def convert_audio_to_stft(src_dir, dest_dir, extension):
"""
Function used to convert audio clips into Short-Time Fourier Transform matrices, and save matrices to files.
:param src_dir: input audio directory
:param dest_dir: output STFT directory
:param extension: desired output file type
"""
paths = prep_utils.get_unprocessed_items(src_dir=src_dir,
dest_dir=dest_dir)
start_time = time.time()
for path in paths:
prep_utils.display_progress_eta(current_item=path,
total_items=paths,
start_time=start_time)
y, sr = librosa.load(path)
# Decompose a spectrogram with NMF
D = librosa.stft(y)
# Separate the magnitude and phase and only use magnitude
S, _ = librosa.magphase(D)
out = dest_dir + prep_utils.get_filename(path) + extension
np.save(out, S)
def make_audio_chunks(seconds, dest_dir):
"""
Function used to convert audio into shorter audio clips, and save audio clips to files.
:param seconds: desired clip length
:param dest_dir: output directory
"""
paths = prep_utils.get_absolute_file_paths(DATASET_DIR, ".wav")
start_time = time.time()
for audio_path in paths:
prep_utils.display_progress_eta(current_item=audio_path,
total_items=paths,
start_time=start_time)
audio = AudioSegment.from_file(audio_path)
chunk_length_ms = seconds * 1000 # 20 seconds
chunks = make_chunks(audio, chunk_length_ms)
chunks.pop(-1)
# Export all of the individual chunks as wav files
for i, chunk in enumerate(chunks):
_, chunk_name = os.path.split(
os.path.splitext(audio_path)[0] + "_chunk_{0}.wav".format(i))
chunk.export(dest_dir + chunk_name, format="wav")
print("\n\nChunks export completed.")
def convert_stft_to_images_grayscale(src_dir, dest_dir, ext=".png", size=None):
"""
Convert STFT matrices into grayscale images
:param src_dir: source directory that stores STFT matrices
:param dest_dir: destination where converted images are being saved
:param ext: image extension
:param size: resize dimension
"""
paths = prep_utils.get_unprocessed_items(src_dir=src_dir,
dest_dir=dest_dir)
start_time = time.time()
for path in paths:
prep_utils.display_progress_eta(current_item=path,
total_items=paths,
start_time=start_time)
S = np.load(path)
S_scaled = prep_utils.increase_brightness(S)
if size:
S_scaled = cv2.resize(S_scaled, (size, size),
interpolation=cv2.INTER_CUBIC)
out_path = dest_dir + prep_utils.get_filename(path) + ext
cv2.imwrite(out_path, S_scaled)
def convert_stft_to_images(src_dir, dest_dir, ext=".png", size=None):
"""
Function used to convert STFT matrices to images, and saves them to destination folder
:param src_dir: source folder where STFT matrices are stored
:param dest_dir: output images folder
:param ext: image format, defaulted to .png
:param size: dimension of desired square image
"""
paths = prep_utils.get_unprocessed_items(src_dir=src_dir,
dest_dir=dest_dir)
start_time = time.time()
for path in paths:
prep_utils.display_progress_eta(current_item=path,
total_items=paths,
start_time=start_time)
S_norm = np.load(path)
S_norm = normalize_stft(S_norm)
if size:
S_norm = cv2.resize(S_norm, (size, size),
interpolation=cv2.INTER_CUBIC)
out_path = dest_dir + prep_utils.get_filename(path) + ext
plt.imsave(out_path, S_norm)
image = cv2.imread(out_path)
cv2.imwrite(out_path, image)
def audio_reconstruction_test(src_dir, dest_dir, ext=".png", size=None):
"""
Test different approaches to image to audio conversion
:param src_dir: image directory
:param dest_dir: audio directory
:param ext: image type
:param size: desired dimension for resizing
"""
paths = prep_utils.get_unprocessed_items(src_dir=src_dir,
dest_dir=dest_dir)
start_time = time.time()
for path in paths:
prep_utils.display_progress_eta(current_item=path,
total_items=paths,
start_time=start_time)
S = np.load(path)
cv2.imshow("image", S)
cv2.waitKey(0)
pd.DataFrame(S).to_csv(dest_dir + "S.csv", header=None, index=False)
S_scaled = prep_utils.increase_brightness(S)
cv2.imshow("image", S_scaled)
cv2.waitKey(0)
pd.DataFrame(S_scaled).to_csv(dest_dir + "S_scaled.csv",
header=None,
index=False)
out_path = dest_dir + "gray" + ext
cv2.imwrite(out_path, S_scaled)
S = cv2.imread(out_path, 0)
S = np.array(S, dtype=np.float32)
S_recovered = S
if size:
S_recovered = cv2.resize(S_recovered, (size, size),
interpolation=cv2.INTER_CUBIC)
pd.DataFrame(S_recovered).to_csv(dest_dir + "S_recovered.csv",
header=None,
index=False)
out_path = dest_dir + "resized" + ext
cv2.imwrite(out_path, S_recovered)
S_audio = np.genfromtxt(dest_dir + "S_recovered.csv", delimiter=',')
S_audio = np.array(S_audio, dtype=np.float32)
S_audio = cv2.resize(S_audio, (431, 1025),
interpolation=cv2.INTER_CUBIC)
y = librosa.griffinlim(S_audio)
out = dest_dir + "s.wav"
# Save reconstructed data
scipy.io.wavfile.write(out, 22050, y)
break
def audio_reconstruction_stylegan(src_dir,
dest_dir,
resize_h,
resize_w,
mode="RGB"):
"""
Image to Audio reconstruction post StyleGAN image generation.
