def gain(self, level, mode: 'multiply'):
''' Increases the volume of the audio by a specified amount '''
amplifier = {
'multiply':
lambda d, l: d * l,
'percent':
lambda d, l: d * l / 100.0,
'additive':
lambda d, l: d + l,
'subtractive':
lambda d, l: d - l,
'dB_additive':
lambda d, l: librosa.db_to_amplitude(
(librosa.amplitude_to_db(d) + l)),
'dB_subtractive':
lambda d, l: librosa.db_to_amplitude(
(librosa.amplitude_to_db(d) - l)),
'dB_multiply':
lambda d, l: (10**(l / 20)) * d
}
data = amplifier[mode](self.data, level)
return Audio().populate(
data, self.sample_rate, self.source_path,
self.operations + [_gainOperation(level, mode)])
def pesq_on_batch(y_denoised, ytest, test_phase, sr=16000):
pesqvalue = 1
try:
y_denoised = np.squeeze(y_denoised, axis=3)
y_denoised = np.squeeze(y_denoised, axis=0)
y_denoised = librosa.db_to_amplitude(y_denoised)
ytest = librosa.db_to_amplitude(ytest)
denoised = y_denoised * test_phase
original = ytest * test_phase
denoised = librosa.istft(denoised)
original = librosa.istft(original)
#print(denoised)
#print(original)
denoised = librosa.util.normalize(denoised)
original = librosa.util.normalize(original)
#pesqvalue=pesq(sr, original, denoised, 'wb')
pesqvalue = pesq(sr, original, denoised, 'wb')
#print(pesqvalue)
except:
print("pesq didnt work")
presqvalue = 1
return pesqvalue
def stoi_on_batch(y_denoised,ytest,test_phase,sr=16000):
stoivalue=0
y_denoised = np.squeeze(y_denoised,axis=3)
y_denoised = np.squeeze(y_denoised,axis=0)
y_denoised = librosa.db_to_amplitude(y_denoised)
ytest = librosa.db_to_amplitude(ytest)
denoised = y_denoised*test_phase
original = ytest*test_phase
denoised = librosa.istft(denoised)
original = librosa.istft(original)
#print(denoised)
#print(original)
denoised = librosa.util.normalize(denoised)
original = librosa.util.normalize(original)
#pesqvalue=pesq(sr, original, denoised, 'wb')
stoivalue=stoi( original, denoised,sr, 'wb')
#print(pesqvalue)
#print("stoi didnt work")
#stoivalue=0
return stoivalue
def pesq_from_fft(noisy,phase_noisy,clean,phase_clean,out=False):
"""
Calculate PESQ Metric on stft batch
"""
phase_noisy=np.array(phase_noisy)
noisy=librosa.db_to_amplitude(noisy)
noisy=noisy*phase_noisy
noisy = librosa.istft(noisy)
clean=np.array(clean)
phase_clean=np.array(phase_clean)
clean=librosa.db_to_amplitude(clean)
clean=clean*phase_clean
clean = librosa.istft(clean)
if out==True:
global cn
scipy.io.wavfile.write(path+'\\gvepre\\predictGVE'+str(cn)+'.wav',16000,noisy)
cn=cn+1
sr =16000
pesqvalue=pesq(sr, clean, noisy, 'wb')
#print(pesqvalue)
return pesqvalue
def source_to_distortion(batch_predicted,target_gt):
batch_predicted = librosa.db_to_amplitude(batch_predicted)
target_gt = librosa.db_to_amplitude(target_gt)
distortion = (batch_predicted-target_gt)**2
return 10*np.log10(np.divide(target_gt, distortion, out=(np.ones_like(Noisy))*50, where=distortion!=0))
def extract_audio(
Z, feature,
params): # if normalized Z: unnormalize first, then pass to func.
