def h_2(raw_sample):
w = windowed(raw_sample)
s = np.fft.fft(w.data, n=N)
p = np.sqrt(1 + (KV / K_C)**2)
pp = np.concatenate([p, np.ones(len(s) - 2 * K_MAX), p[::-1]])
s = s * pp
f = np.fft.fftfreq(len(w.data), T_E)[:K_MAX]
n = len(s)
c, ndelay = complex_cepstrum(s)
rc = real_cepstrum(s)
K_C_MIN_SEARCH = int(np.round(4 / (200 * T_E)))
K_C_MAX_SEARCH = int(np.round(4 / (75 * T_E)))
k_f_p = np.argmax(rc[K_C_MIN_SEARCH:K_C_MAX_SEARCH]) + K_C_MIN_SEARCH
f_p = 4 / (k_f_p * T_E)
print(f_p)
k_f_p -= 200
print(k_f_p)
# plt.plot(rc)
# plt.axvline(x=K_C_MIN_SEARCH, color="b")
# plt.axvline(x=K_C_MAX_SEARCH, color="r")
# plt.axvline(x=k_f_p + 200, color="g", alpha=0.2)
# plt.axvline(x=k_f_p, color="y")
# plt.show()
c_altered = np.concatenate(
[c[:k_f_p], np.zeros(len(c) - 2 * k_f_p), c[-k_f_p:]])
# print(len(c))
assert (len(c) == len(c_altered))
new_s = inverse_complex_cepstrum(c_altered, ndelay)
return normalize(
Spectrum(data=np.concatenate(
[np.zeros(K_MIN), np.abs(new_s[K_MIN:K_MAX])]),
freq=f,
file_name=raw_sample.file_name,
phoneme=raw_sample.phoneme))
def biased_exp(spectrum: Spectrum, beta: float) -> Spectrum:
"""TODO
"""
data = np.exp(spectrum.data) - beta
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme)
def enhance_high_freqs(spectrum: Spectrum) -> Spectrum:
"""TODO
"""
p = np.sqrt(1 + (KV / K_C)**2)
data = spectrum.data * p
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme)
def zero_low_frequencies(spectrum: Spectrum) -> Spectrum:
"""TODO
"""
data = np.concatenate([np.zeros(K_MIN), spectrum.data[K_MIN:]])
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme)
def biased_log(spectrum: Spectrum) -> (Spectrum, float):
"""TODO
"""
avg = np.average(spectrum.data)
beta = RO * avg
data = np.log(spectrum.data + beta)
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme), beta
def normalize(spectrum: Spectrum) -> Spectrum:
"""TODO
"""
data = np.copy(spectrum.data)
norm = np.sqrt(np.sum(np.power(data, 2)))
data /= norm
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme)
def load_samples():
data_bank = []
for r, d, f in os.walk(NEW_SAMPLES_STORE):
for filepath in f:
filename = path.basename(filepath)
if filename.startswith("."):
continue
phoneme = filename.split("_")[0]
assert (phoneme in PHONEMES)
data_bank.append(
Spectrum(data=np.load(path.join(NEW_SAMPLES_STORE, filepath)),
freq=None,
file_name=filepath,
phoneme=Phoneme(phoneme)))
return data_bank
def smooth(spectrum: Spectrum) -> Spectrum:
"""TODO
"""
# normalize weight function
weight_func = (1 / 2) * (1 + np.cos(2 * pi * (QSV - Q / 2) / Q))
total = np.sum(weight_func)
weight_func = weight_func / total
# smooth the signal
data = np.convolve(spectrum.data, weight_func, mode='same')
return Spectrum(data=data,
file_name=spectrum.file_name,
freq=spectrum.freq,
phoneme=spectrum.phoneme)
def spectrum_of(sample: Sample) -> Spectrum:
"""Applies the discrete Fourier Transform on the current sample.
Frequencies above K_MAX / T (about 3000 Hz) are cut.
:remark: Spectrum.data only contains the modulus of the fft spectrum
:param sample: audio sample to work on
:returns: the generated Spectrum object
"""
data = np.abs(np.fft.fft(sample.data, n=N))[:K_MAX]
freq = np.fft.fftfreq(N, T_E)[:K_MAX]
return Spectrum(data=data,
file_name=sample.file_name,
freq=freq,
phoneme=sample.phoneme)