Medix-AI/coqui-tts
Medix-AI
/
coqui-tts
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

audio_lws update, suppress warnings

This commit is contained in:
Eren G 2018-08-02 14:30:05 +02:00
parent ed56dd68f2
commit 3238ffa3e6
1 changed files with 25 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
utils/audio_lws.py
View File

@ -13,6 +13,7 @@ class AudioProcessor(object):
def __init__(self, sample_rate, num_mels, min_level_db, frame_shift_ms, def __init__(self, sample_rate, num_mels, min_level_db, frame_shift_ms,
frame_length_ms, ref_level_db, num_freq, power, preemphasis, frame_length_ms, ref_level_db, num_freq, power, preemphasis,
min_mel_freq, max_mel_freq, griffin_lim_iters=None, ): min_mel_freq, max_mel_freq, griffin_lim_iters=None, ):
print(" > Setting up Audio Processor...")
self.sample_rate = sample_rate self.sample_rate = sample_rate
self.num_mels = num_mels self.num_mels = num_mels
self.min_level_db = min_level_db self.min_level_db = min_level_db
@ -25,8 +26,7 @@ class AudioProcessor(object):
self.max_mel_freq = max_mel_freq self.max_mel_freq = max_mel_freq
self.griffin_lim_iters = griffin_lim_iters self.griffin_lim_iters = griffin_lim_iters
self.preemphasis =preemphasis self.preemphasis =preemphasis
_, self.hop_length, self.win_length = self._stft_parameters() self.n_fft, self.hop_length, self.win_length = self._stft_parameters()
self.num_freq = (self.win_length // 2) + 1
if preemphasis == 0: if preemphasis == 0:
print(" | > Preemphasis is deactive.") print(" | > Preemphasis is deactive.")
@ -35,14 +35,21 @@ class AudioProcessor(object):
librosa.output.write_wav(path, wav.astype(np.float), self.sample_rate, norm=True) librosa.output.write_wav(path, wav.astype(np.float), self.sample_rate, norm=True)
def _stft_parameters(self, ): def _stft_parameters(self, ):
n_fft = (self.num_freq - 1) * 2 n_fft = int((self.num_freq - 1) * 2)
hop_length = int(self.frame_shift_ms / 1000.0 * self.sample_rate) hop_length = int(self.frame_shift_ms / 1000.0 * self.sample_rate)
win_length = int(self.frame_length_ms / 1000.0 * self.sample_rate) win_length = int(self.frame_length_ms / 1000.0 * self.sample_rate)
return n_fft, hop_length, win_length if n_fft % hop_length != 0:
hop_length = n_fft / 8
print(" | > hop_length is set to default ({}).".format(hop_length))
if n_fft % win_length != 0:
win_length = n_fft / 2
print(" | > win_length is set to default ({}).".format(win_length))
print(" | > fft size: {}, hop length: {}, win length: {}".format(n_fft, hop_length, win_length))
return int(n_fft), int(hop_length), int(win_length)
def _lws_processor(self): def _lws_processor(self):
try: try:
return lws.lws(self.win_length, self.hop_length ,mode="speech") return lws.lws(self.win_length, self.hop_length, fftsize=self.n_fft, mode="speech")
except: except:
raise RuntimeError(" !! WindowLength({}) is not multiple of HopLength({}).".format(self.win_length, self.hop_length)) raise RuntimeError(" !! WindowLength({}) is not multiple of HopLength({}).".format(self.win_length, self.hop_length))
@ -70,15 +77,22 @@ class AudioProcessor(object):
return signal.lfilter([1], [1, -self.preemphasis], x) return signal.lfilter([1], [1, -self.preemphasis], x)
def spectrogram(self, y): def spectrogram(self, y):
f = open(os.devnull, 'w')
old_out = sys.stdout
sys.stdout = f
if self.preemphasis: if self.preemphasis:
D = self._lws_processor().stft(self.apply_preemphasis(y)).T D = self._lws_processor().stft(self.apply_preemphasis(y)).T
else: else:
D = self._lws_processor().stft(y).T D = self._lws_processor().stft(y).T
S = self._amp_to_db(np.abs(D)) - self.ref_level_db S = self._amp_to_db(np.abs(D)) - self.ref_level_db
sys.stdout = old_out
return self._normalize(S) return self._normalize(S)
def inv_spectrogram(self, spectrogram): def inv_spectrogram(self, spectrogram):
'''Converts spectrogram to waveform using librosa''' '''Converts spectrogram to waveform using librosa'''
f = open(os.devnull, 'w')
old_out = sys.stdout
sys.stdout = f
S = self._denormalize(spectrogram) S = self._denormalize(spectrogram)
S = self._db_to_amp(S + self.ref_level_db) # Convert back to linear S = self._db_to_amp(S + self.ref_level_db) # Convert back to linear
processor = self._lws_processor() processor = self._lws_processor()
@ -87,6 +101,7 @@ class AudioProcessor(object):
# Reconstruct phase # Reconstruct phase
if self.preemphasis: if self.preemphasis:
return self.apply_inv_preemphasis(y) return self.apply_inv_preemphasis(y)
sys.stdout = old_out
return y return y
def _linear_to_mel(self, spectrogram): def _linear_to_mel(self, spectrogram):
@ -96,13 +111,17 @@ class AudioProcessor(object):
return np.dot(_mel_basis, spectrogram) return np.dot(_mel_basis, spectrogram)
def _build_mel_basis(self, ): def _build_mel_basis(self, ):
return librosa.filters.mel(self.sample_rate, self.win_length, n_mels=self.num_mels) return librosa.filters.mel(self.sample_rate, self.n_fft, n_mels=self.num_mels)
# fmin=self.min_mel_freq, fmax=self.max_mel_freq) # fmin=self.min_mel_freq, fmax=self.max_mel_freq)
def melspectrogram(self, y): def melspectrogram(self, y):
f = open(os.devnull, 'w')
old_out = sys.stdout
sys.stdout = f
if self.preemphasis: if self.preemphasis:
D = self._lws_processor().stft(self.apply_preemphasis(y)).T D = self._lws_processor().stft(self.apply_preemphasis(y)).T
else: else:
D = self._lws_processor().stft(y).T D = self._lws_processor().stft(y).T
S = self._amp_to_db(self._linear_to_mel(np.abs(D))) - self.ref_level_db S = self._amp_to_db(self._linear_to_mel(np.abs(D))) - self.ref_level_db
sys.stdout = old_out
return self._normalize(S) return self._normalize(S)