# convert to audio
if feature == "Stft":
# undo log-magnitude scaling
S = librosa.db_to_amplitude(Z)
# upsample
S = _upsample_fft(S, params["fft_sample_rate"],
params["stft_window_length"])
yhat = librosa.griffinlim(S, hop_length=params["stft_hop_length"])
elif feature == "Mel":
# undo log-power scaling
S = librosa.db_to_power(Z)
yhat = librosa.feature.inverse.mel_to_audio(
S,
sr=params["fft_sample_rate"],
n_fft=params["stft_window_length"],
hop_length=params["stft_hop_length"],
)
elif feature == "Cqt":
# undo log-amplitude scaling
S = librosa.db_to_amplitude(Z)
yhat = librosa.griffinlim_cqt(
S,
sr=params["fft_sample_rate"],
hop_length=params["stft_hop_length"],
fmin=librosa.note_to_hz(params["cqt_min_frequency"]),
)
elif feature == "Mfcc":
yhat = librosa.feature.inverse.mfcc_to_audio(
Z,
n_mels=params["frequency_bins"],
sr=params["fft_sample_rate"],
n_fft=params["stft_window_length"],
hop_length=params["stft_hop_length"],
)
else:
print("Error: feature invalid")
# throw/raise something
return -1
return yhat, params["fft_sample_rate"]
def spectrogram_to_audio_db(data, data_recon, output_dir):
data_np = (data.squeeze(0).to(torch.device("cpu"))).detach().numpy()
data_np = librosa.db_to_amplitude(data_np)
data_griffin_lim = librosa.griffinlim(data_np)
data_recon_np = (data_recon.squeeze(0).to(
torch.device("cpu"))).detach().numpy()
data_recon_np = librosa.db_to_amplitude(data_recon_np)
data_recon_griffin_lim = librosa.griffinlim(data_recon_np)
source_aud_path = output_dir + '_input_' + '.wav'
target_aud_path = output_dir + '_output_' + '.wav'
librosa.output.write_wav(source_aud_path, data_griffin_lim, 16384)
librosa.output.write_wav(target_aud_path, data_recon_griffin_lim, 16384)
return source_aud_path, target_aud_path
def main():
sample_rate = 44100
n_fft = 512
n_frame = 1000
# create arbitrary spectrogram as Numpy array
# for convenience, create spectrogram in range of dB and convert to power
spec = np.random.normal(-20, 2, [1 + n_fft // 2, n_frame])
spec[20:30] += 50
spec = librosa.db_to_amplitude(spec)
# plot created spectrogram
librosa.display.specshow(librosa.amplitude_to_db(spec))
plt.colorbar()
plt.show()
# inverse STFT and playback
wave = librosa.istft(spec)
print('wave shape: {}'.format(wave.shape))
sd.play(wave, samplerate=sample_rate, blocking=True)
def mel_spectro2wav(mel_spectro, preemphasize=hparams.PREEMPHASIZE,
ref_db=hparams.REF_DB,
max_db=hparams.MAX_DB,
n_iter_griffin_lim=hparams.N_ITER_GRIFFIN_LIM,
gl_power=hparams.GL_POWER,
sample_rate=hparams.SAMPLE_RATE,
n_fft=hparams.N_FFT,
n_mels=hparams.SYNTHESIZER_N_MELS,
hop_length=hparams.HOP_LENGTH,
win_length=hparams.WIN_LENGTH,
window=hparams.WINDOW):
mel_spectro = mel_spectro.T
mel_spectro = (np.clip(mel_spectro, 0, 1) * max_db) - max_db + ref_db
amp_mel = librosa.db_to_amplitude(mel_spectro)
inv_mel_basis = np.linalg.pinv(librosa.filters.mel(sample_rate, n_fft=n_fft, n_mels=n_mels))
mag_spectro = np.maximum(1e-10, np.dot(inv_mel_basis, amp_mel))
mag_spectro = mag_spectro ** gl_power
wav = griffin_lim(mag_spectro,
n_iter_griffin_lim=n_iter_griffin_lim,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window)
wav = signal.lfilter([1], [1, -preemphasize], wav)
wav, _ = librosa.effects.trim(wav)
return wav.astype(np.float32)
def random_eq(audio, fs, fraction, order, limits_freq, distribtion):
bands, _, _ = filterbank(audio, fs, fraction, order, limits_freq)
bands = np.stack(bands, axis=0)
coeffs = distribtion(size=(bands.shape[0], 1))
eq_coeffs = librosa.db_to_amplitude(coeffs)
bands *= eq_coeffs
return librosa.util.normalize(np.sum(bands, axis=0))
def fetch_spectrogram_image(row):
""" Fetch spectrogram at specific path from cloud storage """
try:
filename = row['filename'].replace('.mp3', '.wav') + ".png"
