wassname/TTS
1
0
Fork 0
mirror of https://github.com/wassname/TTS.git synced 2026-09-09 11:16:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Batch update after data-loss

Browse Source
This commit is contained in:
Eren Golge
2018-11-02 16:13:51 +01:00
parent f53f9cb360
commit c8a552e627
18 changed files with 1362 additions and 607 deletions
Download Patch File Download Diff File Expand all files Collapse all files
utils/audio.py
+136 -33
View File
@@ -3,26 +3,34 @@ import librosa
import pickle
import copy
import numpy as np
import scipy
from scipy import signal
_mel_basis = None
from pprint import pprint
from scipy import signal, io
class AudioProcessor(object):
def __init__(self,
sample_rate,
num_mels,
min_level_db,
frame_shift_ms,
frame_length_ms,
ref_level_db,
num_freq,
power,
preemphasis,
griffin_lim_iters=None):
bits=None,
sample_rate=None,
num_mels=None,
min_level_db=None,
frame_shift_ms=None,
frame_length_ms=None,
ref_level_db=None,
num_freq=None,
power=None,
preemphasis=None,
signal_norm=None,
symmetric_norm=None,
max_norm=None,
mel_fmin=None,
mel_fmax=None,
clip_norm=True,
griffin_lim_iters=None,
**kwargs):
print(" > Setting up Audio Processor...")
self.bits = bits
self.sample_rate = sample_rate
self.num_mels = num_mels
self.min_level_db = min_level_db
@@ -33,33 +41,79 @@ class AudioProcessor(object):
self.power = power
self.preemphasis = preemphasis
self.griffin_lim_iters = griffin_lim_iters
self.signal_norm = signal_norm
self.symmetric_norm = symmetric_norm
self.mel_fmin = 0 if mel_fmin is None else mel_fmin
self.mel_fmax = mel_fmax
self.max_norm = 1.0 if max_norm is None else float(max_norm)
self.clip_norm = clip_norm
self.n_fft, self.hop_length, self.win_length = self._stft_parameters()
if preemphasis == 0:
print(" | > Preemphasis is deactive.")
print(" | > Audio Processor attributes.")
members = vars(self)
pprint(members)
def save_wav(self, wav, path):
wav_norm = wav * (32767 / max(0.01, np.max(np.abs(wav))))
# librosa.output.write_wav(path, wav_norm.astype(np.int16), self.sample_rate)
scipy.io.wavfile.write(path, self.sample_rate, wav_norm.astype(np.int16))
io.wavfile.write(path, self.sample_rate, wav_norm.astype(np.int16))
def _linear_to_mel(self, spectrogram):
global _mel_basis
if _mel_basis is None:
_mel_basis = self._build_mel_basis()
_mel_basis = self._build_mel_basis()
return np.dot(_mel_basis, spectrogram)
def _mel_to_linear(self, mel_spec):
inv_mel_basis = np.linalg.pinv(self._build_mel_basis())
return np.maximum(1e-10, np.dot(inv_mel_basis, mel_spec))
def _build_mel_basis(self, ):
n_fft = (self.num_freq - 1) * 2
if self.mel_fmax is not None:
assert self.mel_fmax <= self.sample_rate // 2
return librosa.filters.mel(
self.sample_rate, n_fft, n_mels=self.num_mels)
self.sample_rate,
n_fft,
n_mels=self.num_mels,
fmin=self.mel_fmin,
fmax=self.mel_fmax)
def _normalize(self, S):
return np.clip((S - self.min_level_db) / -self.min_level_db, 0, 1)
"""Put values in [0, self.max_norm] or [-self.max_norm, self.max_norm]"""
if self.signal_norm:
S_norm = ((S - self.min_level_db) / - self.min_level_db)
if self.symmetric_norm:
S_norm = ((2 * self.max_norm) * S_norm) - self.max_norm
if self.clip_norm :
S_norm = np.clip(S_norm, -self.max_norm, self.max_norm)
return S_norm
else:
S_norm = self.max_norm * S_norm
if self.clip_norm:
S_norm = np.clip(S_norm, 0, self.max_norm)
return S_norm
else:
return S
def _denormalize(self, S):
return (np.clip(S, 0, 1) * -self.min_level_db) + self.min_level_db
"""denormalize values"""
S_denorm = S
if self.signal_norm:
if self.symmetric_norm:
if self.clip_norm:
S_denorm = np.clip(S_denorm, -self.max_norm, self.max_norm)
S_denorm = ((S_denorm + self.max_norm) * -self.min_level_db / (2 * self.max_norm)) + self.min_level_db
