import os
import random
import numpy as np
from PIL import Image
import fnmatch
from math import ceil
import sox
import tempfile
import shutil
import sys
from pydub import AudioSegment
if (sys.version_info >= (3,0)):
from queue import Queue
else:
from Queue import Queue
NOISE_FILES_LENGTH = [118, 14, 20, 46, 126, 8, 124]
def recursive_glob(path, pattern):
for root, dirs, files in os.walk(path):
for basename in files:
if fnmatch.fnmatch(basename, pattern):
filename = os.path.abspath(os.path.join(root, basename))
if os.path.isfile(filename):
yield filename
class VinylBackgroundSpectrogramGenerator(object):
def __init__(self, source, config, shuffle=False, max_size=100, run_only_once=False):
self.source = source
self.config = config
self.queue = Queue(max_size)
self.shuffle = shuffle
self.run_only_once = run_only_once
if os.path.isdir(self.source):
files = []
files.extend(recursive_glob(self.source, "*.wav"))
files.extend(recursive_glob(self.source, "*.mp3"))
files.extend(recursive_glob(self.source, "*.m4a"))
else:
files = [self.source]
self.files = files
def audioToSpectrogram(self, file, pixel_per_sec, height):
noise_file_index = random.randint(1, len(NOISE_FILES_LENGTH))
noise_file_name = "vinyl_noise/normalized-noise{}.wav".format(noise_file_index)
with tempfile.NamedTemporaryFile(suffix='.wav') as noisy_speech_file:
noise = AudioSegment.from_file(noise_file_name)
speech = AudioSegment.from_file(file)
speech.apply_gain(noise.dBFS - speech.dBFS)
noisy_speech = speech.overlay(noise - 10, loop=True)
noisy_speech.export(noisy_speech_file.name, format="wav")
# shutil.copyfile(noisy_speech_file.name, os.path.join("/extra/tom/news2/debug", "mixed_" + os.path.basename(noisy_speech_file.name)))
with tempfile.NamedTemporaryFile(suffix='.png') as image_file:
command = "{} -n remix 1 rate 10k spectrogram -y {} -X {} -m -r -o {}". format(noisy_speech_file.name, height, pixel_per_sec, image_file.name)
sox.core.sox([command])
# spectrogram can be inspected at image_file.name
image = Image.open(image_file.name)
return np.array(image)
def get_generator(self):
start = 0
while True:
file = self.files[start]
try:
target_height, target_width, target_channels = self.config["input_shape"]
image = self.audioToSpectrogram(file, self.config["pixel_per_second"], target_height)
image = np.expand_dims(image, -1) # add dimension for mono channel
height, width, channels = image.shape
assert target_height == height, "Heigh mismatch {} vs {}".format(target_height, height)
num_segments = width // target_width
for i in range(0, num_segments):
slice_start = i * target_width
slice_end = slice_start + target_width
slice = image[:, slice_start:slice_end]
# Ignore black images
if slice.max() == 0 and slice.min() == 0:
continue
yield slice
except Exception as e:
print("VinylBackgroundSpectrogramGenerator Exception: ", e, file)
pass
finally:
start += 1
if start >= len(self.files):
if self.run_only_once:
break
start = 0
if self.shuffle:
np.random.shuffle(self.files)
def get_num_files(self):
return len(self.files)
if __name__ == "__main__":
a = VinylBackgroundSpectrogramGenerator("/Users/therold/Uni/master-thesis/datasets/EUSpeech/english/input.wav", {"pixel_per_second": 50, "input_shape": [129, 100, 1], "batch_size": 32, "num_classes": 4}, shuffle=True)
gen = a.get_generator()
for a in gen:
pass
