

In [1]:

    
%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

Speech analysis and synthesis with World



In [2]:

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



In [3]:

    
world = World(samplingrate)



In [4]:

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



In [5]:

    
plt.plot(f0)









    Out[5]:





[<matplotlib.lines.Line2D at 0x1193c7a20>]



In [6]:

    
spec = spec_mat[300]



In [7]:

    
plt.plot(np.log(spec))









    Out[7]:





[<matplotlib.lines.Line2D at 0x10d151b00>]



In [8]:

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



In [9]:

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









    Out[9]:





[<matplotlib.lines.Line2D at 0x1193c7f98>]



In [10]:

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

Mel-Cepstrum Analysis



In [11]:

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



In [12]:

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









    Out[12]:





0.41



In [13]:

    
mcep = spec2mcep(spec, 24, alpha)



In [14]:

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



In [15]:

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









    Out[15]:





[<matplotlib.lines.Line2D at 0x11973ea58>]

from fft



In [16]:

    
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[16]:





[<matplotlib.lines.Line2D at 0x116bdbfd0>]



In [17]:

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



In [18]:

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









    Out[18]:





[<matplotlib.lines.Line2D at 0x1197c9320>]

Convert Mel-Cepstrum to Coefficients for MLSA Degital Filter



In [19]:

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



In [20]:

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









    Out[20]:





[<matplotlib.lines.Line2D at 0x119828908>]

Generate excited pulse from F0 sequence



In [21]:

    
from pysas.excite import ExcitePulse



In [22]:

    
ep = ExcitePulse(16000, 80, False)



In [23]:

    
plt.plot(ep.gen(f0))









    Out[23]:





[<matplotlib.lines.Line2D at 0x119940080>]



In [24]:

    
display(Audio(data=np.int16(ep.gen(f0) * 700), rate=16000))

Speech Synthesis WIth Mel-Cepstrum Analysis and MLSA Filter



In [25]:

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



In [26]:

    
cpestrum_dim = 128
mcep_mat = spec2mcep_from_matrix(spec_mat, cpestrum_dim, alpha)



In [27]:

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



In [28]:

    
mlsa = MLSAFilter(cpestrum_dim, alpha, 5)
syn = Synthesis(80, mlsa)
pulse = ep.gen(f0)



In [29]:

    
synth = syn.synthesis(pulse, coef_mat)



In [30]:

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









    Out[30]:





[<matplotlib.lines.Line2D at 0x119ac69e8>]

Original



In [31]:

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

Synthesized



In [32]:

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

voice changer



In [33]:

    
mlsa = MLSAFilter(cpestrum_dim, alpha*1.3 , 5)
syn = Synthesis(80, mlsa)
pulse = ep.gen(f0 * 0.7)
a = syn.synthesis(pulse, coef_mat)
display(Audio(data=np.int16(a * 32767.0), rate=16000))



In [ ]: