Stil need a long time to synthesize a new speech about fastspeech HOT 9 CLOSED

@Liujingxiu23 Thanks, I just found I calculate time in a wrong way, to generate mel and mel_postnet, it takes about 0.05 using P100.

from fastspeech.

How about your wave synthesized? I synthesized wavs using model checkpoint "checkpoint_40000.pth.tar". The quality is bad.

from fastspeech.

Total loss is as follows (I cut top 30000):

from fastspeech.

I will put result wav here soon.

from fastspeech.

@xcmyz Can you tell me how long do you need to synthesis a speech? Thanks!

from fastspeech.

@xcmyz Thank you for your reply. How many iters the model need to converge ? I use the LJSpeech dataset

from fastspeech.

@Liujingxiu23 My loss is about 0.2, not good either. How long do you need to synthesis a new speech?

from fastspeech.

mel_output, mel_output_postnet = model(src_seq, src_pos)
mel = mel_output_postnet[k].detach().cpu().numpy()
wav = audio.inv_mel_spectrogram(mel.T)

For batch-size=10, seq_len=75, using one GPU, the time spend:

1. to generate mel and mel_postnet : 0.159 sec
2. to generate mel, mel_postnet and wavs: 37.625 sec

from fastspeech.

@xcmyz I tried your laest code, the acoustic quality improve much, early the same as tacotron2 I think.
The TTS corpus I use is chinese, and I keep the default hparams setting.
My loss seems not as good as yours, the postnet-mel-loss converge to about 0.5, the duration loss about 0.8. I don not know why?
By the way the pronunciation as well as the tone is not that good. For example, in same wavs "zhang" read like "zhan"， “tao3” read like “tao2” ，why this happens? Do you have any suggest to solve this problem?

from fastspeech.