@inproceedings{choi2017kapre,
  title={Kapre: On-GPU Audio Preprocessing Layers for a Quick Implementation of Deep Neural Network Models with Keras},
  author={Choi, Keunwoo and Joo, Deokjin and Kim, Juho},
  booktitle={Machine Learning for Music Discovery Workshop at 34th International Conference on Machine Learning},
  year={2017},
  organization={ICML}
}

 * n_dft: int > 0 [scalar]
   - The number of DFT points, presumably power of 2.
   - Default: `512`
 * n_hop: int > 0 [scalar]
   - Hop length between frames in sample,  probably <= `n_dft`.
   - Default: `None` (`n_dft / 2` is used)
 * padding: str, `'same'` or `'valid'`.
   - Padding strategies at the ends of signal.
   - Default: `'same'`
 * power_spectrogram: float [scalar],
   - `2.0` to get power-spectrogram, `1.0` to get amplitude-spectrogram.
   - Usually `1.0` or `2.0`.
   - Default: `2.0`
 * return_decibel_spectrogram: bool,
    - Whether to return in decibel or not, i.e. returns log10(amplitude spectrogram) if `True`.
    - Recommended to use `True`, although it's not by default.
    - Default: `False`
 * trainable_kernel: bool
   -  Whether the kernels are trainable or not.
   -  If `True`, Kernels are initialised with DFT kernels and then trained.
   -  Default: `False`
* image_data_format: string, `'channels_first'` or `'channels_last'`.
   -  The returned spectrogram follows this image_data_format strategy.
   -  If `'default'`, it follows the current Keras session's setting.
   -  Setting is in `./keras/keras.json`.
   -  Default: `'default'`

 * The input should be a 2D array, `(audio_channel, audio_length)`.
 * E.g., `(1, 44100)` for mono signal, `(2, 44100)` for stereo signal.
 * It supports multichannel signal input, so `audio_channel` can be any positive integer.

 * abs(Spectrogram) in a shape of 2D data, i.e.,
 * `(None, n_channel, n_freq, n_time)` if `'channels_first'`,
 * `(None, n_freq, n_time, n_channel)` if `'channels_last'`,

 * sr: integer > 0 [scalar]
   - sampling rate of the input audio signal.
   - Default: `22050`
 * n_mels: int > 0 [scalar]
   - The number of mel bands.
   - Default: `128`
 * fmin: float > 0 [scalar]
   - Minimum frequency to include in Mel-spectrogram.
   - Default: `0.0`
 * fmax: float > `fmin` [scalar]
   - Maximum frequency to include in Mel-spectrogram.
   - If `None`, it is inferred as `sr / 2`.
   - Default: `None`
 * power_melgram: float [scalar]
   - Power of `2.0` if power-spectrogram,
   - `1.0` if amplitude spectrogram.
   - Default: `1.0`
 * return_decibel_melgram: bool
   - Whether to return in decibel or not, i.e. returns log10(amplitude spectrogram) if `True`.
   - Recommended to use `True`, although it's not by default.
   - Default: `False`
 * trainable_fb: bool
   - Whether the spectrogram -> mel-spectrogram filterbanks are trainable.
   - If `True`, the frequency-to-mel matrix is initialised with mel frequencies but trainable.
   - If `False`, it is initialised and then frozen.
   - Default: `False`
 * **kwargs:
   - The keyword arguments of `Spectrogram` such as `n_dft`, `n_hop`,
   - `padding`, `trainable_kernel`, `image_data_format`.

 * The input should be a 2D array, `(audio_channel, audio_length)`.
E.g., `(1, 44100)` for mono signal, `(2, 44100)` for stereo signal.
 * It supports multichannel signal input, so `audio_channel` can be any positive integer.

 * abs(mel-spectrogram) in a shape of 2D data, i.e.,
 * `(None, n_channel, n_mels, n_time)` if `'channels_first'`,
 * `(None, n_mels, n_time, n_channel)` if `'channels_last'`,

* amin: float [scalar]
    - Noise floor. Default: 1e-10
* top_db: float [scalar]
    - Dynamic range of output. Default: 80.0

* input_shape: tuple of ints
    - E.g., `(None, n_ch, n_row, n_col)` if theano.
* str_axis: str
    - used ONLY IF `int_axis` is `None`.
    - `'batch'`, `'data_sample'`, `'channel'`, `'freq'`, `'time')`
    - Even though it is optional, actually it is recommended to use
    - `str_axis` over `int_axis` because it provides more meaningful
    - and image data format-robust interface.
* int_axis: int
    - axis index that along which mean/std is computed.
    - `0` for per data sample, `-1` for per batch.
    - `1`, `2`, `3` for channel, row, col (if channels_first)
    - if `int_axis is None`, `str_axis` SHOULD BE set.

Input/output are 2D image format.
E.g., if channel_first,
    - input_shape: ``(None, n_ch, n_freqs, n_time)``
    - output_shape: ``(None, n_ch, n_mels, n_time)``

* n_fbs: int
   - Number of filterbanks
* sr: int
    - sampling rate. It is used to initialize `freq_to_mel`.
* init: str
    - if `'mel'`, init with mel center frequencies and stds.
* fmin: float
    - min frequency of filterbanks.
    - If `init == 'log'`, fmin should be > 0. Use `None` if you got no idea.
* fmax: float
    - max frequency of filterbanks.
    - If `init == 'log'`, fmax is ignored.
* trainable_fb: bool,
    - Whether the filterbanks are trainable or not.

* power: float [scalar]
    - The power of noise. std if it's white noise.
    - Default: `0.1`
* random_gain: bool
    - Whether the noise gain is random or not.
    - If `True`, gain is sampled from `uniform(low=0.0, high=power)` in every batch.
    - Default: `False`
* noise_type; str,
    - Specify the type of noise. It only supports `'white'` now.
    - Default: `white`