:param src_dir: directory of fake images generated by StyleGAN
:param dest_dir: destination directory where converted audio will be saved
:param resize_h: height of the desired image dimension
:param resize_w: width of the desired image dimension
:param mode: "RGB" or "grayscale", generated image type by StyleGAN
"""
src_dir, sub_dir = ar_utils.select_images_iteration(directory=src_dir)
paths = prep_utils.get_absolute_file_paths(src_dir)
out_dir = dest_dir + sub_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
start_time = time.time()
for path in paths:
prep_utils.display_progress_eta(current_item=path,
total_items=paths,
start_time=start_time)
out_path = out_dir + prep_utils.get_filename(path)
if mode == "RGB":
image = cv2.imread(path)
image_gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
S_recovered = np.array(image_gray, dtype=np.float32)
S_recovered = cv2.resize(S_recovered, (resize_w, resize_h),
interpolation=cv2.INTER_CUBIC)
S = (S_recovered - np.min(S_recovered)) / (
np.max(S_recovered) - np.min(S_recovered)) * 2 - 1
pd.DataFrame(S).to_csv(out_dir + prep_utils.get_filename(path) +
"_norm.csv",
header=None,
index=False)
plt.imsave(out_dir + prep_utils.get_filename(path) + "_norm.png",
S)
S = ar_utils.unnormalize_stft(s=S)
pd.DataFrame(S).to_csv(out_dir + prep_utils.get_filename(path) +
"_reconstruct.csv",
header=None,
index=False)
plt.imsave(
out_dir + prep_utils.get_filename(path) + "_reconstruct.png",
S)
y = librosa.griffinlim(S)
out = out_dir + prep_utils.get_filename(path) + ".wav"
scipy.io.wavfile.write(out, 22050, y)
rate, data = scipy.io.wavfile.read(out)
reduced_noise = nr.reduce_noise(audio_clip=data,
noise_clip=data,
verbose=False)
out = out_dir + prep_utils.get_filename(path) + "_nr.wav"
sf.write(out, reduced_noise, rate)
elif mode == "grayscale":
S = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
S = np.array(S, dtype=np.float32)
S_recovered = S
pd.DataFrame(S_recovered).to_csv(out_path + "_original.csv",
header=None,
index=False)
cv2.imwrite(out_path + "_original.png", S_recovered)
S_recovered = cv2.resize(S_recovered, (resize_w, resize_h),
interpolation=cv2.INTER_CUBIC)
S_recovered = ar_utils.decrease_brightness(S_recovered)
pd.DataFrame(S_recovered).to_csv(out_path + "_recovered.csv",
header=None,
index=False)
cv2.imwrite(out_path + "_recovered.png", S_recovered)
y = librosa.griffinlim(S_recovered)
out = out_dir + prep_utils.get_filename(path) + "_recovered.wav"
scipy.io.wavfile.write(out, 22050, y)
rate, data = scipy.io.wavfile.read(out)
reduced_noise = nr.reduce_noise(audio_clip=data,
noise_clip=data,
verbose=False)
out = out_dir + prep_utils.get_filename(path) + "_nr.wav"
sf.write(out, reduced_noise, rate)