Python hparams.hparams.num_freq() Examples

def _build_mel_basis():
    n_fft = (hparams.num_freq - 1) * 2
    return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels) 

def spectrogram2wav(spectrogram, n_iter=hparams.griffin_lim_iters, n_fft=(hparams.num_freq - 1) * 2,
                    win_length=int(hparams.frame_length_ms / 1000 * hparams.sample_rate),
                    hop_length=int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)):
    '''Converts spectrogram into a waveform using Griffin-lim's raw.
    '''

    def invert_spectrogram(spectrogram):
        '''
        spectrogram: [t, f]
        '''
        spectrogram = tf.expand_dims(spectrogram, 0)
        inversed = tf.contrib.signal.inverse_stft(spectrogram, win_length, hop_length, n_fft)
        squeezed = tf.squeeze(inversed, 0)
        return squeezed

    spectrogram = tf.transpose(spectrogram)

    spectrogram = tf.cast(spectrogram, dtype=tf.complex64)  # [t, f]
    X_best = tf.identity(spectrogram)
    for i in range(n_iter):
        X_t = invert_spectrogram(X_best)
        est = tf.contrib.signal.stft(X_t, win_length, hop_length, n_fft, pad_end=False)  # (1, T, n_fft/2+1)
        phase = est / tf.cast(tf.maximum(1e-8, tf.abs(est)), tf.complex64)  # [t, f]
        X_best = spectrogram * phase  # [t, t]
    X_t = invert_spectrogram(X_best)
    y = tf.real(X_t)

    return y 

def _stft_parameters():
  n_fft = (hparams.num_freq - 1) * 2
  hop_length = int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)
  win_length = int(hparams.frame_length_ms / 1000 * hparams.sample_rate)
  return n_fft, hop_length, win_length


# Conversions: 

def _build_mel_basis():
  n_fft = (hparams.num_freq - 1) * 2
  return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels) 

def __init__(self, coordinator, metadata_filename, hparams):
		super(Feeder, self).__init__()
		self._coord = coordinator
		self._hparams = hparams
		self._cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
		self._offset = 0

		# Load metadata
		self._mel_dir = os.path.join(os.path.dirname(metadata_filename), 'mels')
		self._linear_dir = os.path.join(os.path.dirname(metadata_filename), 'linear')
		with open(metadata_filename, encoding='utf-8') as f:
			self._metadata = [line.strip().split('|') for line in f]
			frame_shift_ms = hparams.hop_size / hparams.sample_rate
			hours = sum([int(x[4]) for x in self._metadata]) * frame_shift_ms / (3600)
			log('Loaded metadata for {} examples ({:.2f} hours)'.format(len(self._metadata), hours))

		# Create placeholders for inputs and targets. Don't specify batch size because we want
		# to be able to feed different batch sizes at eval time.
		self._placeholders = [
		tf.placeholder(tf.int32, shape=(None, None), name='inputs'),
		tf.placeholder(tf.int32, shape=(None, ), name='input_lengths'),
		tf.placeholder(tf.float32, shape=(None, None, hparams.num_mels), name='mel_targets'),
		tf.placeholder(tf.int32,[None],'mel_lengths'),
		tf.placeholder(tf.float32, shape=(None, None), name='token_targets'),
		tf.placeholder(tf.float32, shape=(None, None, hparams.num_freq), name='linear_targets'),
		]

		# Create queue for buffering data
		queue = tf.FIFOQueue(8, [tf.int32, tf.int32, tf.float32, tf.int32, tf.float32, tf.float32], name='input_queue')
		self._enqueue_op = queue.enqueue(self._placeholders)
		self.inputs, self.input_lengths, self.mel_targets, self.mel_lengths, self.token_targets, self.linear_targets = queue.dequeue()
		self.inputs.set_shape(self._placeholders[0].shape)
		self.input_lengths.set_shape(self._placeholders[1].shape)
		self.mel_targets.set_shape(self._placeholders[2].shape)
		self.mel_lengths.set_shape(self._placeholders[3].shape)
		self.token_targets.set_shape(self._placeholders[4].shape)
		self.linear_targets.set_shape(self._placeholders[5].shape) 

def _stft_parameters():
  n_fft = (hparams.num_freq - 1) * 2
  hop_length = int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)
  win_length = int(hparams.frame_length_ms / 1000 * hparams.sample_rate)
  return n_fft, hop_length, win_length


