Parallel Training of TN and ITN Models Through CycleGAN for Improved Sequence to Sequence Learning Performance

Text Normalization (TN) and Inverse Text Normalization (ITN) both are essential for conversational systems like text-to-speech (TTS) and automatic speech recognition (ASR). Traditionally the rule-based approach was used for both TN and ITN, but nowadays the data-driven method is being adopted as a viable solution due to the advancement of deep learning techniques. The data-driven models are trained on a large corpus of parallel texts by minimizing an end-to-end loss, where TN and ITN are usually treated as separate tasks. In this work, we have integrated TN and ITN in the same pipeline utilizing the CycleGAN framework in order to train both models in a parallel fashion. The proposed approach not only trains both networks simultaneously but also improves the performance of the models compared to models trained individually. Incorporating cycle loss and adversarial loss with end-to-end loss benefits the overall performance of the networks by forming a multi-task training framework.