Replies: 2 comments
-
Thanks for your email. @Emrys365, can you follow up on this discussion? |
Beta Was this translation helpful? Give feedback.
-
Sorry for the late response. In most existing NN-based multichannel speech enhancement (SE) models, it is likely that they learned to process a fixed array geometry depending on the training data. For example, CHiME-4 data features a 6-mic rectangular array. Instead, there are three major solutions.
|
Beta Was this translation helpful? Give feedback.
-
Hi,
I've noticed many of the speech enhancement algorithms have multiple channels (5 or 10 even).
Some of my audio data when processed from soundfile/librosa has 2 channels. However, the # of channels isn't equivalent to the expected input of the multi-channel speech enhancement model.
What are the canonical ways of expanding the data to more channels without having to train or build out a new model?
For example, I've tried adding 0s to the extra channels or copied and pasted the data in the original channels but that hasn't worked too well.
Beta Was this translation helpful? Give feedback.
All reactions