def load_spec(audio_path, mode=0):
"""
takes audio path and mode to return various audio 2D representation with the
actual audio and sample rate as tensor
use mode=1 to get melspectrogram
and mode=2 to get mfcc
Default mode=0 for Spectrogram
"""
wave, sr = librosa.load(audio_path, sr=None, mono=True)
# trim silent edges(below 60 db by default), change the threashold by passing `top_db`
# The threshold (in decibels) below reference to consider as silence (default : 60 db)
s, _ = librosa.effects.trim(wave, top_db=60)
# convert to tensor
wave = torch.FloatTensor(s).unsqueeze(0)
# generate (mel)spectrogram / mfcc
if (mode == 1):
# s = torchaudio.transforms.MelSpectrogram(sample_rate=sr)(wave)
s = librosa.feature.melspectrogram(y=s, sr=sr, hop_length=512)
elif (mode == 2):
# s = torchaudio.transforms.MFCC(sample_rate=sr)(wave)
s = librosa.feature.mfcc(y=s, sr=sr, n_mfcc=40)
else:
# s = torchaudio.transforms.Spectrogram()(wave)
freqs, times, s = librosa.reassigned_spectrogram(y=s,
sr=sr,
hop_length=512)
s = torch.FloatTensor(s).unsqueeze(0)
return s, wave, sr
def get_spectrogram(files):
Spec = []
for filename in files:
y = np.load(filename.replace('.wav', '.npy'))
sr = GLOBAL_SAMPLING_RATE
freqs, times, mags = librosa.reassigned_spectrogram(y, sr)
mags_db = librosa.power_to_db(mags, ref=np.max)
Spec.append(mags_db)
return Spec
def reassigned_spectrogram(y, fs, hparams):
freqs, times, mags = librosa.reassigned_spectrogram(
y=preemphasis(y, hparams),
sr=fs,
n_fft=hparams.n_fft,
hop_length=int(hparams.hop_length_ms / 1000 * fs),
win_length=int(hparams.win_length_ms / 1000 * fs),
center=False,
)
S = librosa.amplitude_to_db((freqs > 0) * (times > 0) * mags,
ref=hparams.ref_level_db)
S = _normalize(S, hparams)
return S
def stft_reassign_from_sig(
sig_wf: np.ndarray, frequency_sample_rate_hz: float, band_order_Nth: float
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray,
np.ndarray]:
"""
Librosa STFT is complex FFT grid, not power
Reassigned frequencies are not the same as the standard mesh frequencies
:param sig_wf: array with input signal
:param frequency_sample_rate_hz: sample rate of frequency in Hz
:param band_order_Nth: Nth order of constant Q bands
:return: six numpy ndarrays with STFT, STFT_bits, time_stft_s, frequency_stft_hz, time_stft_rsg_s,
frequency_stft_rsg_hz
"""
sig_duration_s = len(sig_wf) / frequency_sample_rate_hz
_, min_frequency_hz = scales.from_duration(band_order_Nth, sig_duration_s)
order_Nth, cycles_M, quality_Q, \
frequency_center, frequency_start, frequency_end = \
scales.frequency_bands_g2f1(scale_order_input=band_order_Nth,
frequency_low_input=min_frequency_hz,
frequency_sample_rate_input=frequency_sample_rate_hz)
# Choose the spectral resolution as the key parameter
frequency_resolution_min_hz = np.min(frequency_end - frequency_start)
frequency_resolution_max_hz = np.max(frequency_end - frequency_start)
frequency_resolution_hz_geo = np.sqrt(frequency_resolution_min_hz *
frequency_resolution_max_hz)
stft_time_duration_s = 1 / frequency_resolution_hz_geo
stft_points_per_seg = int(frequency_sample_rate_hz * stft_time_duration_s)
# From CQT
stft_points_hop, _, _, _, _ = \
scales.cqt_frequency_bands_g2f1(band_order_Nth,
min_frequency_hz,
frequency_sample_rate_hz,
is_power_2=False)
print('Reassigned STFT Duration, NFFT, HOP:', len(sig_wf),
stft_points_per_seg, stft_points_hop)
STFT_Scaling = 2 * np.sqrt(np.pi) / stft_points_per_seg
# Reassigned frequencies require a 'best fit' solution.
frequency_stft_rsg_hz, time_stft_rsg_s, STFT_mag = \
librosa.reassigned_spectrogram(sig_wf, sr=frequency_sample_rate_hz,
n_fft=stft_points_per_seg,
hop_length=stft_points_hop, win_length=None,
window='hann', center=False, pad_mode='reflect')
# Must be scaled to match scipy psd
STFT_mag *= STFT_Scaling
STFT_bits = utils.log2epsilon(STFT_mag)
# Standard mesh times and frequencies for plotting - nice to have both
time_stft_s = librosa.times_like(STFT_mag,
sr=frequency_sample_rate_hz,
hop_length=stft_points_hop)
frequency_stft_hz = librosa.core.fft_frequencies(
sr=frequency_sample_rate_hz, n_fft=stft_points_per_seg)
# Reassigned frequencies are not the same as the standard mesh frequencies
return STFT_mag, STFT_bits, time_stft_s, frequency_stft_hz, time_stft_rsg_s, frequency_stft_rsg_hz
def tfr_multi(y_multi):
y, sr = y_multi
return librosa.reassigned_spectrogram(y, fill_nan=True)
