def display_spectrogram():
"""
Function used to generate and display sample spectrogram from audio files.
"""
paths = prep_utils.get_absolute_file_paths(AUDIO_CHUNKS_20S_DIR)[:3]
for path in paths:
y, sr = librosa.load(path)
# Decompose a spectrogram with NMF
# Short-time Fourier transform underlies most analysis.
# librosa.stft returns a complex matrix D.
# D[f, t] is the FFT value at frequency f, time (frame) t.
D = librosa.stft(y)
# Separate the magnitude and phase and only use magnitude
S, phase = librosa.magphase(D)
print("S Shape: ", S.shape)
melspec_log = librosa.feature.melspectrogram(S=np.log(S), sr=sr)
print("MelSpec Shape: ", melspec_log.shape)
plt.figure()
librosa.display.specshow(melspec_log, y_axis='mel', x_axis='time')
plt.colorbar()
plt.show()
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 downsample():
"""
Downsample and resize to 256x256, and save them locally
"""
paths = prep_utils.get_absolute_file_paths(PROCESSED_STFT_DIR)
for path in paths:
S = np.load(path)
S_downsample = skimage.transform.resize(S, (256, 256),
anti_aliasing=True)
out = RESIZED_STFT_DIR + prep_utils.get_filename(path) + ".npy"
np.save(out, S_downsample)
def audio_reconstruction():
"""
Function used to reconstruct sample audio clips from STFT matrices, and save audio to file.
"""
paths = prep_utils.get_absolute_file_paths(STFT_ARRAY_DIR)
for path in paths:
S = np.load(path)
y = librosa.griffinlim(S)
out = AUDIO_OUT_DIR + prep_utils.get_filename(path) + ".wav"
# Save reconstructed data
scipy.io.wavfile.write(out, 22050, y)
def record_mean_std():
"""
Record mean and std of all STFT matrices and save them locally
"""
paths = prep_utils.get_absolute_file_paths(STFT_ARRAY_DIR)
mean_list = []
std_list = []
for path in paths:
S = np.load(path)
S = np.log(S)
mag_mean = np.mean(S)
mag_std = np.std(S)
mean_list.append(mag_mean)
std_list.append(mag_std)
print("Finished:", path)
data = {"mean": mean_list, "std": std_list, "path": paths}
df = pd.DataFrame.from_dict(data)
df.to_csv("./data/saved_mean_std.csv")
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)