

In [94]:

    
%load_ext autoreload
%matplotlib inline
%autoreload 2
import matplotlib.pyplot as plt
from IPython.display import Audio, display
import numpy as np
from pysas import World, waveread









    



The autoreload extension is already loaded. To reload it, use:
  %reload_ext autoreload

Speech analysis and synthesis with World



In [159]:

    
signal, samplingrate, _ = waveread("test/cmu_arctic/arctic_a0001.wav")  # from http://festvox.org/cmu_arctic/dbs_bdl.html



In [3]:

    
world = World(samplingrate)



In [97]:

    
f0, spec_mat, aperiod_mat = world.analyze(signal)



In [98]:

    
plt.plot(f0)









    Out[98]:





[<matplotlib.lines.Line2D at 0x11a506710>]



In [99]:

    
spec = spec_mat[300]



In [100]:

    
plt.plot(np.log(spec))









    Out[100]:





[<matplotlib.lines.Line2D at 0x11a510940>]



In [117]:

    
out = world.synthesis(f0, spec_mat, aperiod_mat)



In [118]:

    
plt.plot(signal)
plt.plot(out)









    Out[118]:





[<matplotlib.lines.Line2D at 0x11a87e0f0>]

Mel-Cepstrum Analysis



In [125]:

    
from pysas.mcep import spec2mcep, mcep2spec, mcep2coef, coef2mcep, estimate_alpha



In [121]:

    
alpha = round(estimate_alpha(samplingrate), 3) 
alpha









    Out[121]:





0.41



In [122]:

    
mcep = spec2mcep(spec, 24, alpha)



In [123]:

    
spec2 = mcep2spec(mcep, alpha, world.fftsize())



In [124]:

    
plt.plot(np.log(spec))
plt.plot(np.log(spec2[:world.envelopesize()]))









    Out[124]:





[<matplotlib.lines.Line2D at 0x11ad042b0>]

from fft



In [126]:

    
i = 80* 300
windowsize = 1024
sig = signal[i:i+windowsize] * np.hanning(windowsize)
power_spectrum = (np.absolute(np.fft.fft(sig)) ** 2)[:(windowsize>>1) + 1]
plt.plot(np.log(power_spectrum))









    Out[126]:





[<matplotlib.lines.Line2D at 0x11affe7f0>]



In [127]:

    
fft_mcep = spec2mcep(power_spectrum, 20, alpha)
reconst_pspec = mcep2spec(fft_mcep, alpha, windowsize)



In [128]:

    
plt.plot(np.log(spec))
plt.plot(np.log(spec2))
plt.plot(np.log(power_spectrum))
plt.plot(np.log(reconst_pspec))









    Out[128]:





[<matplotlib.lines.Line2D at 0x11b499b38>]

Convert Mel-Cepstrum to Coefficients for MLSA Degital Filter



In [129]:

    
coef = mcep2coef(fft_mcep, alpha)
reconst_mcep = coef2mcep(coef, alpha)



In [130]:

    
plt.plot(fft_mcep)
plt.plot(reconst_mcep)
plt.plot(coef)









    Out[130]:





[<matplotlib.lines.Line2D at 0x11b557cc0>]

Generate excited pulse from F0 sequence



In [131]:

    
from pysas.excite import ExcitePulse



In [114]:

    
ep = ExcitePulse(16000, 80, False)



In [115]:

    
plt.plot(ep.gen(f0))









    Out[115]:





[<matplotlib.lines.Line2D at 0x11a7fa438>]

Speech Synthesis WIth Mel-Cepstrum Analysis and MLSA Filter



In [298]:

    
from pysas.synthesis.mlsa import MLSAFilter
from pysas.synthesis import Synthesis
from pysas.mcep import spec2mcep_from_matrix



In [299]:

    
mlsa = MLSAFilter(20, 0.41, 5)



In [300]:

    
pulse = ep.gen(f0)



In [301]:

    
mcep_mat = spec2mcep_from_matrix(spec_mat, 20, 0.41)ct



In [302]:

    
coef_mat = []
for i in range(mcep_mat.shape[0]):
    coef_mat.append(mcep2coef(mcep_mat[i], 0.41))
coef_mat = np.array(coef_mat)



In [303]:

    
syn = Synthesis(80, mlsa)



In [296]:

    
synth = syn.synthesis(pulse, coef_mat)



In [288]:

    
plt.plot(synth)
plt.plot(signal)









    Out[288]:





[<matplotlib.lines.Line2D at 0x11d630208>]

Original



In [265]:

    
display(Audio(data=np.int16(signal * 32767.0), rate=16000))

Synthesized



In [297]:

    
display(Audio(data=np.int16(synth * 32767.0), rate=16000))