# filename = row['filename']
logging.info("Fetching " + filename)
blob = pool_bucket.get_blob(filename)
if not blob:
logging.error(filename + " not found")
tmp = tempfile.NamedTemporaryFile(suffix=".png")
blob.download_to_file(tmp)
tmp.seek(0)
img = np.asarray(Image.open(tmp))
assert blob.metadata['db_min']
assert blob.metadata['db_max']
img_mapped = np.interp(
img, (0, 255),
(int(blob.metadata['db_min']), int(blob.metadata['db_max'])))
cqt = librosa.db_to_amplitude(img_mapped)
cqt = np.flipud(cqt).T
tmp.close()
return [cqt, blob.metadata]
except Exception as e:
logging.error("Could not download " + filename)
logging.error(e)
return
def spectrogram2wav(mag):
'''# Generate wave file from spectrogram'''
# transpose
mag = mag.T
# de-noramlize
mag = (np.clip(mag, 0, 1) * hp.max_db) - hp.max_db + hp.ref_db
# to amplitude
mag = librosa.db_to_amplitude(mag)
# print(np.max(mag), np.min(mag), mag.shape)
# (1025, 812, 16)
# wav reconstruction
wav = griffin_lim(mag)
#wav = butter_bandpass_filter(wav, hp.lowcut, hp.highcut, hp.sr, order=6)
# de-preemphasis
wav = signal.lfilter([1], [1, -hp.preemphasis], wav)
# trim
wav, _ = librosa.effects.trim(wav)
return wav
def postpro_and_gen(S,
phase,
returnS=0,
dBscale=1,
denormalize=1,
complex_phase=0,
clip_phase=0): # T, F
if dBscale:
if denormalize:
# denormalization
S = S * hparams.max_db - hparams.max_db + hparams.ref_db
S = librosa.db_to_amplitude(S)
# pad with 0
Sfull = np.concatenate((S, np.zeros(shape=(S.shape[0], 1))), axis=-1)
if clip_phase:
phase = np.concatenate((phase, np.zeros(shape=(2, phase.shape[1], 1))),
axis=-1)
# generate waveform
wav = genWaveclip(Sfull, phase, complex_phase)
if not returnS:
return wav
else:
return wav, Sfull, phase
def extract_loudness(self, audio):
S = librosa.stft(audio)
power = np.abs(S)**2
p_mean = np.sum(power, axis=0, keepdims=True)
db = librosa.power_to_db(p_mean, ref=np.max(power))
amp = librosa.db_to_amplitude(db)
return amp[0]
def make_test_folder(X_in, X_out, name, step, n_samples=5):
path_to_step_folder = mkdir(path_to_log, str(step + 1))
path_to_step_folder_name = mkdir(path_to_step_folder, name)
for i in range(n_samples):
plt.subplot(2, n_samples, 1 + i)
plt.axis('off')
plt.imshow(X_in[i].reshape(spec_dim, spec_dim), cmap='gray')
imageio.imwrite(
path.join(path_to_step_folder_name,
str(i) + "_real_" + name.split('_')[0] + ".jpg"),
X_in[i].reshape(spec_dim, spec_dim))
im = cv2.imread(
path.join(path_to_step_folder_name,
str(i) + "_real_" + name.split('_')[0] + ".jpg"), -1)
im = im[:fft_len // 2 + 1, :fft_len // 2 + 1]
im = (im * 80.0 / 255.0) - 80.0
im = librosa.db_to_amplitude(im)
y2 = griffinlim(im, hop_length=hop_length)
write(
path.join(path_to_step_folder_name,
str(i) + "_real_" + name.split('_')[0] + ".wav"), 16000,
y2 * 1.5)
# plot translated image
for i in range(n_samples):
plt.subplot(2, n_samples, 1 + n_samples + i)
plt.axis('off')
plt.imshow(X_out[i].reshape(spec_dim, spec_dim), cmap='gray')
imageio.imwrite(
path.join(path_to_step_folder_name,
str(i) + "_generated_" + name.split('_')[2] + ".jpg"),
X_out[i].reshape(spec_dim, spec_dim))
im = cv2.imread(
path.join(path_to_step_folder_name,
str(i) + "_generated_" + name.split('_')[2] + ".jpg"),
-1)
im = im[:fft_len // 2 + 1, :fft_len // 2 + 1]
im = (im * 80.0 / 255.0) - 80.0
im = librosa.db_to_amplitude(im)
y2 = griffinlim(im, hop_length=hop_length)
write(
path.join(path_to_step_folder_name,
str(i) + "_generated_" + name.split('_')[2] + ".wav"),
16000, y2 * 1.5)
# save plot to file
filename1 = '%s_generated_plot_%06d.png' % (name, (step + 1))
plt.savefig(path.join(path_to_step_folder, filename1), dpi=300)
plt.close()
def extend_envelope(envelope_estimated, noise_floor_onset):
