def prepare_file2():
filepath = "./noise_raw/degradations/*.wav"
print("Loading: ambient sounds")
data, sr = load_and_concat(filepath)
print("Extracting features: " + filepath)
features = extract_features_melspec(data, sr)
print(features.shape)
np.save("./noise/ambient-sounds", features)
def prepare_file():
filepath = "./noise_raw/ambient-silence.wav"
print("Loading: " + filepath)
data, sr = librosa.load(filepath)
print("Extracting features: " + filepath)
features = extract_features_melspec(data, sr)
print(features.shape)
np.save("./noise/ambient-silence", features)
def prepare_librispeech():
libredir = "./voice_raw/librispeech/dev-clean/"
speaker_count = 0
for speaker in os.listdir(libredir):
speaker_dir = libredir + speaker + "/"
if os.path.isdir(speaker_dir):
concated, sr = load_and_concat(speaker_dir + "/**/*.flac")
features = extract_features_melspec(concated, sr)
write_to_disk(features, speaker_count)
print("Extracted features - LIBRE - speaker: %i" % speaker_count)
speaker_count += 1
def main():
audio_filename = "./samples/speech-test.wav"
data, sr = librosa.load(audio_filename)
print("DATA", data.shape, data.shape[0] / sr)
features = extract_features_melspec(data, sr)
plt.figure(figsize=(12, 4))
librosa.display.specshow(features, sr=sr, x_axis='time', y_axis='mel')
plt.title('Mel-frekvenču spektrogramma')
plt.colorbar(format='%+02.0f dB')
plt.tight_layout()
plt.show()
def prepare_timit():
speaker_count = 0
for dataset in datasets:
dataset_dir = basepath + dataset + "/"
for group in os.listdir(dataset_dir):
group_dir = dataset_dir + group + "/"
if os.path.isdir(group_dir):
for speaker in os.listdir(group_dir):
speaker_dir = group_dir + speaker + "/"
if os.path.isdir(speaker_dir):
concated, sr = load_and_concat(speaker_dir + "/*.WAV")
with_degraded = degrade(concated, sr)
features = extract_features_melspec(with_degraded, sr)
write_to_disk(features, speaker_count)
print("Extracted features - TIMIT - speaker: %i" %
speaker_count)
speaker_count += 1
def prepare_urbansounds():
features = []
files = glob.glob("./noise_raw/urbansounds/data/**/*.wav")
for i, file in enumerate(files):
print(str(i + 1) + " loading: " + file)
try:
data, _sr = librosa.load(file)
except:
continue
sr = _sr
f = extract_features_melspec(data, sr)
features.append(f)
if i > 0 and i % 100 == 0:
np.save("./noise/vad_noise_" + str(i), flatten(features))
shuffle(features)
features = flatten(features)
print(features.shape)
np.save("./noise/vad_noise", features)
from keras.models import load_model
import matplotlib.pyplot as plt
import datetime
from utils import extract_features_melspec, extract_features_mfcc, flatten
name = "speech"
run = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M")
# audio_filename = "noise_raw/degradations/applause.wav"
audio_filename = "./samples/speech-test.wav"
data, sr = librosa.load(audio_filename)
print("SAMPLE RATE", sr)
print("DATA SHAPE", data.shape)
features = extract_features_melspec(data, sr)
np.save("./samples/" + name + "_" + run, features)
print("FEATURES SHAPE", features.shape)
model = load_model('models/vad2_2018-05-25_13-32/model_vad2.33.hdf5')
timeseries_length = 100
hop_length = 25
length = 0
remainder = len(features)
while remainder >= timeseries_length:
length += 1
remainder -= hop_length
x = np.ndarray((length, timeseries_length, features.shape[1]))
def prepare_ljspeech():
data, sr = load_and_concat("./voice_raw/ljspeech/wavs/LJ01*.wav")
features = extract_features_melspec(data, sr)
write_to_disk(features, 0)
import numpy as np
from matplotlib import pyplot as plt
import librosa
from postprocess_utils import seg_metrics
from utils import extract_features_melspec
audio_filename = "./samples/seg-test16.wav"
features_filename = "./samples/seg-test_features.npy"
# predictions_filename = "samples/predictions_2018-05-24_17-48.npy"
audio, sr = librosa.load(audio_filename, sr=16000)
# predictions = np.load(predictions_filename)
# features = np.load(features_filename)
features = extract_features_melspec(audio, sr)
print("AUDIO", audio.shape)
# print("PREDICTIONS", predictions.shape)
print("FEATURES", features.shape)
timeseries_length = 100
hop_length = 25
# preds = deoverlap_predictions(predictions, features, hop_length)
# norm_preds = defragment_vad(preds)
# reference = [(6.42, 6.85), (13.49, 13.78)]
reference = [(0, 6.42), (6.42, 13.49), (13.49, 20.43)]
# lium = [(13.55, 13.67)]
lium = [(0, 13.55), (13.55, 20.43)]