def log_mel_spec_to_audio(log_spec):
"""
Estimate audio signal from log-mel spectrogram
using Griffin-Lim algorithm.
"""
spec = np.exp(log_spec)
return mel_to_audio(spec, n_iter=32, **LIBROSA_SPEC_KWARGS)
def log_melspectrogram_to_audio(mels, sr, hop_length, n_iter):
n_fft = hop_length * 4
# Use floating point
mels = mels.astype(float) / numpy.max(mels)
# inverse of log
# FIXME: totaly fails to converge, leading to a lot of white noise
mels = numpy.exp(mels + 1e-6) - 1.0
print(numpy.min(mels), numpy.max(mels), numpy.mean(mels))
# invert to get waveform
y = inverse.mel_to_audio(mels,
sr=sr,
hop_length=hop_length,
n_fft=n_fft,
n_iter=n_iter)
# Normalize audio amplitude?
print(numpy.min(y), numpy.max(y), numpy.mean(y))
norm_factor = 0.5 / max(numpy.min(y), numpy.max(y))
print('norm', norm_factor)
y = y * norm_factor
return y
def save_sample(self, n=5):
for idxs, labels, data in self.data_loader:
labels = labels[:n]
data = data[:n]
break
x = data.type('torch.FloatTensor').to(self.device)
o_size = x.size()
x = self._reshape(x)
x_recon, *_ = self.model(x)
# Revert the shape
x = x.view(*o_size).detach().cpu().numpy()
x_recon = x_recon.view(*o_size).detach().cpu().numpy()
# Join chunks
x = x.transpose(0, 2, 1, 3).reshape(o_size[0], o_size[2], -1)
x_recon = x_recon.transpose(0, 2, 1,
3).reshape(o_size[0], o_size[2], -1)
for i in range(x.shape[0]):
idx = idxs[i]
fpath = self.data_loader.dataset.path_to_data[idx]
fname = path.splitext(path.basename(fpath))[0]
# save spectrogram
np.save(path.join(self.config.sample_dir, fname + ".npy"), x[i])
np.save(path.join(self.config.sample_dir, fname + "_recon.npy"),
x_recon[i])
try:
# save waveform
au = inverse.mel_to_audio(core.db_to_power(x[i]),
sr=22050,
n_fft=2048,
hop_length=735)
recon_au = inverse.mel_to_audio(core.db_to_power(x_recon[i]),
sr=22050,
n_fft=2048,
hop_length=735)
rate = len(au) // 5
wavfile.write(
path.join(self.config.sample_dir, fname + ".wav"), rate,
au)
wavfile.write(
path.join(self.config.sample_dir, fname + "_recon.wav"),
rate, recon_au)
except Exception as e:
self.logger.debug(str(e))
self.logger.info("Cannot save from mel to audio")
def save_wav(spectrogram: t.Any, path: t.Union[str, Path]) -> None:
signal = mel_to_audio(spectrogram)
write_wav(path, signal, sr=22050)
def __call__(self, x: t.Any) -> t.Any:
return mel_to_audio(x, sr=self.sr)
from librosa.feature import inverse
import numpy as np
import torch
from scipy.io import wavfile
if __name__ == '__main__':
mel = np.load('data/mel/p225/p225_013.npy').astype(np.float32)
audio_arr = inverse.mel_to_audio(mel, sr=48000, n_iter=10)
wavfile.write('test.wav', 48000, audio_arr)
def pointwise_mel_to_audio(x):
return mel_to_audio(x,
sr=melspectrogram_transform_parameters["sample_rate"],
n_fft=melspectrogram_transform_parameters["n_fft"],
hop_length=melspectrogram_transform_parameters["hop_length"],
n_iter=4)