# Fit straight line in dB scaling, return exponential, extended envelope.
env_db = librosa.amplitude_to_db(envelope_estimated, ref=1.0)
a = (env_db[noise_floor_onset] - env_db[0]) / np.max(
(noise_floor_onset, 1.))
b = env_db[0]
return librosa.db_to_amplitude(a * np.arange(len(envelope_estimated)) + b)
def inv_spectrogram_librosa(spectrogram, fs, hparams):
"""Converts spectrogram to waveform using librosa"""
S_denorm = _denormalize(spectrogram, hparams)
S = librosa.db_to_amplitude(
S_denorm + hparams.ref_level_db
) # Convert back to linear
# Reconstruct phase
return griffinlim_librosa(S, fs, hparams)
def SDR(original, predicted):
original = librosa.db_to_amplitude(original)
predicted = librosa.db_to_amplitude(predicted)
distortion = predicted - original
original = original**2
distortion = distortion**2
sdr = np.divide(original, distortion)
sdr = np.nan_to_num(sdr, nan=60.0, posinf=60.0, neginf=60.0)
#print(sdr)
sdr = 10 * np.log10(sdr)
return np.mean(sdr)
def SNR(Noisy, Clean, mask=True, Bark=False):
"""Function to Calculate Signal-to-Noise Ratio, mask==True puts out IBM"""
m_ibm = []
if Bark == False:
Noisy = librosa.db_to_amplitude(Noisy)
Clean = librosa.db_to_amplitude(Clean)
N = np.subtract(Noisy, Clean)
m_ibm = 20 * np.log10(
np.divide(Clean, N, out=np.zeros_like(Noisy), where=N != 0))
print("masking output")
if mask == True:
m_ibm = (m_ibm >= 0).astype(int)
return m_ibm
def generateWavFromSTFT(amp, phase, wl, hl):
Mdb_inormed = np.interp(amp, (amp.min(), amp.max()), (-15, 65))
iM = librosa.db_to_amplitude(Mdb_inormed)
iC = iM * np.exp(1j * phase)
iy = librosa.istft(iC, hop_length=hl, win_length=wl)
iy = librosa.util.normalize(iy)
return iy
def spec_to_audio(X_out, name_gen):
imageio.imwrite(path.join(path_to_results, name_gen + ".jpg"),
X_out.reshape(260, 260))
im = cv2.imread(path.join(path_to_results, name_gen + ".jpg"), -1)
im = im[:257, :257]
im = (im * 80.0 / 255.0) - 80.0
im = librosa.db_to_amplitude(im)
y2 = griffinlim(im, hop_length=256)
write(path.join(path_to_results, name_gen + ".wav"), 16000, y2 * 3)
def do_convert(predictor, input_name, logdir2):
convert_s = datetime.datetime.now()
# Load input audio
input_audio, _ = librosa.load(input_name, sr=hp.default.sr, dtype=np.float64)
# Extract F0 from input audio first
input_f0, t_table = pw.dio(input_audio, hp.default.sr)
input_f0 = pw.stonemask(input_audio, input_f0, t_table, hp.default.sr)
