training data for vae

[GooThinker]

The technical report mentions that an audio VAE was trained using over 1 million music-related data samples. However, in my testing, I found that this VAE also performs quite well on speech. This makes me curious: in audio training, is audio quality more important than quantity? Is a dataset of 1 million audio samples sufficient?

[sskalnik]

Datasets of at least 5000 audio files are best, as a general rule.

Quality and quantity are important. Good metadata, like tags and prompts and so on, is perhaps the most important. A dataset of high-quality audio with poor metadata will probably be outperformed by a dataset of "barely decent" audio with great metadata.