class PairSoundFeature:
def __init__(self, params=None):
bw_range_source = params["bw_range_source"]
fw_range_source = params["fw_range_source"]
bw_range_target = params["bw_range_target"]
assert bw_range_source > 0
assert fw_range_source > 0
# assert bw_range_target > 0
self.params = params
self.voice_feature_extractor = VoiceFeatureExtractor(params=params)
self.rough_spectral_envelope_extractor = RoughSpectralEnvelopeExtractor(
params=params)
self.bw_range_source = bw_range_source
self.fw_range_source = fw_range_source
self.bw_range_target = bw_range_target
self.feature_length = (params["n_triangle_function"] * 2 + 1) * (
bw_range_source + fw_range_source + bw_range_target)
def get_feature_length(self):
return self.feature_length
#
def get_sound_pair_features(self, source_path=None, target_path=None):
source_sound, _ = get_mono_left_channel_sound_and_sampling_frequency(
filename=source_path)
target_sound, _ = get_mono_left_channel_sound_and_sampling_frequency(
filename=target_path)
rough_features_source = self.rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
source_sound)
rough_features_target = self.rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
target_sound)
(dtw_matrix,
corresponding_segment_index_list,
local_distance_matrix) = \
get_dtw_matrix_and_corresponding_segments(rough_features_source,
rough_features_target)
corresponding_segment = get_corresponding_segments_function(
corresponding_segment_index_list)
source_sound_features = self.voice_feature_extractor.extract(
sound=source_sound)
target_sound_features = self.voice_feature_extractor.extract(
sound=target_sound)
n_triangle_function = self.params["n_triangle_function"]
source_feature_array_list = []
target_feature_array_list = []
for i in range(len(source_sound_features["period_list"])):
delta_t = corresponding_segment(i) - corresponding_segment(i - 1)
i_target = int(round(corresponding_segment(i)))
feature_source_array = np.zeros([
n_triangle_function * 2 + 1, self.fw_range_source +
self.bw_range_source + self.bw_range_target
])
feature_target_array = np.zeros([n_triangle_function * 2 + 1 + 1])
# target features
feature_target_array[0] = delta_t
feature_target_array[1] = get_element_from_list_constant_outside(
i_target, target_sound_features["period_list"])
feature_target_array[2: 2 + n_triangle_function] = \
get_element_from_list_constant_outside(i_target, target_sound_features["spectral_envelope_coeffs_harmonic_list"])
feature_target_array[2 + n_triangle_function:2 + 2 * n_triangle_function] = \
get_element_from_list_constant_outside(i_target, target_sound_features["spectral_envelope_coeffs_noise_list"])
# source features
for k in range(-self.bw_range_source, self.fw_range_source):
feature_source_array[0, k + self.bw_range_source] = \
get_element_from_list_constant_outside(i + k, source_sound_features["period_list"])
feature_source_array[1: 1 + n_triangle_function, k + self.bw_range_source] = \
get_element_from_list_constant_outside(i + k, source_sound_features["spectral_envelope_coeffs_harmonic_list"])
feature_source_array[1 + n_triangle_function:1 + 2 * n_triangle_function, k + self.bw_range_source] = \
get_element_from_list_constant_outside(i + k, source_sound_features["spectral_envelope_coeffs_noise_list"])
for k in range(-self.bw_range_target, 0):
# We add sound features from the target speaker from previous time steps
# zero outside : keep this to have a neural network with no burn-in time
feature_source_array[0, - k - 1 + self.fw_range_source + self.bw_range_source] = \
get_element_from_list_zero_outside(i_target + k, target_sound_features["period_list"])
feature_source_array[1: 1 + n_triangle_function, - k - 1 + self.fw_range_source + self.bw_range_source] = \
get_element_from_list_zero_outside(i_target + k, target_sound_features["spectral_envelope_coeffs_harmonic_list"])
feature_source_array[1 + n_triangle_function:1 + 2 * n_triangle_function, - k - 1 + self.fw_range_source + self.bw_range_source] = \
get_element_from_list_zero_outside(i_target + k,
target_sound_features["spectral_envelope_coeffs_noise_list"])
# print(feature_source_array)
source_feature_array_list.append(feature_source_array)
target_feature_array_list.append(feature_target_array)
pair_feature_dict = {}
pair_feature_dict["source_file"] = source_path
