/mlpg_c

Primary LanguagePythonMIT LicenseMIT

mlpg_c ====

Maximum Likelihood Parameter Generation (MLPG) [1,2] implementation in C for Python

Install ----

pip install mlpg_c

Usage of mlpg_solve ----

from mlpg_c import mlpg_c as mlpg

param_gen = mlpg.mlpg_solve(jnt_sd_mat, prec_vec, coef_vec)

Variable desc. ----

param_gen: generated parameter trajectory: T x dim

jnt_sd_mat: joint static-delta feature vector sequence: T x (dim*2)

prec_vec: vector of diagonal precision (inverse covariance) matrix: (dim*2) x 1

coef_vec: vector of delta coefficients: n_coeff x 1. e.g.: [-0.5,0.5,0.0]

Usage of mlpg_solve_seq ----

from mlpg_c import mlpg_c as mlpg

param_gen = mlpg.mlpg_solve_seq(jnt_sd_mat, prec_mat, coef_vec)

Variable desc. ----

param_gen: generated parameter trajectory: T x dim

jnt_sd_mat: joint static-delta feature vector sequence: T x (dim*2)

prec_mat: sequence of diagonal precision (inverse covariance) matrices: T x (dim*2)

coef_vec: vector of delta coefficients: n_coeff x 1, e.g.: [-0.5,0.5,0.0]

To-do: ----

  • full precision matrix
  • docs

References: ----[1] K. Tokuda, T. Kobayashi, S. Imai, Speech parameter generation from HMM using dynamic features, in Proc. ICASSP, Detroit, USA, May 1995, pp. 660––663.

[2] T. Toda, A. W. Black, K. Tokuda, Voice conversion based on maximum-likelihood estimation of spectral parameter trajectory, IEEE Trans. Audio Speech Lang. Process., vol. 15, no. 8, 2222-–2235, 2007.