Python python_speech_features.logfbank() Examples

def fbank(wav_path, flat=True):
    """ Currently grabs log Mel filterbank, deltas and double deltas."""

    (rate, sig) = wav.read(wav_path)
    if len(sig) == 0:
        logger.warning("Empty wav: {}".format(wav_path))
    fbank_feat = python_speech_features.logfbank(sig, rate, nfilt=40)
    energy = extract_energy(rate, sig)
    feat = np.hstack([energy, fbank_feat])
    delta_feat = python_speech_features.delta(feat, 2)
    delta_delta_feat = python_speech_features.delta(delta_feat, 2)
    all_feats = [feat, delta_feat, delta_delta_feat]
    if not flat:
        all_feats = np.array(all_feats)
        # Make time the first dimension for easy length normalization padding
        # later.
        all_feats = np.swapaxes(all_feats, 0, 1)
        all_feats = np.swapaxes(all_feats, 1, 2)
    else:
        all_feats = np.concatenate(all_feats, axis=1)

    # Log Mel Filterbank, with delta, and double delta
    feat_fn = wav_path[:-3] + "fbank.npy"
    np.save(feat_fn, all_feats) 

def __mel(audio_data, sampling_rate, win_len, win_step, num_features, n_fft, f_min, f_max):
    """Convert a wav signal into a logarithmically scaled mel filterbank.

    Args:
        audio_data (np.ndarray): Wav signal.
        sampling_rate (int):  Sampling rate.
        win_len (float): Window length in seconds.
        win_step (float): Window stride in seconds.
        num_features (int): Number of features to generate.
        n_fft (int): Number of Fast Fourier Transforms.
        f_min (float): Minimum frequency to consider.
        f_max (float): Maximum frequency to consider.

    Returns:
        np.ndarray: Mel-filterbank. Shape: [time, num_features]
    """
    mel = psf.logfbank(signal=audio_data, samplerate=sampling_rate, winlen=win_len,
                       winstep=win_step, nfilt=num_features, nfft=n_fft,
                       lowfreq=f_min, highfreq=f_max, preemph=0.97)
    return mel 

def __call__(self, pkg, cached_file=None):
        pkg = format_package(pkg)
        wav = pkg['chunk']
        if torch.is_tensor(wav):
            wav = wav.data.numpy().astype(np.float32)
        max_frames = wav.shape[0] // self.hop
        if cached_file is not None:
            # load pre-computed data
            X = torch.load(cached_file)
            beg_i = pkg['chunk_beg_i'] // self.hop
            end_i = pkg['chunk_end_i'] // self.hop
            X = X[:, beg_i:end_i]
            pkg[self.name] = X
        else:
            winlen = (float(self.win) / self.rate)
            winstep = (float(self.hop) / self.rate)
            X = logfbank(wav, self.rate, winlen, winstep,
                         self.n_filters, self.n_fft).T
            expected_frames = len(wav) // self.hop

            if self.der_order > 0 :
                deltas=[X]
                for n in range(1,self.der_order+1):
                    deltas.append(librosa.feature.delta(X,order=n))
                X=np.concatenate(deltas)

            fbank = torch.FloatTensor(X)
            if fbank.shape[1] < expected_frames:
                P = expected_frames - fbank.shape[1]
                # pad repeating borders
                fbank = F.pad(fbank.unsqueeze(0), (0, P), mode='replicate')
                fbank = fbank.squeeze(0)
            pkg[self.name] = fbank
        # Overwrite resolution to hop length
        pkg['dec_resolution'] = self.hop
        return pkg 

def SpeechFeaturesPreprocessor(feature_type: str = "mfcc",
                               delta_order: int = 0,
                               delta_window: int = 2,
                               **kwargs) -> Callable:
    """Calculate speech features.

    First, the given type of features (e.g. MFCC) is computed using a window
    of length `winlen` and step `winstep`; for additional keyword arguments
    (specific to each feature type), see
    http://python-speech-features.readthedocs.io/. Then, delta features up to
    `delta_order` are added.

    By default, 13 MFCCs per frame are computed. To add delta and delta-delta
    features (resulting in 39 coefficients per frame), set `delta_order=2`.

