper focuses on two characteristics of feed search task, blog feed's topical diversity and multifaceted property of query. We propose a novel feed-level selection of local posts which uses only highly relevant local posts in each top-ranked feed, in order to capture the correct and diverse relevant information to a given topic. Experimental results show that the proposed approach outperforms traditional feedback approaches. Especially, the proposed approach gives 2% further increase of nDCG over the best performing result of TREC '08 Blog Distillation Task.