# Get MFCC, Spectral Envelope, and Aperiodicity
mfcc = _get_mfcc(input_audio, hp.default.n_fft, hp.default.win_length, hp.default.hop_length)
mfcc = np.expand_dims(mfcc, axis=0)
input_ap = pw.d4c(input_audio, input_f0, t_table, hp.default.sr, fft_size=hp.default.n_fft)
input_sp_en = _get_spectral_envelope(preemphasis(input_audio, coeff=hp.default.preemphasis), hp.default.n_fft)
plt.imsave('./converted/debug/input_sp_en_original.png', input_sp_en, cmap='binary')
input_sp_en = np.expand_dims(input_sp_en, axis=0)
# Convert Spectral Envelope
output_sp_en, ppgs = convert_spectral_envelope(predictor, mfcc, input_sp_en)
output_sp_en = np.squeeze(output_sp_en.astype(np.float64), axis=0)
preproc_s = datetime.datetime.now()
# Denormalization
output_sp_en = denormalize_db(output_sp_en, hp.default.max_db, hp.default.min_db)
# Db to amp
output_sp_en = librosa.db_to_amplitude(output_sp_en)
# Emphasize the magnitude
output_sp_en = np.power(output_sp_en, hp.convert.emphasis_magnitude)
preproc_e = datetime.datetime.now()
preproc_t = preproc_e - preproc_s
print("Pre-Processing time:{}s".format(preproc_t.seconds))
# F0 transformation with WORLD Vocoder
output_f0 = f0_adapt(input_f0, logdir2)
# Synthesize audio and de-emphasize
output_audio = pw.synthesize(output_f0, output_sp_en, input_ap, hp.default.sr)
output_audio = inv_preemphasis(output_audio, coeff=hp.default.preemphasis)
# Saving output_audio to 32-bit Float wav file
output_audio = output_audio.astype(np.float32)
librosa.output.write_wav(path="./converted/"+input_name,y=output_audio,sr=hp.default.sr)
# Saving PPGS data to Grayscale Image and raw binary file
ppgs = np.squeeze(ppgs, axis=0)
plt.imsave('./converted/debug/'+input_name+'.png', ppgs, cmap='binary')
np.save('./converted/debug/'+input_name+'.npy', ppgs)
convert_e = datetime.datetime.now()
convert_time = convert_e - convert_s
print("Total Converting Time:{}s".format(convert_time.seconds))
def mag(self):
if self._mag is None:
if self._D is not None and self._ph is not None:
if self._ref_mag is None:
self._ref_mag = 1.0
self._mag = librosa.db_to_amplitude(self._D, ref=self._ref_mag)
else:
self._mag, self._ph = librosa.core.magphase(self.F)
return self._mag
def SDR(original,predicted):
"""
This function calculated the source-to-distortion ratio on a batchwise stft basis.
Herefore SDR on all tf-units is found and the mean is put out.