pair_feature_dict["target_file"] = target_path
pair_feature_dict[
"source_feature_array_list"] = source_feature_array_list
pair_feature_dict[
"target_feature_array_list"] = target_feature_array_list
return pair_feature_dict
import matplotlib.pyplot as plt
import numpy as np
params = get_params()
filename = os.path.join(params["project_base_path"],
"data/bruce_willis/Studio/1.wav")
#filename = "/home/monsieur/maria.wav"
sound, sampling_frequency = get_mono_left_channel_sound_and_sampling_frequency(
filename)
from feature_extraction.voice_feature_extractor import VoiceFeatureExtractor
voice_feature_extractor = VoiceFeatureExtractor(params)
sound_features = voice_feature_extractor.extract(sound=sound)
spectral_envelope_coeffs_noise_list = sound_features[
"spectral_envelope_coeffs_noise_list"]
spectral_envelope_coeffs_harmonic_list = sound_features[
"spectral_envelope_coeffs_harmonic_list"]
period_list = sound_features["period_list"]
# period_list_bzz = [sampling_frequency/220]*len(period_list)
frequency_list = [sampling_frequency / p for p in period_list]
from synthesis.voice_synthesizer import generate_periodic_sound, generate_periodic_filtered_sound, \
generate_filtered_noise
from scipy.io.wavfile import write
out_sound_noise = generate_filtered_noise(spectral_envelope_coeffs_noise_list,
class PairSoundFeatureOneVsOne:
def __init__(self, params=None):
self.params = params
self.voice_feature_extractor = VoiceFeatureExtractor(params=params)
self.rough_spectral_envelope_extractor = RoughSpectralEnvelopeExtractor(
params=params)
self.feature_length = params["n_triangle_function"] * 2 + 1
def get_feature_length(self):
return self.feature_length
def get_sound_pair_features(self, source_path=None, target_path=None):
source_sound, _ = get_mono_left_channel_sound_and_sampling_frequency(
filename=source_path)
target_sound, _ = get_mono_left_channel_sound_and_sampling_frequency(
filename=target_path)
rough_features_source = self.rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
source_sound)
rough_features_target = self.rough_spectral_envelope_extractor.get_spectral_envelope_from_sound(
target_sound)
(dtw_matrix,
corresponding_segment_index_list,
local_distance_matrix) = \
get_dtw_matrix_and_corresponding_segments(rough_features_source,
rough_features_target)
corresponding_segment = get_corresponding_segments_function(
corresponding_segment_index_list)
source_sound_features = self.voice_feature_extractor.extract(
sound=source_sound)
target_sound_features = self.voice_feature_extractor.extract(
sound=target_sound)
n_triangle_function = self.params["n_triangle_function"]
source_feature_array_list = []
target_feature_array_list = []
for i in range(len(source_sound_features["period_list"])):
i_target = int(round(corresponding_segment(i)))
feature_source_array = np.zeros([self.feature_length])
feature_target_array = np.zeros([self.feature_length])
# target features
feature_target_array[0] = get_element_from_list_constant_outside(
i_target, target_sound_features["period_list"])
feature_target_array[1: 1 + n_triangle_function] = \
get_element_from_list_constant_outside(i_target, target_sound_features["spectral_envelope_coeffs_harmonic_list"])
feature_target_array[1 + n_triangle_function:1 + 2 * n_triangle_function] = \
get_element_from_list_constant_outside(i_target, target_sound_features["spectral_envelope_coeffs_noise_list"])
# source features
feature_source_array[0] = \
get_element_from_list_constant_outside(i, source_sound_features["period_list"])
feature_source_array[1: 1 + n_triangle_function] = \
get_element_from_list_constant_outside(i, source_sound_features["spectral_envelope_coeffs_harmonic_list"])
feature_source_array[1 + n_triangle_function:1 + 2 * n_triangle_function] = \
get_element_from_list_constant_outside(i, source_sound_features["spectral_envelope_coeffs_noise_list"])
source_feature_array_list.append(feature_source_array)
target_feature_array_list.append(feature_target_array)
pair_feature_dict = {}
pair_feature_dict["source_file"] = source_path
pair_feature_dict["target_file"] = target_path
pair_feature_dict[
"source_feature_array_list"] = source_feature_array_list
pair_feature_dict[
"target_feature_array_list"] = target_feature_array_list
return pair_feature_dict