dcn01 / lipsync-negu93

LipSync by: José Agustín Barrachina & Matias Dwek

LipSync

by: José Agustín Barrachina & Matias Dwek

This programs loads an audio of one or two speakers and performs the following audio processing algorithms.

• Phonems Identifier using a LPC analysis system and formants detection
• Voice recognition using a RNN-LMST [1]
  • Input of the network 13 MFCC (Mel-Frequency Cepstral Coeficients)
  • 3 hidden layers
  • 250 LSTM cells
  • Optimizer: Adam
  • Loss: CTC (Connectionist Temporal Classification)
• Speaker recognition (up to two speakers) using a GMM (Gaussian Mixture Model) [2]
• Speed modifier
  • Uses over/sub sampling to avoid deep/acute deformation
• Pitch modification using FFT

GUI

• Mouth image shows the speaker mouth movement
• Progress line in green (in sync with the playing audio)
• Done with pyQT5

References

[1] Abdel-rahman Mohamed Alex Graves and Geoffrey Hinton (2012). “SPEECH RECOGNITION WITH DEEP RECURRENT NEURAL NETWORKS”. Department of Computer Science, University of Toronto.

[2] Vibha Tiwari (2010). “MFCC and its applications in speaker recognition”. En: International journal on emerging technologies 1.1, p´ags. 19-22.