def generate_filtered_noise(spectral_envelope_coeffs_list=None, params=None):
n_gap = params["n_gap"]
segment_len = params["segment_len"]
triangle = params["triangle_lin_interpol"]
sqrt_triangle = np.sqrt(triangle)
spectral_envelope_extractor = SpectralEnvelopeExtractor(params)
sound_len = len(spectral_envelope_coeffs_list) * n_gap + segment_len
sound = np.zeros(sound_len)
n_gap = params["n_gap"]
smoother = np.ones(params["segment_len"] // 256)
for i_segment in range(len(spectral_envelope_coeffs_list)):
coeffs = spectral_envelope_coeffs_list[i_segment]
full_spectral_envelope = spectral_envelope_extractor.get_full_spectral_envelope_from_coeffs(
coeffs)
segment = np.random.randn(segment_len) * sqrt_triangle
fft_segment = np.fft.fft(segment)
white_fft_segment = fft_segment / fftconvolve(np.abs(fft_segment),
smoother, "same")
fft_filtered_segment = white_fft_segment * full_spectral_envelope
filtered_segment = np.real(np.fft.ifft(fft_filtered_segment))
add_to_segment(i_segment=i_segment,
source=filtered_segment * sqrt_triangle,
dest=sound,
n_gap=n_gap)
return sound
def generate_periodic_filtered_sound(
segment_period_expressed_in_sample_list=None,
spectral_envelope_coeffs_list=None,
params=None):
sawtooth = generate_periodic_sound(segment_period_expressed_in_sample_list=
segment_period_expressed_in_sample_list,
params=params)
filtered_sawtooth = np.zeros(sawtooth.shape[0])
spectral_envelope_extractor = SpectralEnvelopeExtractor(params)
apowin2 = params["apowin2"]
n_gap = params["n_gap"]
smoother = np.ones(params["segment_len"] // 256)
for i_segment in range(len(spectral_envelope_coeffs_list)):
seg = get_segment(sound=sawtooth, i_segment=i_segment, params=params)
fft_seg = np.fft.fft(seg * apowin2)
fft_seg_envelope = fftconvolve(np.abs(fft_seg), smoother, "same")
white_fft_seg = fft_seg / fft_seg_envelope
coeffs = spectral_envelope_coeffs_list[i_segment]
spectral_envelope = spectral_envelope_extractor.get_full_spectral_envelope_from_coeffs(
coeffs)
filtered_seg = np.real(np.fft.ifft(white_fft_seg * spectral_envelope))
add_to_segment(i_segment=i_segment,
source=filtered_seg,
dest=filtered_sawtooth,
n_gap=n_gap)
return filtered_sawtooth
def __init__(self, params):
self.pitch_extractor = PitchEstimator(params=params)
self.period_and_noise_separator = PeriodicAndNoiseSeparator(
params=params)
self.spectral_envelope_extractor = SpectralEnvelopeExtractor(
params=params)
self.params = params
class RoughSpectralEnvelopeExtractor:
def __init__(self, params=None):
self.params = deepcopy(params)
self.params["n_triangle_function"] = 15
self.spectral_features_extractor = SpectralEnvelopeExtractor(params)
def get_spectral_envelope_from_sound(self, sound=None):
rough_spectral_envelope_list = self.spectral_features_extractor.get_spectral_envelope_from_sound(sound=sound)
rough_spectral_envelope_and_delta_list = []
for it in range(len(rough_spectral_envelope_list)):
log_rough_spectral_envelope_prev = (get_element_from_list_zero_outside(it, rough_spectral_envelope_list) + 1)
log_rough_spectral_envelope = (get_element_from_list_zero_outside(it-1, rough_spectral_envelope_list) + 1)
derivative_log_rough_spectral_envelope = log_rough_spectral_envelope - log_rough_spectral_envelope_prev
feature_vector = np.concatenate([1*log_rough_spectral_envelope,
10*derivative_log_rough_spectral_envelope], axis=0)
rough_spectral_envelope_and_delta_list.append(feature_vector)
return rough_spectral_envelope_and_delta_list
from feature_extraction.spectral_enveloppe_extractor import SpectralEnvelopeExtractor
from common.count_segments import count_segments
from database_tools.sound_file_loader import get_segment
import matplotlib.pyplot as plt
import numpy as np
from feature_extraction.voice_feature_extractor import voice_feature_extractor
params = get_params()
filename = "/home/monsieur/projets/voice_conversion/data/bruce_willis/Studio/1.wav"