# Conversions: 

def _build_mel_basis():
  n_fft = (hparams.num_freq - 1) * 2
  return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels) 

def _stft_parameters():
  n_fft = (hparams.num_freq - 1) * 2
  hop_length = int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)
  win_length = int(hparams.frame_length_ms / 1000 * hparams.sample_rate)
  return n_fft, hop_length, win_length


# Conversions: 

def _build_mel_basis():
  n_fft = (hparams.num_freq - 1) * 2
  return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels) 

def __init__(self, coordinator, metadata_filename, hparams):
		super(Feeder, self).__init__()
		self._coord = coordinator
		self._hparams = hparams
		self._cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
		self._offset = 0

		# Load metadata
		self._mel_dir = os.path.join(os.path.dirname(metadata_filename), 'mels')
		self._linear_dir = os.path.join(os.path.dirname(metadata_filename), 'linear')
		with open(metadata_filename, encoding='utf-8') as f:
			self._metadata = [line.strip().split('|') for line in f]
			frame_shift_ms = hparams.hop_size / hparams.sample_rate
			hours = sum([int(x[4]) for x in self._metadata]) * frame_shift_ms / (3600)
			log('Loaded metadata for {} examples ({:.2f} hours)'.format(len(self._metadata), hours))

		# Create placeholders for inputs and targets. Don't specify batch size because we want
		# to be able to feed different batch sizes at eval time.
		self._placeholders = [
		tf.placeholder(tf.int32, shape=(None, None), name='inputs'),
		tf.placeholder(tf.int32, shape=(None, ), name='input_lengths'),
		tf.placeholder(tf.float32, shape=(None, None, hparams.num_mels), name='mel_targets'),
		tf.placeholder(tf.int32,[None],'mel_lengths'),
		tf.placeholder(tf.float32, shape=(None, None), name='token_targets'),
		tf.placeholder(tf.float32, shape=(None, None, hparams.num_freq), name='linear_targets'),
		]

		# Create queue for buffering data
		queue = tf.FIFOQueue(8, [tf.int32, tf.int32, tf.float32, tf.int32, tf.float32, tf.float32], name='input_queue')
		self._enqueue_op = queue.enqueue(self._placeholders)
		self.inputs, self.input_lengths, self.mel_targets, self.mel_lengths, self.token_targets, self.linear_targets = queue.dequeue()
		self.inputs.set_shape(self._placeholders[0].shape)
		self.input_lengths.set_shape(self._placeholders[1].shape)
		self.mel_targets.set_shape(self._placeholders[2].shape)
		self.mel_lengths.set_shape(self._placeholders[3].shape)
		self.token_targets.set_shape(self._placeholders[4].shape)
		self.linear_targets.set_shape(self._placeholders[5].shape) 

def _stft_parameters():
	n_fft = (hps.num_freq - 1) * 2
	hop_length = int(hps.frame_shift_ms / 1000 * hps.sample_rate)
	win_length = int(hps.frame_length_ms / 1000 * hps.sample_rate)
	return n_fft, hop_length, win_length


# Conversions: 

def _build_mel_basis():
	n_fft = (hps.num_freq - 1) * 2
	return librosa.filters.mel(hps.sample_rate, n_fft, n_mels=hps.num_mels) 

def synthesize(self, text, index, out_dir, log_dir, mel_filename, reference_mel):
		cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
		seq = text_to_sequence(text, cleaner_names)
		feed_dict = {
			self.model.inputs: [np.asarray(seq, dtype=np.int32)],
			self.model.input_lengths: np.asarray([len(seq)], dtype=np.int32)
		}

