Python hparams.hparams.silence_threshold() Examples

The following are 15 code examples of hparams.hparams.silence_threshold(). You can vote up the ones you like or vote down the ones you don't like, and go to the original project or source file by following the links above each example. You may also want to check out all available functions/classes of the module hparams.hparams , or try the search function .
Example #1
Source File: audio.py    From cnn_vocoder with MIT License 6 votes vote down vote up 
def adjust_time_resolution(quantized, mel):
    """Adjust time resolution by repeating features

    Args:
        quantized (ndarray): (T,)
        mel (ndarray): (N, D)

    Returns:
        tuple: Tuple of (T,) and (T, D)
    """
    assert len(quantized.shape) == 1
    assert len(mel.shape) == 2

    upsample_factor = quantized.size // mel.shape[0]
    mel = np.repeat(mel, upsample_factor, axis=0)
    n_pad = quantized.size - mel.shape[0]
    if n_pad != 0:
        assert n_pad > 0
        mel = np.pad(mel, [(0, n_pad), (0, 0)], mode="constant", constant_values=0)

    # trim
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)

    return quantized[start:end], mel[start:end, :]
Example #2
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def adjust_time_resolution(quantized, mel):
    """Adjust time resolution by repeating features

    Args:
        quantized (ndarray): (T,)
        mel (ndarray): (N, D)

    Returns:
        tuple: Tuple of (T,) and (T, D)
    """
    assert len(quantized.shape) == 1
    assert len(mel.shape) == 2

    upsample_factor = quantized.size // mel.shape[0]
    mel = np.repeat(mel, upsample_factor, axis=0)
    n_pad = quantized.size - mel.shape[0]
    if n_pad != 0:
        assert n_pad > 0
        mel = np.pad(mel, [(0, n_pad), (0, 0)], mode="constant", constant_values=0)

    # trim
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)

    return quantized[start:end], mel[start:end, :]
Example #3
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def adjust_time_resolution(quantized, mel):
    """Adjust time resolution by repeating features

    Args:
        quantized (ndarray): (T,)
        mel (ndarray): (N, D)

    Returns:
        tuple: Tuple of (T,) and (T, D)
    """
    assert len(quantized.shape) == 1
    assert len(mel.shape) == 2

    upsample_factor = quantized.size // mel.shape[0]
    mel = np.repeat(mel, upsample_factor, axis=0)
    n_pad = quantized.size - mel.shape[0]
    if n_pad != 0:
        assert n_pad > 0
        mel = np.pad(mel, [(0, n_pad), (0, 0)], mode="constant", constant_values=0)

    # trim
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)

    return quantized[start:end], mel[start:end, :]
Example #4
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def adjust_time_resolution(quantized, mel):
    """Adjust time resolution by repeating features

    Args:
        quantized (ndarray): (T,)
        mel (ndarray): (N, D)

    Returns:
        tuple: Tuple of (T,) and (T, D)
    """
    assert len(quantized.shape) == 1
    assert len(mel.shape) == 2

    upsample_factor = quantized.size // mel.shape[0]
    mel = np.repeat(mel, upsample_factor, axis=0)
    n_pad = quantized.size - mel.shape[0]
    if n_pad != 0:
        assert n_pad > 0
        mel = np.pad(mel, [(0, n_pad), (0, 0)], mode="constant", constant_values=0)

    # trim
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)

    return quantized[start:end], mel[start:end, :]
Example #5
Source File: audio.py    From cnn_vocoder with MIT License 5 votes vote down vote up 
def trim(quantized):
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)
    return quantized[start:end]
Example #6
Source File: audio.py    From cnn_vocoder with MIT License 5 votes vote down vote up 
def start_and_end_indices(quantized, silence_threshold=2):
    for start in range(quantized.size):
        if abs(quantized[start] - 127) > silence_threshold:
            break
    for end in range(quantized.size - 1, 1, -1):
        if abs(quantized[end] - 127) > silence_threshold:
            break

    assert abs(quantized[start] - 127) > silence_threshold
    assert abs(quantized[end] - 127) > silence_threshold

    return start, end
Example #7
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def trim(quantized):
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)
    return quantized[start:end]
Example #8
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def start_and_end_indices(quantized, silence_threshold=2):
    for start in range(quantized.size):
        if abs(quantized[start] - 127) > silence_threshold:
            break
    for end in range(quantized.size - 1, 1, -1):
        if abs(quantized[end] - 127) > silence_threshold:
            break

    assert abs(quantized[start] - 127) > silence_threshold
    assert abs(quantized[end] - 127) > silence_threshold

    return start, end
Example #9
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def trim(quantized):
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)
    return quantized[start:end]
Example #10
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def trim(quantized):
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)
    return quantized[start:end]
Example #11
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def start_and_end_indices(quantized, silence_threshold=2):
    for start in range(quantized.size):
        if abs(quantized[start] - 127) > silence_threshold:
            break
    for end in range(quantized.size - 1, 1, -1):
        if abs(quantized[end] - 127) > silence_threshold:
            break

    assert abs(quantized[start] - 127) > silence_threshold
    assert abs(quantized[end] - 127) > silence_threshold

    return start, end
Example #12
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def trim(quantized):
    start, end = start_and_end_indices(quantized, hparams.silence_threshold)
    return quantized[start:end]
Example #13
Source File: audio.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
def start_and_end_indices(quantized, silence_threshold=2):
    for start in range(quantized.size):
        if abs(quantized[start] - 127) > silence_threshold:
            break
    for end in range(quantized.size - 1, 1, -1):
        if abs(quantized[end] - 127) > silence_threshold:
            break

    assert abs(quantized[start] - 127) > silence_threshold
    assert abs(quantized[end] - 127) > silence_threshold

    return start, end
Example #14
Source File: ljspeech.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 4 votes vote down vote up 
def _process_utterance(out_dir, index, wav_path, text):
    # Load the audio to a numpy array:
    wav = audio.load_wav(wav_path)

    if hparams.rescaling:
        wav = wav / np.abs(wav).max() * hparams.rescaling_max