sound, sampling_frequency = get_mono_left_channel_sound_and_sampling_frequency(
filename)
params = get_params()
params["n_triangle_function"] = 40
rough_spectral_envelope_extractor = SpectralEnvelopeExtractor(params=params)
spectrum_as_list = rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
sound)
spectrum = np.zeros([40, len(spectrum_as_list)])
for it in range(len(spectrum_as_list)):
spectrum[:, it] = spectrum_as_list[it]
plt.imshow(np.log(spectrum + 1E-7))
plt.gca().invert_yaxis()
plt.show()
def __init__(self, params=None):
self.params = deepcopy(params)
self.params["n_triangle_function"] = 15
self.spectral_features_extractor = SpectralEnvelopeExtractor(params)
class VoiceFeatureExtractor:
def __init__(self, params):
self.pitch_extractor = PitchEstimator(params=params)
self.period_and_noise_separator = PeriodicAndNoiseSeparator(
params=params)
self.spectral_envelope_extractor = SpectralEnvelopeExtractor(
params=params)
self.params = params
def extract(self, sound=None):
apowin2 = self.params["apowin2"]
n_segment = count_segments(sound=sound, params=self.params)
period_list = []
spectral_envelope_coeffs_noise_list = []
spectral_envelope_coeffs_harmonic_list = []
for i_segment in range(n_segment):
x = get_segment(sound=sound,
i_segment=i_segment,
params=self.params)
x_apodized = x * apowin2
period = self.pitch_extractor.estimate_period(x_apodized)
periodic, noise = self.period_and_noise_separator.separate_components(
x_apodized=x_apodized, period=period)
spectral_envelope_coeffs_periodic = self.spectral_envelope_extractor.get_spectral_envelope_coeffs(
periodic)
spectral_envelope_coeffs_noise = self.spectral_envelope_extractor.get_spectral_envelope_coeffs(
noise)
spectral_envelope_coeffs_noise_list.append(
spectral_envelope_coeffs_noise)
spectral_envelope_coeffs_harmonic_list.append(
spectral_envelope_coeffs_periodic)
period_list.append(period)
return {
"period_list":
period_list,
"spectral_envelope_coeffs_harmonic_list":
spectral_envelope_coeffs_harmonic_list,
"spectral_envelope_coeffs_noise_list":
spectral_envelope_coeffs_noise_list
}
def extract_as_array(self, sound=None):
feature_dict = self.extract(sound=sound)
period_array = np.array(feature_dict["period_list"]).reshape(
[-1, 1]) # shape = [n_segment, 1]
spectral_envelope_coeffs_harmonic_array = np.array(
feature_dict["spectral_envelope_coeffs_harmonic_list"]
) # shape = [n_segment, n_triangle_function]
spectral_envelope_coeffs_noise_array = np.array(
feature_dict["spectral_envelope_coeffs_noise_list"]
) # shape = [n_segment, n_triangle_function]
feature_array = np.concatenate([
period_array, spectral_envelope_coeffs_harmonic_array,
spectral_envelope_coeffs_noise_array
],
axis=1)
return feature_array
from feature_extraction.voice_feature_extractor import voice_feature_extractor
from feature_extraction.spectral_enveloppe_extractor import SpectralEnvelopeExtractor
from database_tools.sound_file_loader import get_mono_left_channel_sound_and_sampling_frequency
from alignment.dynamic_time_warping import get_dtw_matrix_and_corresponding_segments
from params.params import get_params
import numpy as np
from alignment.get_corresponding_segments_as_function import get_corresponding_segments_function
import matplotlib.pyplot as plt
params = get_params()
params["n_triangle_function"] = 15
rough_spectral_envelope_extractor = SpectralEnvelopeExtractor(params=params)
file_1 = "/home/monsieur/projets/voice_conversion/data/bruce_willis/Studio/1.wav"
file_2 = "/home/monsieur/projets/voice_conversion/data/pierre_sendorek/Studio/1.wav"
sound_1, _ = get_mono_left_channel_sound_and_sampling_frequency(file_1)
sound_2, _ = get_mono_left_channel_sound_and_sampling_frequency(file_2)
voice_1_features = rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
sound=sound_1)
voice_2_features = rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
sound=sound_2)
def distance(a, b):
return np.linalg.norm(np.log(a) - np.log(b))