Synthetic Data Generation Method for Turkish Automatic Speech Recognition Systems

Automatic speech recognition system is the understanding of human voice by computers. Automatic speech recognition is an important field of study for human-computer interaction. With the development of technology, the number of studies in this field is increasing. The m odels developed for the current open source automatic speech recognition task give quite successful results in widely used languages. However, uncommon languages such as Turkish are still an area of development. In this study, a synthetic data generation system is proposed for the problem of recognizing proper names, which is an important challenge for automatic speech recognition systems. In the synthetic data generation process of the proposed system, fine-tuning is performed on the XTTS-V2 model based on the transformer. Synthetic data is generated using the obtained model and the determined reference sounds, and fine-tuning is performed on the Whisper model with these data. In the findings obtained, it is concluded that the synthetic data generated with the reference sound shows improvements in ASR model performance.