If I understand this paper and FlowSeq correctly, the normalizing flow is used to model the dependence of text X (from the posterior P(Z|X, Y)). As GAN can also model the distribution, can I use a GAN-based network to replace the flow-based prior P(Z|X)?

Hi thank you for your awesome work. I tried to synthesize using inference script and checkpoint u provided the result sound robotic, why is that ? do I missing something ?

在运行 preprocess.py 这步时出现异常

Traceback (most recent call last): File "train.py", line 329, in main() File "train.py", line 257, in main for fids, texts, mels, t_lengths, m_lengths in train_set.take(1): File "/home/wuyx/miniconda3/envs/vc/lib/python3.6/site-packages/tensorflow/python/data/ops/iterator_ops.py", line 761, in next return self._next_internal() File "/home/wuyx/miniconda3/envs/vc/lib/python3.6/site-packages/tensorflow/python/data/ops/iterator_ops.py", line 747, in _next_internal output_shapes=self._flat_output_shapes) File "/home/wuyx/miniconda3/envs/vc/lib/python3.6/site-packages/tensorflow/python/ops/gen_dataset_ops.py", line 2728, in iterator_get_next _ops.raise_from_not_ok_status(e, name) File "/home/wuyx/miniconda3/envs/vc/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 6897, in raise_from_not_ok_status six.raise_from(core._status_to_exception(e.code, message), None) File "", line 3, in raise_from tensorflow.python.framework.errors_impl.InvalidArgumentError: Type mismatch between parsed tensor (float) and dtype (double) [[{{node ParseTensor_1}}]] [Op:IteratorGetNext]

in theposterior.py code: time_level_log_probs = -0.5 * (tf.cast(dim, tf.float32) * tf.math.log(2 * np.pi)+ tf.reduce_sum(expanded_logvar + normalized_samples ** 2., axis=3))

but the log_prob of gaussian: log_probs = log(1.0 / (sqrt(2.0 * pi) * std) * exp(-0.5 * (x-u) ** 2 / std ** 2)) = -0.5 * (log(2.0 * pi) + 2.0 * log(std) + (x-u) ** 2 / std ** 2)

so you miss a constant value 2.0 of expanded_logvar although it doesn't matter? time_level_log_probs = -0.5 * (tf.cast(dim, tf.float32) * tf.math.log(2 * np.pi)+ tf.reduce_sum(2.0 * expanded_logvar + normalized_samples ** 2., axis=3))

I downloaded the pretrained model of databaker and synthesized wavs using inference.py. The results are not very good, I mean the alignment is not right especially when the input text is long. For example, "失恋的人特别喜欢往人烟罕至的角落里钻。", the synthesized wavs sounds like: 失恋的人特别喜欢往人烟罕至的_角角落里钻钻钻钻_

For longer input text，the synthesized wavs are totally wrong

Hi, I want to know what config does you use when you train the hifigan model of DataBaker to get the samples in the webset https://light1726.github.io/vaenar-tts/.
With these parameters clearified, we can better compare the quality of the synthsized wavs with other SOTA acoustic models.

I mean the following three parameter in config file. "upsample_rates":
"upsample_kernel_sizes":
"upsample_initial_channel":

Thank you!