Tensorflowtts: What do you think about FastPitch ?

[Yablon]

I found that NVIDIA released a paper named FastPitch: Parallel Text-to-speech with Pitch Prediction.

Do you have any plan to try and release some samples to compare with tacotron based models and fastspeech ?

[dathudeptrai]

@Yablon samples here https://tensorspeech.github.io/TensorflowTTS/. FastSpeech2 ~ Tacotron2 interms of quality but much faster, i think :D. Nvidia FastPitch is the same as FastSpeech2 except fastspeech2 use energy while fastpitch doesn't :)). You can hear our samples and judging by yourself :))

[Yablon]

@dathudeptrai Thank you for your quick relpy !

[dathudeptrai]

@Yablon i think u have trained some FastPitch model, hope you can make some comments after hear our samples :). Note that nvidia use Waveglow which is much slower and larger compared with our Mb-Melgan.

[Yablon]

@dathudeptrai Yes, I have trained the FastPitch model using NVIDIA repo and its recommended parameters, here are some audio samples. The vocoder is their pretrained WaveGlow.
wavs_devset10-150.zip

Following is audio samples from this repo,
fastspeech2.100k.zip

Both of the two models are trained on same LJSpeech Dataset. Sample rate is the same, 22050Hz.

I think that FastPitch samples is better than FastSpeech2 in the sense of audio quality. The shortage may come from the difference of the WaveGlow and MelGan vocoder.

What is your opinion ?

[dathudeptrai]

@Yablon yeah, waveglow is better than mb-melgan but above samples is not the same so we can't judge anything. It's much larger and much slower too. BTW, can you generate FastPitch + Waveglow samples same as my samples on my demo page ? (https://tensorspeech.github.io/TensorflowTTS/) (pls see the FastSpeech-2 samples (BBC news) section). Adition, i remember that my pretrained model is 150k fastspeech2 -_- not 100k :D.

[Yablon]

@dathudeptrai Okay, I will generate them later.

[dathudeptrai]

@Yablon can you update :))

[Yablon]

@dathudeptrai Oh，I will update today

[Yablon]

wavs_tts_1500_pace_0.8.zip
wavs_tts_1500.zip
wavs_tts_1500_pace_0.8_pitch_shift60hz.zip

Finally I got this. @dathudeptrai
wavs_tts_1500_pace_0.8.zip is 0.8 pace of wavs_tts_1500.zip. wavs_tts_1500_pace0.8_pitch_shift60hz.zip has pitch shifted down by 60hz.

The model was trained with code of this repo for 1500 epochs using LJSpeech-1.1 data as the repo recommended.
The vocoder is the official waveglow

The audios are generated using the following file. Text in the file is copied from your demo page. Mels in first column of the file is not used.
tts.txt

[dathudeptrai]

@Yablon thanks, i think it's comparable. Seem that FastSpeech2 produce better mel than FastPitch but waveglow reconstruct audio better than mb-melgan.