There is a contraction in WaveRNNInferenceWrapper during inference when batched

[yangarbiter]

The waveform above is the ground truth and the one below is the generated waveform with batched set to true.

See this colab example (internal for more information).

[mthrok]

Looking at the code of WaveRNNInferenceWrapper, the same overlap value is used for folding and unfolding.

audio/examples/pipeline_wavernn/wavernn_inference_wrapper.py

Lines 193 to 194 in 19f53cf

	if batched:
	specgram = _fold_with_overlap(specgram, timesteps, overlap)

audio/examples/pipeline_wavernn/wavernn_inference_wrapper.py

Lines 204 to 205 in 19f53cf

	if batched:
	output = _xfade_and_unfold(output, overlap)

But the former is spectrogram while the later is waveform, and they do not have the same number of samples over time axis, even though they should have a similar wall time range. I think that is the cause.

In the example notebook, the shapes of the tensors look like the following;

1. Input spectrogram [1, 80, 164]
2. Spectrogram after xfold [2, 80, 110]
3. Output waveform [2, 1, 29150]
4. Waveform after unfolding [1, 58295]