5/1/2018 Tue lab meeting

Post date: May 2, 2018 10:34:46 AM

We have discussed what to do for CeSLea (digital companion) project.

• Text-dependent speaker recognition

  • need keyword spotting using wordnet

  • background model for rejection - research topic

• Text-independent speaker recognition

  • i-vector

  • ergodic (fully connected) HMM

  • RNN / LSTM

  • needs variety of training sentences

• Speaker segmentation

  • finds when the enrolled speakers are talking

  • finds when there is speaker change (not enrolled)

• Speech separation

  • using extra information as Google did recently on Youtube videos

  • general source separation

• Rejection

  • likelihood ratio test (log difference test) - difference between 1st and 2nd speakers, or 1st and average scores

  • normalized Z-test: z = (score-mean)/std > threshold

    • 90% CI 1.645

    • 95% CI 1.960

    • 99% CI 2.576

    • 99.5% CI 2.807