GMM

Voice conversion using Gaussian Mixture Models

Converting voice from a source speaker to a target speaker using a parallel dataset and GMMs

Dataset used:

• CMU Arctic dataset
• Our own dataset made from an audiobook sampled at 2/3 seconds
  • 2 second sampled
  • 3 second sampled

Sys params/Constants

• Sampling frequency: 16000
• FFT size: 1024
• Max utterances to be used together: 100
• Test size: 0.03% of training set
• frame_period = 5
• The number of samples between successive frames/hop length = 80
• Order of mel-cepstrum = 24

Software libraries used

• sklearn (for gaussian mixture model)
• nnmnkwii (for DTWAligner, Dataset_Model
• librosa (for audio file management)
• pysptk (for audio synthesis)

References

• T. Toda, A. W. Black, and K. Tokuda, “Voice conversion based on maximum likelihood estimation of spectral parameter trajectory,” IEEETrans. Audio, Speech, Lang. Process, vol. 15, no. 8, pp. 2222–2235, Nov. 2007.
• Kobayashi, Kazuhiro, et al. “Statistical Singing Voice Conversion with Direct Waveform Modification based on the Spectrum Differential.” Fifteenth Annual Conference of the International Speech Communication Association. 2014.