39 karma · joined March 2, 2016
For applied ML, my tips are: make sure you learn the dark side of BatchNorm and Dropout, start with simple and elegant baselines instead of complex SOTA algorithms, spend more time on understanding your data than trying algorithms, be aware that SOTA in a related task will often suck at your task, be data driven. Also, most of your ideas will not work but you have to try and conduct experiments carefully.
1. Speech distortion is extremely detrimental to ASR models while human listeners may not be able to notice. Noise reduction models such as RNNNoise and DeepFilterNet try to reduce "perceptible" noises. And doing that will create imperceptible distortion which ASR models does not like at all.
2. Many noise reduction models apply on raw spectrograms or ERB band such as RNNNoise and DeepFilterNet (equivalent rectangular bands). On the other hands, ASR models mostly run on melspectrograms. This mismatch tends to create problem. I have seen many papers from Google claimed that reducing noise in melspectrogram often helps their keyword spotting.
Source: I work in this research area.
Publish or perish!
I got downvoted when I expressed similar opinion with regards to MiniGPT4. I guess HN crowd value usefulness more than real contribution.