Training w/ WAV Files Converted From MP3?

[automationist]

I'm wondering which one of these options would result in a better model:

1. Adding more data (MP3)
2. Not adding the MP3 data, and only training with the data that has never been compressed.

What do you guys think?

[thorstenMueller]

Not being an audio expert, but i would use the original wavefiles (if you have access to them). Compressing to mp3 might remove audio details that will not come back by converting back to wave. If that's your question.

[domcross]

When the MP3s have high quality (bitrate of 192KBits/sec or higher) it should be good enough to give it a try. As @thorstenMueller noted the best option is using original WAV and not using lossy compression (like MP3 is) at all, though.