ulmo

Large acoustic models

Things to investigate

• alternatives to K-means and better tokenization
  • Self supervised pretext tasks such as next frame prediction, infill, ABX (STFT, Mel Spec, CQT, CWT, DWT, WPT choice tasks)
• long context to discover repeated impulses
• adding an token
• Try a denoising autoencoder in place of kmeans to create labels.
• Data augmentation by using sets of spectrogram parameters (vary SR +- 10%)
• methods of self supervision such as SimCLR, wav2vec2,and BYOL