def analysis_spec(
wav,
fs,
f0s,
shift=0.005, # Usually 5ms
dftlen=4096, # You can adapt this one according to your pipeline
verbose=1):
'''
Estimate the amplitude spectral envelope.
'''
if sp.pystraight.isanalysiseavailable():
# Use STRAIGHT's envelope if available (as in PML's publications)
SPEC = sigproc.pystraight.analysis_spec(wav,
fs,
f0s,
shift,
dftlen,
keeplen=True)
elif sigproc.interfaces.worldvocoder_is_available():
warnings.warn(
'''\n\nWARNING: straight_mcep is not available,
but WORLD vocoder has been detected and will be used instead.
Note that PML-related publications present results using STRAIGHT vocoder.
The results might be thus different.
''', RuntimeWarning)
# Then try WORLD vocoder
import pyworld
#_f0, ts = pyworld.dio(x, fs, frame_period=shift*1000) # raw pitch extractor # Use REAPER instead
pwts = np.ascontiguousarray(f0s[:, 0])
pwf0 = pyworld.stonemask(wav, np.ascontiguousarray(f0s[:, 1]), pwts,
fs) # pitch refinement
SPEC = pyworld.cheaptrick(
wav, pwf0, pwts, fs,
fft_size=dftlen) # extract smoothed spectrogram
SPEC = 10.0 * np.sqrt(
SPEC
) # TODO Best gain correction I could find. Hard to find the good one between PML and WORLD different syntheses
else:
# Estimate the sinusoidal parameters at regular intervals in order
# to build the amplitude spectral envelope
sinsreg, f0sps = sp.sinusoidal.estimate_sinusoidal_params(
wav, fs, f0s, nbper=3, quadraticfit=True, verbose=verbose - 1)
warnings.warn(
'''\n\nWARNING: Neither straight_mcep nor WORLD's cheaptrick spectral envelope estimators are available.
Thus, a SIMPLISTIC Linear interpolation will be used for the spectral envelope.
Do _NOT_ use this envelope for speech synthesis!
Please use a better one (e.g. STRAIGHT's).
If you use this simplistic envelope, the TTS quality will
be lower than that in the results reported.
''', RuntimeWarning)
SPEC = sp.multi_linear(sinsreg, fs, dftlen)
SPEC = np.exp(SPEC) * np.sqrt(float(dftlen))
return SPEC
def analysis_spec(wav, fs, f0s,
shift=0.005, # Usually 5ms
dftlen=4096, # You can adapt this one according to your pipeline
verbose=1):
'''
Estimate the amplitude spectral envelope.
'''
if 0 and sigproc.pystraight.isanalysiseavailable():
SPEC = sigproc.pystraight.analysis_spec(wav, fs, f0s, shift, dftlen, keeplen=True)
else:
# Estimate the sinusoidal parameters at regular intervals in order
# to build the amplitude spectral envelope
sinsreg, f0sps = sp.sinusoidal.estimate_sinusoidal_params(wav, fs, f0s, nbper=3, quadraticfit=True, verbose=verbose-1)
# Estimate the amplitude spectral envelope
print(" WARNING: straight_mcep is unavailable.")
print(" A SIMPLISTIC Linear interpolation is used for the amplitude envelope.")
print(" Please replace it with a better one (e.g. STRAIGHT's).")
print(" If you use this simplistic envelope, the TTS quality is")
print(" very likely to be lower than that in the results reported.")
SPEC = sp.multi_linear(sinsreg, fs, dftlen)
SPEC = np.exp(SPEC)*np.sqrt(dftlen)
return SPEC