def decibel_statistics(wav, sampling_rate):
"""
Calculate (min, max) values for the decibel values of both
the linear scale magnitude spectrogram and a
mel scale magnitude spectrogram.
Arguments:
wav (np.ndarray):
Audio time series.
The shape is expected to be shape=(n,).
sampling_rate (int):
Sampling rate using in the calculation of `wav`.
Returns:
np.ndarray:
Min and max values of the decibel representations.
Calculation: np.array[min(linear_db), max(linear_db), min(mel_db), max(mel_db)]
"""
n_fft = 1024
hop_length = n_fft // 4
win_length = n_fft
n_mels = 80
# Get the linear scale spectrogram.
linear_spec = linear_scale_spectrogram(wav,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length)
# Get the mel scale spectrogram.
mel_spec = mel_scale_spectrogram(wav,
n_fft=n_fft,
sampling_rate=sampling_rate,
n_mels=n_mels,
fmin=0,
fmax=sampling_rate // 2,
hop_length=hop_length,
win_length=win_length,
power=1)
# 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, 20, 100)
# 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, -7.7, 95.8)
return np.array([
np.min(linear_mag_db),
np.max(linear_mag_db),
np.min(mel_mag_db),
np.max(mel_mag_db)
])
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 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')