Bidirectional Spoken-Written Text Conversion with Large Language Models

Traditional ASR systems normalize transcriptions into spoken form for training, leading to spoken form outputs. In contrast, modern Transformer-based models directly map speech-to-text, preserving the written form. However, existing speech databases mainly contain spoken form, causing inconsistencies in recognition results. To address this, Inverse Text Normalization (ITN) is required, and dual transcription data is crucial for effective training. However, constructing such datasets is costly and time-consuming. This study proposes a data augmentation method leveraging LLMs to automatically generate dual transcription data with minimal effort. It employs iterative learning to expand the dataset through supervised and semi-supervised methods. Additionally, the bidirectional text conversion (BTC) model supports both ITN and TN within a unified framework. Experimental results demonstrate that the proposed method achieved an ERR of 13.4% and 4.7%, outperforming prior approaches.