    Arguments:
        feature_type: mfcc, fbank, logfbank or ssc (default is mfcc)
        delta_order: maximum order of the delta features (default is 0)
        delta_window: window size for delta features (default is 2)
        **kwargs: keyword arguments for the appropriate function from
            python_speech_features

    Returns:
        A numpy array of shape [num_frames, num_features].
    """

    if feature_type not in FEATURE_TYPES:
        raise ValueError(
            "Unknown speech feature type '{}'".format(feature_type))

    def preprocess(audio: Audio) -> np.ndarray:
        features = [FEATURE_TYPES[feature_type](
            audio.data, samplerate=audio.rate, **kwargs)]

        for _ in range(delta_order):
            features.append(delta(features[-1], delta_window))

        return np.concatenate(features, axis=1)

    return preprocess 

def SpeechFeaturesPreprocessor(feature_type: str = "mfcc",
                               delta_order: int = 0,
                               delta_window: int = 2,
                               **kwargs) -> Callable:
    """Calculate speech features.

    First, the given type of features (e.g. MFCC) is computed using a window
    of length `winlen` and step `winstep`; for additional keyword arguments
    (specific to each feature type), see
    http://python-speech-features.readthedocs.io/. Then, delta features up to
    `delta_order` are added.

    By default, 13 MFCCs per frame are computed. To add delta and delta-delta
    features (resulting in 39 coefficients per frame), set `delta_order=2`.

    Arguments:
        feature_type: mfcc, fbank, logfbank or ssc (default is mfcc)
        delta_order: maximum order of the delta features (default is 0)
        delta_window: window size for delta features (default is 2)
        **kwargs: keyword arguments for the appropriate function from
            python_speech_features

    Returns:
        A numpy array of shape [num_frames, num_features].
    """

    if feature_type not in FEATURE_TYPES:
        raise ValueError(
            "Unknown speech feature type '{}'".format(feature_type))

    def preprocess(audio: Audio) -> np.ndarray:
        features = [FEATURE_TYPES[feature_type](
            audio.data, samplerate=audio.rate, **kwargs)]

        for _ in range(delta_order):
            features.append(delta(features[-1], delta_window))

        return np.concatenate(features, axis=1)

    return preprocess 

def SpeechFeaturesPreprocessor(feature_type: str = "mfcc",
                               delta_order: int = 0,
                               delta_window: int = 2,
                               **kwargs) -> Callable:
    """Calculate speech features.

    First, the given type of features (e.g. MFCC) is computed using a window
    of length `winlen` and step `winstep`; for additional keyword arguments
    (specific to each feature type), see
    http://python-speech-features.readthedocs.io/. Then, delta features up to
    `delta_order` are added.

    By default, 13 MFCCs per frame are computed. To add delta and delta-delta
    features (resulting in 39 coefficients per frame), set `delta_order=2`.

    Arguments:
        feature_type: mfcc, fbank, logfbank or ssc (default is mfcc)
        delta_order: maximum order of the delta features (default is 0)
        delta_window: window size for delta features (default is 2)
        **kwargs: keyword arguments for the appropriate function from
            python_speech_features

    Returns:
        A numpy array of shape [num_frames, num_features].
    """

    if feature_type not in FEATURE_TYPES:
        raise ValueError(
            "Unknown speech feature type '{}'".format(feature_type))

    def preprocess(audio: Audio) -> np.ndarray:
        features = [FEATURE_TYPES[feature_type](
            audio.data, samplerate=audio.rate, **kwargs)]

        for _ in range(delta_order):
            features.append(delta(features[-1], delta_window))

        return np.concatenate(features, axis=1)

    return preprocess 

def get_fbank_feature(wavsignal, fs):
    '''
    输入为wav文件数学表示和采样频率，输出为语音的FBANK特征+一阶差分+二阶差分；
    '''
    feat_fbank = logfbank(wavsignal, fs, nfilt=40)
    feat_fbank_d = delta(feat_fbank, 2)
    feat_fbank_dd = delta(feat_fbank_d, 2)
    wav_feature = np.column_stack((feat_fbank, feat_fbank_d, feat_fbank_dd))
    return wav_feature