Clean/target data are reverberant speech and not clean speech

[dr-costas]

Hi,

At the specifications of the challenge it is written:

EEE ICASSP 2023 Deep Noise Suppression (DNS) grand challenge is the 5th edition of Microsoft DNS challenges with focus on deep speech enhancement achieved by suppressing background noise, reverberation and neighboring talkers and enhancing the signal quality

Thus, this means that the clean/target speech signals must be without reverberation. Though, at the data creation file, noisyspeech_synthesizer_singleprocess.py and at line 192, we can see that the clean signal is processed and made reverberant.

Can you please elaborate on this? Since the script for data creation comes from the challenge organisers, does the above processing of the clean data (i.e. making the clean speech signals reverberant) mean that the target speech signals are reverberant signals? If yes, then what is meant by

EEE ICASSP 2023 Deep Noise Suppression (DNS) grand challenge is the 5th edition of Microsoft DNS challenges with focus on deep speech enhancement achieved by suppressing background noise, reverberation and neighboring talkers and enhancing the signal quality

?

Thank you in advance for the info.

[XuWink]

我也遇到了这个疑问