Evaluation Protocol

[MoayedHajiAli]

Hello,
I noticed in your code that you have n_candidate_per_text to be set to 3 by default. I am wondering if that was used during the evaluation as it was not mentioned in the paper?
Additionally, what CLAP backbone did you use to calculate the CLAP Score in comparison with other methods such as Make-an-Audio, as the scale of the number is very different. Thank you for your help.

Best,

[haoheliu]

Hi @MoayedHajiAli

Yes we use n_candidate_per_text=3 during evaluation. But we found there is no significant improvement on metrics score with this technique.

The CLAP backbone we use is music_speech_audioset_epoch_15_esc_89.98.pt from laion-clap. You can find more details in the following links:
https://github.com/haoheliu/audioldm_eval/blob/passt_replace_panns/audioldm_eval/metrics/clap_score.py
https://github.com/haoheliu/audioldm_eval/blob/passt_replace_panns/audioldm_eval/eval.py#L63

Thanks!