Function mfsc

python/mlx/data/features/audio.py:162–267

Returns a function that computes spectrogram features from audio in particular mel-frequency spectral coefficients (MFSCs). .. note:: This feature extractor operates on mono audio provided as a 1D array. Meaning the input should have shape ``(N,)`` and not ``(N, 1)``. You may

(
    n_filterbank,
    sampling_freq,
    frame_size_ms=25,
    frame_stride_ms=10,
    pre_emphasis_coeff=0.97,
    window_type=WindowType.Hamming,
    low_freq=0,
    high_freq=-1,
    mel_floor=1.0,
    freq_scale=FrequencyScale.MEL,
    post_process=None,
)

def mfsc(
    n_filterbank,
    sampling_freq,
    frame_size_ms=25,
    frame_stride_ms=10,
    pre_emphasis_coeff=0.97,
    window_type=WindowType.Hamming,
    low_freq=0,
    high_freq=-1,
    mel_floor=1.0,
    freq_scale=FrequencyScale.MEL,
    post_process=None,
):
    """Returns a function that computes spectrogram features from audio in
    particular mel-frequency spectral coefficients (MFSCs).

    .. note::
       This feature extractor operates on mono audio provided as a 1D array.
       Meaning the input should have shape ``(N,)`` and not ``(N, 1)``. You may
       want to look into :meth:`~mlx.data.Buffer.squeeze` to transform arrays of shape
       ``(N, 1)`` to ``(N,)``.

    The featurization function

    1. computes a sliding window of the input audio
    2. applies a pre-emphasis filter
    3. applies a windowing function
    4. computes the power spectrum
    5. computes triangular filterbank features
    6. compute the log of the features
    7. apply any post processing function that may be provided

    The following example loads the librispeech dataset and computes MFSC
    features:

    .. code-block:: python

        from mlx.data.datasets import load_librispeech
        from mlx.data.features import mfsc

        dset = (
            load_librispeech()
            .squeeze("audio")
            .key_transform("audio", mfsc(80, 16000))
            .to_stream()
            .prefetch(16, 8)
            .batch(16)
            .prefetch(2, 1)
        )

    Args:
        n_filterbank (int): How many frequency bands to use. This number will
            be the dimensionality of the resulting features.
        sampling_freq (int): The sampling frequency of the input audio in Hz.
        frame_size_ms (int): Each output feature will correspond to that many
            milliseconds of input audio. (default: 25)
        frame_stride_ms (int): Two consecutive features will correspond to
            audio windows that are that many milliseconds apart. (default: 10)