This paper introduces a large-scale spontaneous speech corpus of Korean, named KsponSpeech. This corpus contains 969 h of general open-domain dialog utterances, spoken by about 2000 native Korean speakers in a clean environment. All data were constructed by recording the dialogue of two people freely conversing on a variety of topics and manually transcribing the utterances. The transcription provides a dual transcription consisting of orthography and pronunciation, and disfluency tags for spontaneity of speech, such as filler words, repeated words, and word fragments. This paper also presents the baseline performance of an end-to-end speech recognition model trained with KsponSpeech. In addition, we investigated the performance of standard end-to-end architectures and the number of sub-word units suitable for Korean. We investigated issues that should be considered in spontaneous speech recognition in Korean. KsponSpeech is publicly available on an open data hub site of the Korea government.
Korean and Japanese have different writing scripts but share the same Subject-Object-Verb (SOV) word order. In this study, we pre-train a language-generation model using a Masked Sequence-to-Sequence pre-training (MASS) method on Korean and Japanese monolingual corpora. When building the pre-trained generation model, we allow the smallest number of shared vocabularies between the two languages. Then, we build an unsupervised Neural Machine Translation (NMT) system between Korean and Japanese based on the pre-trained generation model. Despite the different writing scripts and few shared vocabularies, the unsupervised NMT system performs well compared to other pairs of languages. Our interest is in the common characteristics of both languages that make the unsupervised NMT perform so well. In this study, we propose a new method to analyze cross-attentions between a source and target language to estimate the language differences from the perspective of machine translation. We calculate cross-attention measurements between Korean–Japanese and Korean–English pairs and compare their performances and characteristics. The Korean–Japanese pair has little difference in word order and a morphological system, and thus the unsupervised NMT between Korean and Japanese can be trained well even without parallel sentences and shared vocabularies.
Since building Neural Machine Translation (NMT) systems requires a large parallel corpus, various data augmentation techniques have been adopted, especially for low-resource languages. In order to achieve the best performance through data augmentation, the NMT systems should be able to evaluate the quality of augmented data. Several studies have addressed data weighting techniques to assess data quality. The basic idea of data weighting adopted in previous studies is the loss value that a system calculates when learning from training data. The weight derived from the loss value of the data, through simple heuristic rules or neural models, can adjust the loss used in the next step of the learning process. In this study, we propose EvalNet, a data evaluation network, to assess parallel data of NMT. EvalNet exploits a loss value, a cross-attention map, and a semantic similarity between parallel data as its features. The cross-attention map is an encoded representation of cross-attention layers of Transformer, which is a base architecture of an NMT system. The semantic similarity is a cosine distance between two semantic embeddings of a source sentence and a target sentence. Owing to the parallelism of data, the combination of the cross-attention map and the semantic similarity proved to be effective features for data quality evaluation, besides the loss value. EvalNet is the first NMT data evaluator network that introduces the cross-attention map and the semantic similarity as its features. Through various experiments, we conclude that EvalNet is simple yet beneficial for robust training of an NMT system and outperforms the previous studies as a data evaluator.
In this paper, we propose a new method for code-switching (CS) automatic speech recognition (ASR) in Korean. First, the phonetic variations in English pronunciation spoken by Korean speakers should be considered. Thus, we tried to find a unified pronunciation model based on phonetic knowledge and deep learning. Second, we extracted the CS sentences semantically similar to the target domain and then applied the language model (LM) adaptation to solve the biased modeling toward Korean due to the imbalanced training data. In this experiment, training data were AI Hub (1033 h) in Korean and Librispeech (960 h) in English. As a result, when compared to the baseline, the proposed method improved the error reduction rate (ERR) by up to 11.6% with phonetic variant modeling and by 17.3% when semantically similar sentences were applied to the LM adaptation. If we considered only English words, the word correction rate improved up to 24.2% compared to that of the baseline. The proposed method seems to be very effective in CS speech recognition.