def collect_decibel_statistics(path_listing):
"""
Calculate the average (min, max) values for the decibel values of
both the linear scale magnitude spectrogram's and a mel scale
magnitude spectrogram's of a list of wav files.
Arguments:
path_listing (list):
List of wav file paths.
Returns:
np.ndarray:
Average min and max values of the decibel representations.
Calculation: (avg(linear_min_db), avg(linear_max_db), avg(mel_min_db), avg(mel_max_db)).
"""
# (min_linear, max_linear, min_mel, max_mel)
stats = np.zeros(4)
# Accumulate statistics for a list of wav files.
for path in path_listing:
wav, sampling_rate = load_wav(path)
# Accumulate the calculated min and max values.
stats += decibel_statistics(wav, sampling_rate)
# Calculate the average min and max values.
n_files = len(path_listing)
stats /= n_files
return stats
def load_audio(file_path):
# Window length in audio samples.
win_len = ms_to_samples(model_params.win_len,
model_params.sampling_rate)
# Window hop in audio samples.
hop_len = ms_to_samples(model_params.win_hop,
model_params.sampling_rate)
# Load the actual audio file.
wav, sr = load_wav(file_path.decode())
# TODO: Determine a better silence reference level for the CMU_ARCTIC dataset (See: #9).
# Remove silence at the beginning and end of the wav so the network does not have to learn
# some random initial silence delay after which it is allowed to speak.
wav, _ = librosa.effects.trim(wav)
# Calculate the linear scale spectrogram.
# Note the spectrogram shape is transposed to be (T_spec, 1 + n_fft // 2) so dense layers
# for example are applied to each frame automatically.
linear_spec = linear_scale_spectrogram(wav, model_params.n_fft,
hop_len, win_len).T
# Calculate the Mel. scale spectrogram.
# Note the spectrogram shape is transposed to be (T_spec, n_mels) so dense layers for
# example are applied to each frame automatically.
mel_spec = mel_scale_spectrogram(wav, model_params.n_fft, sr,
model_params.n_mels,
model_params.mel_fmin,
model_params.mel_fmax, hop_len,
win_len, 1).T
# Convert the linear spectrogram into decibel representation.
linear_mag = np.abs(linear_spec)
linear_mag_db = magnitude_to_decibel(linear_mag)
linear_mag_db = normalize_decibel(linear_mag_db,
CMUDatasetHelper.linear_ref_db,
CMUDatasetHelper.linear_mag_max_db)
# => linear_mag_db.shape = (T_spec, 1 + n_fft // 2)
# Convert the mel spectrogram into decibel representation.
mel_mag = np.abs(mel_spec)
mel_mag_db = magnitude_to_decibel(mel_mag)
mel_mag_db = normalize_decibel(mel_mag_db,
CMUDatasetHelper.mel_mag_ref_db,
CMUDatasetHelper.mel_mag_max_db)
# => mel_mag_db.shape = (T_spec, n_mels)
# Tacotron reduction factor.
if model_params.reduction > 1:
mel_mag_db, linear_mag_db = DatasetHelper.apply_reduction_padding(
mel_mag_db, linear_mag_db, model_params.reduction)
return np.array(mel_mag_db).astype(np.float32), \
np.array(linear_mag_db).astype(np.float32)
def collect_duration_statistics(dataset_name, path_listing):
durations = []
print("Collecting duration statistics for {} files ...".format(
len(path_listing)))
for path in path_listing:
# Load the audio file.
wav, sampling_rate = load_wav(path)
# Get the duration in seconds.
duration = get_duration(wav, sampling_rate)
# Collect durations.
durations.append(duration)
durations_sum = sum(durations)
durations_avg = durations_sum / len(durations)
durations_min = min(durations)
durations_max = max(durations)
print("durations_sum: {} sec.".format(durations_sum))
print("durations_avg: {} sec.".format(durations_avg))
print("durations_min: {} sec.".format(durations_min))
print("durations_max: {} sec.".format(durations_max))
from matplotlib import rc
rc('font', **{'family': 'serif', 'serif': ['Computer Modern'], 'size': 13})
rc('text', usetex=True)
# Create a histogram of the individual file durations.
fig = plt.figure(figsize=(1.5 * 14.0 / 2.54, 7.7 / 2.54), dpi=100)
plt.hist(durations, bins=100, normed=False, color="#6C8EBF")
plt.grid(linestyle='dashed')
plt.xlim([0, 21])
# plt.title('"{}" file duration distribution'.format(dataset_name))
plt.xlabel("Duration (seconds)")
plt.ylabel("Count")
plt.show()
