SEAugment

Data augmentations for speech enhancement

What's in it

• SpecTransform was proposed in RNNoise, which is achieved by filtering the noise and speech signal independently for each training example using a second order filter
  

• MixTransform uses different snr combine speech samples and noise samples, which is a common method for data augment in speech enhencement
  

• VolTransform apply step gains to target audio, which simulates different microphone volumes
  

• FilterTransform design a biquad peaking equalizer filter and perform filtering on samples, which simulates different microphone frequency response
  

• ClipTransform truncate samples whose amplitude larger than a given level, which simulates clipping effect
  

• ReverbTransform uses convolve to simulates reverberation. RIR datasets: OpenSLR26 and OpenSLR28
  

• BreakTransform use time-axis mask to simulates frame drop in communication
  

• HowlingTransform

• DynamicTransform