dhgrs / chainer-clarinet Goto Github PK

Hi,

I have trained both teacher/student networks.
Teacher has very nice quality although it takes way too long to generate (1s takes ~6 minutes on gpu).
Student results sound a bit robotic and have "noise" where should be silence.
Do you have any tips on that? (I haven't changed any parameters besides path to trained teacher model).

UPDATE:
Here I my examples for 2 short (~1s) files:
clarinet-results.zip

Thanks.

Hello
I would like to experiment with 48kHz sampling frequency.
What kind of parameter setting is good?

I also want to put the acoustic features from outside the script.
Is there any good way to do it?

Hi,

the Clarinet paper mentions also Text-to-Wave Architecture for end-to-end TTS.

Do you have any suggestions what would I need to do full TTS once the student network is trained?

Should I use some pre-trained model to produce mel-spetrograms like Tacotron2 or DeepVoice3? Or something else entirely?

Thanks!

Hi,

is there a reason why

StudentGaussianIAF/teacher_params.py has length=12000
and
StudentGaussianIAF/params.py has length=24000

while all other preprocessing params stay the same?

Thanks.

Hi,

Thanks for sharing.

It seems that there is an extra sampling step after the student network at https://github.com/dhgrs/chainer-ClariNet/blob/master/StudentGaussianIAF/net.py#L58 to get student samples. However, Algorithm 1 in https://arxiv.org/pdf/1807.07281.pdf uses z from the last IAF flow as student samples. What is the motivation of making such change?

Thanks,

Jian

Hi,

I have trained AutoregressiveWaveNet using command from README.md

python train.py -g 0

Then I have tried to generate audio using command from README.md

python generate.py -i ../../LJSpeech-1.1/wavs/LJ001-0001.wav -o result.wav -m 2018_09_27_16_03_22/snapshot_iter_500000 -g 0

Generated audio was completely silent, do you have any tip what could have gone wrong?

Thanks.

Hello
I would like to experiment with 48kHz sampling frequency.
What kind of parameter setting is good?

I also want to put the acoustic features from outside the script.
Is there any good way to do it?

Could you provide your loss plot from training autoregressive wavenet?
I've started training model with your implementation and have got following loss "jump". Is it OK?

Hello.

I would like to experiment with another acoustic features already extracted from speechs.
I am trying it in several ways, but for now it does not work well.
Could you tell me if there is any good way?

Hello
I would like to experiment with 48kHz sampling frequency.
What kind of parameter setting is good?

I also want to put the acoustic features from outside the script.
Is there any good way to do it?

I want to analysis the samples, but seems that they cannot be downloaded from nana-music. Is there any chance you can share your samples in a different way?

Thanks,

Hi,
Do you plan subj?
thanks