The development of Large Language Models (LLMs) has expanded beyond text-based tasks to speech applications such as Automatic Speech Recognition (ASR) and Automated Speech Translation (AST). However, training speech language models based on LLMs requires large-scale datasets, which are challenging to construct. To address data scarcity, previous studies have explored synthetic data generation using ASR for transcribing unlabeled speech and Text-to-Speech (TTS) for generating speech from text. While synthetic data enables large-scale dataset construction without human intervention, concerns persist regarding quality degradation and its impact on model performance. This study investigates the effects of synthetic data on ASR and AST tasks. Experimental results indicate that synthetic data alone may degrade performance, whereas combining it with real data can enhance performance, demonstrating its potential when integrated with other data sources.