    # Mu-law quantize
    if is_mulaw_quantize(hparams.input_type):
        # [0, quantize_channels)
        out = P.mulaw_quantize(wav, hparams.quantize_channels)

        # Trim silences
        start, end = audio.start_and_end_indices(out, hparams.silence_threshold)
        wav = wav[start:end]
        out = out[start:end]
        constant_values = P.mulaw_quantize(0, hparams.quantize_channels)
        out_dtype = np.int16
    elif is_mulaw(hparams.input_type):
        # [-1, 1]
        out = P.mulaw(wav, hparams.quantize_channels)
        constant_values = P.mulaw(0.0, hparams.quantize_channels)
        out_dtype = np.float32
    else:
        # [-1, 1]
        out = wav
        constant_values = 0.0
        out_dtype = np.float32

    # Compute a mel-scale spectrogram from the trimmed wav:
    # (N, D)
    mel_spectrogram = audio.melspectrogram(wav).astype(np.float32).T
    # lws pads zeros internally before performing stft
    # this is needed to adjust time resolution between audio and mel-spectrogram
    l, r = audio.lws_pad_lr(wav, hparams.fft_size, audio.get_hop_size())

    # zero pad for quantized signal
    out = np.pad(out, (l, r), mode="constant", constant_values=constant_values)
    N = mel_spectrogram.shape[0]
    assert len(out) >= N * audio.get_hop_size()

    # time resolution adjustment
    # ensure length of raw audio is multiple of hop_size so that we can use
    # transposed convolution to upsample
    out = out[:N * audio.get_hop_size()]
    assert len(out) % audio.get_hop_size() == 0

    timesteps = len(out)

    # Write the spectrograms to disk:
    audio_filename = 'ljspeech-audio-%05d.npy' % index
    mel_filename = 'ljspeech-mel-%05d.npy' % index
    np.save(os.path.join(out_dir, audio_filename),
            out.astype(out_dtype), allow_pickle=False)
    np.save(os.path.join(out_dir, mel_filename),
            mel_spectrogram.astype(np.float32), allow_pickle=False)

    # Return a tuple describing this training example:
    return (audio_filename, mel_filename, timesteps, text)
Example #15
Source File: ljspeech.py    From representation_mixing with BSD 3-Clause "New" or "Revised" License 4 votes vote down vote up 
def _process_utterance(out_dir, index, wav_path, text):
    # Load the audio to a numpy array:
    wav = audio.load_wav(wav_path)

    if hparams.rescaling:
        wav = wav / np.abs(wav).max() * hparams.rescaling_max

    # Mu-law quantize
    if is_mulaw_quantize(hparams.input_type):
        # [0, quantize_channels)
        out = P.mulaw_quantize(wav, hparams.quantize_channels)

        # Trim silences
        start, end = audio.start_and_end_indices(out, hparams.silence_threshold)
        wav = wav[start:end]
        out = out[start:end]
        constant_values = P.mulaw_quantize(0, hparams.quantize_channels)
        out_dtype = np.int16
    elif is_mulaw(hparams.input_type):
        # [-1, 1]
        out = P.mulaw(wav, hparams.quantize_channels)
        constant_values = P.mulaw(0.0, hparams.quantize_channels)
        out_dtype = np.float32
    else:
        # [-1, 1]
        out = wav
        constant_values = 0.0
        out_dtype = np.float32

    # Compute a mel-scale spectrogram from the trimmed wav:
    # (N, D)
    mel_spectrogram = audio.melspectrogram(wav).astype(np.float32).T
    # lws pads zeros internally before performing stft
    # this is needed to adjust time resolution between audio and mel-spectrogram
    l, r = audio.lws_pad_lr(wav, hparams.fft_size, audio.get_hop_size())

    # zero pad for quantized signal
    out = np.pad(out, (l, r), mode="constant", constant_values=constant_values)
    N = mel_spectrogram.shape[0]
    assert len(out) >= N * audio.get_hop_size()

    # time resolution adjustment
    # ensure length of raw audio is multiple of hop_size so that we can use
    # transposed convolution to upsample
    out = out[:N * audio.get_hop_size()]
    assert len(out) % audio.get_hop_size() == 0

    timesteps = len(out)

    # Write the spectrograms to disk:
    audio_filename = 'ljspeech-audio-%05d.npy' % index
    mel_filename = 'ljspeech-mel-%05d.npy' % index
    np.save(os.path.join(out_dir, audio_filename),
            out.astype(out_dtype), allow_pickle=False)
    np.save(os.path.join(out_dir, mel_filename),
            mel_spectrogram.astype(np.float32), allow_pickle=False)

    # Return a tuple describing this training example:
    return (audio_filename, mel_filename, timesteps, text)