return S_denorm
else:
if self.clip_norm:
S_denorm = np.clip(S_denorm, 0, self.max_norm)
S_denorm = (S_denorm * -self.min_level_db /
self.max_norm) + self.min_level_db
return S_denorm
else:
return S
def _stft_parameters(self, ):
"""Compute necessary stft parameters with given time values"""
n_fft = (self.num_freq - 1) * 2
hop_length = int(self.frame_shift_ms / 1000.0 * self.sample_rate)
win_length = int(self.frame_length_ms / 1000.0 * self.sample_rate)
@@ -92,8 +146,16 @@ class AudioProcessor(object):
S = self._amp_to_db(np.abs(D)) - self.ref_level_db
return self._normalize(S)
def melspectrogram(self, y):
if self.preemphasis != 0:
D = self._stft(self.apply_preemphasis(y))
else:
D = self._stft(y)
S = self._amp_to_db(self._linear_to_mel(np.abs(D))) - self.ref_level_db
return self._normalize(S)
def inv_spectrogram(self, spectrogram):
'''Converts spectrogram to waveform using librosa'''
"""Converts spectrogram to waveform using librosa"""
S = self._denormalize(spectrogram)
S = self._db_to_amp(S + self.ref_level_db) # Convert back to linear
# Reconstruct phase
@@ -102,6 +164,16 @@ class AudioProcessor(object):
else:
return self._griffin_lim(S**self.power)
def inv_mel_spectrogram(self, mel_spectrogram):
'''Converts mel spectrogram to waveform using librosa'''
D = self._denormalize(mel_spectrogram)
S = self._db_to_amp(D + self.ref_level_db)
S = self._mel_to_linear(S) # Convert back to linear
if self.preemphasis != 0:
return self.apply_inv_preemphasis(self._griffin_lim(S**self.power))
else:
return self._griffin_lim(S**self.power)
def _griffin_lim(self, S):
angles = np.exp(2j * np.pi * np.random.rand(*S.shape))
S_complex = np.abs(S).astype(np.complex)
@@ -111,20 +183,17 @@ class AudioProcessor(object):
y = self._istft(S_complex * angles)
return y
def melspectrogram(self, y):
if self.preemphasis != 0:
D = self._stft(self.apply_preemphasis(y))
else:
D = self._stft(y)
S = self._amp_to_db(self._linear_to_mel(np.abs(D))) - self.ref_level_db
return self._normalize(S)
def _stft(self, y):
return librosa.stft(
y=y, n_fft=self.n_fft, hop_length=self.hop_length, win_length=self.win_length)
y=y,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=self.win_length,
)
def _istft(self, y):
return librosa.istft(y, hop_length=self.hop_length, win_length=self.win_length)
return librosa.istft(
y, hop_length=self.hop_length, win_length=self.win_length)
def find_endpoint(self, wav, threshold_db=-40, min_silence_sec=0.8):
window_length = int(self.sample_rate * min_silence_sec)
@@ -134,3 +203,37 @@ class AudioProcessor(object):
if np.max(wav[x:x + window_length]) < threshold:
return x + hop_length
return len(wav)
# WaveRNN repo specific functions
# def mulaw_encode(self, wav, qc):
# mu = qc - 1
# wav_abs = np.minimum(np.abs(wav), 1.0)
# magnitude = np.log(1 + mu * wav_abs) / np.log(1. + mu)
# signal = np.sign(wav) * magnitude
# # Quantize signal to the specified number of levels.
# signal = (signal + 1) / 2 * mu + 0.5
# return signal.astype(np.int32)
# def mulaw_decode(self, wav, qc):
# """Recovers waveform from quantized values."""
# mu = qc - 1
# # Map values back to [-1, 1].
# casted = wav.astype(np.float32)
# signal = 2 * (casted / mu) - 1
# # Perform inverse of mu-law transformation.
# magnitude = (1 / mu) * ((1 + mu) ** abs(signal) - 1)
# return np.sign(signal) * magnitude
def load_wav(self, filename, encode=False):
x, sr = librosa.load(filename, sr=self.sample_rate)
assert self.sample_rate == sr
return x
def encode_16bits(self, x):
return np.clip(x * 2**15, -2**15, 2**15 - 1).astype(np.int16)
def quantize(self, x):
return (x + 1.) * (2**self.bits - 1) / 2
def dequantize(self, x):
return 2 * x / (2**self.bits - 1) - 1