Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge


Computes the inverse modified DCT of mdcts.

To reconstruct an original waveform, the same window function should be used with mdct and inverse_mdct.

Example usage:

def compare_round_trip():
  samples = 1000
  frame_length = 400
  halflen = frame_length // 2
  waveform = tf.random.normal(dtype=tf.float32, shape=[samples])
  waveform_pad = tf.pad(waveform, [[halflen, 0],])
  mdct = tf.signal.mdct(waveform_pad, frame_length, pad_end=True,
  inverse_mdct = tf.signal.inverse_mdct(mdct,
  inverse_mdct = inverse_mdct[halflen: halflen + samples]
  return waveform, inverse_mdct
waveform, inverse_mdct = compare_round_trip()
np.allclose(waveform.numpy(), inverse_mdct.numpy(), rtol=1e-3, atol=1e-4)

Implemented with TPU/GPU-compatible ops and supports gradients.

mdcts A float32/float64 [..., frames, frame_length // 2] Tensor of MDCT bins representing a batch of frame_length // 2-point MDCTs.
window_fn A callable that takes a frame_length and a dtype keyword argument and returns a [frame_length] Tensor of samples in the provided datatype. If set to None, a rectangular window with a scale of 1/sqrt(2) is used. For perfect reconstruction of a signal from mdct followed by inverse_mdct, please use tf.signal.vorbis_window, tf.signal.kaiser_bessel_derived_window or None. If using another window function, make sure that w[n]^2 + w[n + frame_length // 2]^2 = 1 and w[n] = w[frame_length - n - 1] for n = 0,...,frame_length // 2 - 1 to achieve perfect reconstruction.
norm If "ortho", orthonormal inverse DCT4 is performed, if it is None, a regular dct4 followed by scaling of 1/frame_length is performed.
name An optional name for the operation.

A [..., samples] Tensor of float32/float64 signals representing the inverse MDCT for each input MDCT in mdcts where samples is (frames - 1) * (frame_length // 2) + frame_length.

ValueError If mdcts is not at least rank 2.