One of the main challenges for punctuation restoration for a low-resource language such as Cantonese is data scarcity. While its spoken and written forms are very different, current Cantonese datasets are mostly from formal written text. Naturally spoken data are very scarce. To address this gap, we leverage LLM to annotate naturally spoken Cantonese transcripts sourced from YouTube. Then, we fine-tune pre-trained language models for punctuation restoration using the LLM-annotated transcripts. Our experiments show that models trained on LLM-annotated transcripts outperform those trained solely on formal written text, despite the smaller dataset size. Our best-performing model achieves performance on par with the strongest LLM evaluated on a benchmark dataset, while being significantly smaller. These findings highlight the potential of LLM-generated data for improving NLP tasks in low-resource languages. Our data and code are publicly available.