05/24/2018 Thr Lab Meeting

Post date: May 29, 2018 5:19:40 PM

Attendees: Zahra, 진석, 지호

• Database

  • TIMIT of 430 speakers

  • Ignore Train/Test/DR slits

  • Do cross-validation: 2 test / 1 val / 7 training, rotate it. The number of total speech sounds per speaker is 10.

• Zahra is working on implementing CNN

  • 1-d CNN with sliding windows with small length (5-10 frames) and whole frequency component

  • Feature: MFCC, spectrogram, spectral flux, spectral tile

  • We found a python version of praat, which can provide various kinds of speech features

• RNN implementation

  • Use masking to define loss function

  • RNN / LSTM / BLSTM

• Method: Boosting

  • Find pairs of speakers who are hard to distinguish (determined by validation data)

  • Learn models and choose features to classify the hard cases

  • It can be cascaded in many levels, but try with narrows ones first