A Comparative Study on Neural Architectures and Training Methods for Japanese Speech Recognition

End-to-end (E2E) modeling is advantageous for automatic speech recognition\n(ASR) especially for Japanese since word-based tokenization of Japanese is not\ntrivial, and E2E modeling is able to model character sequences directly. This\npaper focuses on the latest E2E modeling techniques, and investigates their\nperformances on character-based Japanese ASR by conducting comparative\nexperiments. The results are analyzed and discussed in order to understand the\nrelative advantages of long short-term memory (LSTM), and Conformer models in\ncombination with connectionist temporal classification, transducer, and\nattention-based loss functions. Furthermore, the paper investigates on\neffectivity of the recent training techniques such as data augmentation\n(SpecAugment), variational noise injection, and exponential moving average. The\nbest configuration found in the paper achieved the state-of-the-art character\nerror rates of 4.1%, 3.2%, and 3.5% for Corpus of Spontaneous Japanese (CSJ)\neval1, eval2, and eval3 tasks, respectively. The system is also shown to be\ncomputationally efficient thanks to the efficiency of Conformer transducers.\n