# DEBUG: Dump plot into a pdf file.
fig.savefig("/tmp/durations.pdf", bbox_inches='tight')
# DEBUG: Dump statistics into a csv file.
np.savetxt("/tmp/durations.csv",
durations,
delimiter=",",
fmt='%s',
header="duration")
def collect_reconstruction_error(path_listing, n_iters):
mse_errors = []
n_fft = 2048
# Window length in ms.
win_len = 50.0
# Window stride in ms.
win_hop = 12.5
print("Collecting reconstruction statistics for {} files ...".format(
len(path_listing)))
for path in path_listing:
# Load the audio file.
wav, sampling_rate = load_wav(path)
win_len_samples = ms_to_samples(win_len, sampling_rate=sampling_rate)
win_hop_samples = ms_to_samples(win_hop, sampling_rate=sampling_rate)
stft = linear_scale_spectrogram(wav,
win_length=win_len_samples,
hop_length=win_hop_samples,
n_fft=n_fft)
mag = np.abs(stft)
# mag = np.power(mag, 1.2)
_, mse = griffin_lim_v2(spectrogram=mag,
win_length=win_len_samples,
hop_length=win_hop_samples,
n_fft=n_fft,
n_iter=n_iters)
# Collect mean-squared errors.
mse_errors.append(mse)
# For debugging purposes only.
# print('"{}" => iters: {}, mse: {}'.format(path, n_iters, mse))
total_mse = sum(mse_errors) / len(mse_errors)
print('Dataset MSE with {} iterations: {}'.format(n_iters, total_mse))
return total_mse
def plot_liner_mel_spec_comparasion():
ms_win_len = 50.0
ms_win_hop = 12.5
n_fft = 1024
wav_path = '/thesis/datasets/blizzard_nancy/wav/RURAL-02198.wav'
wav, sr = load_wav(wav_path)
win_len = ms_to_samples(ms_win_len, sampling_rate=sr)
hop_len = ms_to_samples(ms_win_hop, sampling_rate=sr)
linear_spec = linear_scale_spectrogram(wav, n_fft, hop_len, win_len).T
mel_spec = mel_scale_spectrogram(wav,
n_fft=n_fft,
sampling_rate=sr,
n_mels=80,
fmin=0,
fmax=sr // 2,
hop_length=hop_len,
win_length=win_len,
power=1).T
# ==================================================================================================
# Convert the linear spectrogram into decibel representation.
# ==================================================================================================
linear_mag = np.abs(linear_spec)
linear_mag_db = magnitude_to_decibel(linear_mag)
# ==================================================================================================
# Convert the mel spectrogram into decibel representation.
# ==================================================================================================
mel_mag = np.abs(mel_spec)
mel_mag_db = magnitude_to_decibel(mel_mag)
rc('font', **{'family': 'serif', 'serif': ['Computer Modern'], 'size': 13})
rc('text', usetex=True)
y_formater = ticker.FuncFormatter(
lambda x, pos: '{:.0f}'.format(x / 1000.0))
linear_mag_db = linear_mag_db[int((0.20 * sr) / hop_len):int((1.85 * sr) /
hop_len), :]
fig = plot_spectrogram(linear_mag_db.T,
sr,
hop_len,
0.0,
sr // 2.0,
'linear',
figsize=((1.0 / 1.35) * (14.0 / 2.54), 7.7 / 2.54),
_formater=y_formater)
fig.savefig("/tmp/linear_spectrogram_raw_mag_db.pdf", bbox_inches='tight')
def __tmp_fmt(x):
if x == 0.0:
return '{:.0f}'.format(x / 1000.0)
elif x < 1000:
return '{:.1f}'.format(x / 1000.0)
else:
return '{:.0f}'.format(math.floor(x / 1000.0))
y_formater = ticker.FuncFormatter(lambda x, pos: __tmp_fmt(x))
mel_mag_db = mel_mag_db[int((0.20 * sr) / hop_len):int((1.85 * sr) /
hop_len), :]
fig = plot_spectrogram(mel_mag_db.T,
sr,
hop_len,
0.0,
sr // 2.0,
'mel',
figsize=((1.025 / 1.35) * (14.0 / 2.54),
7.7 / 2.54),
_formater=y_formater)
fig.savefig("/tmp/mel_spectrogram_raw_mag_db.pdf", bbox_inches='tight')