"""
original = librosa.db_to_amplitude(original)
predicted = librosa.db_to_amplitude(predicted)
distortion = predicted-original
# power spectrum:
original = original**2
distortion = distortion**2
sdr=np.divide(original,distortion)
# Fixing NAN Values:
sdr= np.nan_to_num(sdr,nan=60.0,posinf=60.0,neginf=60.0)
sdr = 10*np.log10(sdr)
return np.mean(sdr)
def convert_specs_to_audio(self, spectrograms, min_max_values):
signals = []
for spec, min_max_value in zip(spectrograms, min_max_values):
log_spec = spec[:, :, 0]
denorm_log_spec = self.min_max_normalizer.denormalise(
log_spec, min_max_value["min"], min_max_value["max"])
spectr = librosa.db_to_amplitude(denorm_log_spec)
signal = librosa.istft(spectr, hop_length=self.hop_length)
signals.append(signal)
return signals
def img2audio(self, mode): # DECODER
magma_dif_list = self.readMagmaDiff()
mtx_rgb_sum = self.readImg(self.trans_img_path)
mtx_value = self.curver(mtx_rgb_sum, magma_dif_list)
mtx_unit = np.ones(mtx_value.shape)
mtx_db = (mtx_value / np.max(mtx_value) - mtx_unit) * float(
abs(self.limit)) # get "normalized" db matrix
mtx_amp = librosa.db_to_amplitude(mtx_db, ref=1.0)
return self.reconstructer(mtx_amp, mode)
def generateWavFromCQT(amp, phase, sr, hl):
Mdb_inormed = np.interp(amp, (amp.min(), amp.max()), (-15, 65))
iM = librosa.db_to_amplitude(Mdb_inormed)
bins_per_octave = 12 * 12
D = iM * np.exp(1j*phase)
iy = librosa.icqt(C=D, sr=sr, hop_length=hl, bins_per_octave=bins_per_octave)
iy = librosa.util.normalize(iy)
return iy
def pesq_on_batch(y_denoised,ytest,test_phase,sr=16000):
pesqvalue=1
try:
y_denoised = np.squeeze(y_denoised,axis=3)
y_denoised = np.squeeze(y_denoised,axis=0)
y_denoised = librosa.db_to_amplitude(y_denoised)
ytest = librosa.db_to_amplitude(ytest)
denoised = y_denoised*test_phase
original = ytest*test_phase
denoised = librosa.istft(denoised)
original = librosa.istft(original)
#print(denoised)
#print(original)
denoised = librosa.util.normalize(denoised)
original = librosa.util.normalize(original)
#pesqvalue=pesq(sr, original, denoised, 'wb')
pmsqe.init_constants(Fs=sr, Pow_factor=pmsqe.perceptual_constants.Pow_correc_factor_Hann, apply_SLL_equalization=True,
apply_bark_equalization=True, apply_on_degraded=True, apply_degraded_gain_correction=True)
#pesqvalue=pesq(sr, original, denoised, 'wb')
pesqvalue=per_frame_PMSQE(original,denoised)
#print(pesqvalue)
except:
print("pesq didnt work")
presqvalue=1
return pesqvalue
def magnitude_db_and_phase_to_audio(frame_length, hop_length_fft,
stftaudio_magnitude_db, stftaudio_phase):
stftaudio_magnitude_rev = librosa.db_to_amplitude(stftaudio_magnitude_db,
ref=1.0)
audio_reverse_stft = stftaudio_magnitude_rev * stftaudio_phase
audio_reconstruct = librosa.core.istft(audio_reverse_stft,
hop_length=hop_length_fft,
length=frame_length)
return audio_reconstruct
def rebuild_audio_from_spectro_clips(spectrogram_clips, is_dB_format=False):
"""rebuild waveform solely from magnitude spectrogram"""
# audio spectrogram format:
# 1. normal stft spectromgram
# 2. dB-scaled spectrogram log(epilon + S*2)
spectrogram = np.concatenate(spectrogram_clips, axis=1)
if is_dB_format:
spectrogram = librosa.db_to_amplitude(spectrogram)
waveform = librosa.istft(spectrogram,
hop_length=HOP_LEN,
win_length=WIN_LEN)
return waveform
def test_db_to_amplitude():
srand()
NOISE_FLOOR = 1e-6
# Make some noise
x = np.abs(np.random.randn(1000)) + NOISE_FLOOR
db = librosa.amplitude_to_db(x, top_db=None)
x2 = librosa.db_to_amplitude(db)
assert np.allclose(x, x2)
def spectrogram2wav(mag):
'''# Generate wave file from spectrogram'''
# transpose
mag = mag.T
# de-noramlize
mag = (np.clip(mag, 0, 1) * hp.max_db) - hp.max_db + hp.ref_db
# to amplitude
mag = librosa.db_to_amplitude(mag)
# print(np.max(mag), np.min(mag), mag.shape)
# (1025, 812, 16)
# wav reconstruction
wav = griffin_lim(mag)
# de-preemphasis
wav = signal.lfilter([1], [1, -hp.preemphasis], wav)
# trim
wav, _ = librosa.effects.trim(wav)
return wav
def __test(ref):
db = librosa.amplitude_to_db(xp, ref=ref, top_db=None)
xp2 = librosa.db_to_amplitude(db, ref=ref)
assert np.allclose(xp, xp2)