		if self.gta:
			feed_dict[self.model.mel_targets] = np.load(mel_filename).reshape(1, -1, 80)
			feed_dict[self.model.reference_mel] = np.load(mel_filename).reshape(1, -1, 80)
		elif hparams.use_vae:
			reference_mel = [np.asarray(reference_mel, dtype=np.float32)]
			feed_dict[self.model.reference_mel] = reference_mel


		if self.gta or not hparams.predict_linear:
			mels, alignment = self.session.run([self.mel_outputs, self.alignment], feed_dict=feed_dict)

		else:
			linear, mels, alignment = self.session.run([self.linear_outputs, self.mel_outputs, self.alignment], feed_dict=feed_dict)
			linear = linear.reshape(-1, hparams.num_freq)

		mels = mels.reshape(-1, hparams.num_mels) #Thanks to @imdatsolak for pointing this out

		# Write the spectrogram to disk
		# Note: outputs mel-spectrogram files and target ones have same names, just different folders
		mel_filename = os.path.join(out_dir, 'speech-mel-{:05d}.npy'.format(index))
		np.save(mel_filename, mels, allow_pickle=False)

		if log_dir is not None:
			#save wav (mel -> wav)
			wav = audio.inv_mel_spectrogram(mels.T)
			audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-mel.wav'.format(index)))

			if hparams.predict_linear:
				#save wav (linear -> wav)
				wav = audio.inv_linear_spectrogram(linear.T)
				audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-linear.wav'.format(index)))

			#save alignments
			plot.plot_alignment(alignment, os.path.join(log_dir, 'plots/speech-alignment-{:05d}.png'.format(index)),
				info='{}'.format(text), split_title=True)

			#save mel spectrogram plot
			plot.plot_spectrogram(mels, os.path.join(log_dir, 'plots/speech-mel-{:05d}.png'.format(index)),
				info='{}'.format(text), split_title=True)

		return mel_filename 

def synthesize(self, text, index, out_dir, log_dir, mel_filename, reference_mel):
		cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
		seq = text_to_sequence(text, cleaner_names)
		feed_dict = {
			self.model.inputs: [np.asarray(seq, dtype=np.int32)],
			self.model.input_lengths: np.asarray([len(seq)], dtype=np.int32)
		}

		if self.gta:
			feed_dict[self.model.mel_targets] = np.load(mel_filename).reshape(1, -1, 80)
			feed_dict[self.model.reference_mel] = np.load(mel_filename).reshape(1, -1, 80)
		elif hparams.use_vae:
			reference_mel = [np.asarray(reference_mel, dtype=np.float32)]
			feed_dict[self.model.reference_mel] = reference_mel


		if self.gta or not hparams.predict_linear:
			mels, alignment = self.session.run([self.mel_outputs, self.alignment], feed_dict=feed_dict)

		else:
			linear, mels, alignment = self.session.run([self.linear_outputs, self.mel_outputs, self.alignment], feed_dict=feed_dict)
			linear = linear.reshape(-1, hparams.num_freq)

		mels = mels.reshape(-1, hparams.num_mels) #Thanks to @imdatsolak for pointing this out

		# Write the spectrogram to disk
		# Note: outputs mel-spectrogram files and target ones have same names, just different folders
		mel_filename = os.path.join(out_dir, 'speech-mel-{:05d}.npy'.format(index))
		np.save(mel_filename, mels, allow_pickle=False)

		if log_dir is not None:
			#save wav (mel -> wav)
			wav = audio.inv_mel_spectrogram(mels.T)
			audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-mel.wav'.format(index)))

			if hparams.predict_linear:
				#save wav (linear -> wav)
				wav = audio.inv_linear_spectrogram(linear.T)
				audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-linear.wav'.format(index)))

			#save alignments
			plot.plot_alignment(alignment, os.path.join(log_dir, 'plots/speech-alignment-{:05d}.png'.format(index)),
				info='{}'.format(text), split_title=True)

			#save mel spectrogram plot
			plot.plot_spectrogram(mels, os.path.join(log_dir, 'plots/speech-mel-{:05d}.png'.format(index)),
				info='{}'.format(text), split_title=True)

		return mel_filename