Questions about environment robustness eval

[shuowander]

• For the SD-QA set, I noticed that the score range in the leaderboard is [0,100], while for speaker and environment variations, the score range is 1,5. Could you clarify if the same test set is used for robustness evaluation and the leaderboard? Are the same metrics applied in both cases?

• In Figure 5, you illustrate the impact of various environments. Regarding babble and white noise, I observed that the x-axis represents noise levels ranging from [-5,20]. Could you explain how noise levels are defined and why both negative and positive values are included? Additionally, could you provide more details on how noise is added to the data? And how many data do you use to test the noise level variation?