WaveGlow sythensis result

[tingyang01]

I tried TTS with WaveGlow as follow, but i've got noise result
Could you explain the reason for me?
def synthesis_waveglow(text_seq, model, waveglow, alpha=1.0, mode=""):
denoiser = Denoiser(waveglow)
text = text_to_sequence(text_seq, hp.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)

pos = torch.stack([torch.Tensor([i+1 for i in range(text.size(1))])])
pos = pos.long().to(device)

model.eval()
with torch.no_grad():
    _, mel_postnet = model(text, pos, alpha=alpha)
with torch.no_grad():
    #wav = waveglow.infer(mel_postnet, sigma=0.666)
    wav = waveglow.infer(torch.transpose(mel_postnet,1,2).type(torch.cuda.HalfTensor), sigma=0.666)
print("Wav Have Been Synthesized.")

if not os.path.exists("results"):
    os.mkdir("results")

wav_denoised = denoiser(wav, strength=0.01)[:, 0]
#audio.save_wav(wav[0].data.cpu().numpy(), os.path.join(
#    "results", text_seq + mode + ".wav"))
audio.save_wav(wav_denoised[0].cpu().numpy(), os.path.join(
    "results", text_seq + mode + ".wav"))

Thank you

[TakoYuxin]

I tried TTS with WaveGlow as follow, but i've got noise result
Could you explain the reason for me?
def synthesis_waveglow(text_seq, model, waveglow, alpha=1.0, mode=""):
denoiser = Denoiser(waveglow)
text = text_to_sequence(text_seq, hp.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)

pos = torch.stack([torch.Tensor([i+1 for i in range(text.size(1))])])
pos = pos.long().to(device)

model.eval()
with torch.no_grad():
    _, mel_postnet = model(text, pos, alpha=alpha)
with torch.no_grad():
    #wav = waveglow.infer(mel_postnet, sigma=0.666)
    wav = waveglow.infer(torch.transpose(mel_postnet,1,2).type(torch.cuda.HalfTensor), sigma=0.666)
print("Wav Have Been Synthesized.")

if not os.path.exists("results"):
    os.mkdir("results")

wav_denoised = denoiser(wav, strength=0.01)[:, 0]
#audio.save_wav(wav[0].data.cpu().numpy(), os.path.join(
#    "results", text_seq + mode + ".wav"))
audio.save_wav(wav_denoised[0].cpu().numpy(), os.path.join(
    "results", text_seq + mode + ".wav"))

Thank you

I also got noisy results. I think the reason is WaveGlow uses a slightly different audio processing method so these two models are trained on different scaled mel spectrograms, thus not compatible. I tried to convert the mel spectrogram predicted by FastSpeech into the same scale used by WaveGlow but failed...... If you happened to get good results, please let me know

[angelorobo]

@TakoYuxin Can you please share how did you "convert the mel spectrogram predicted by FastSpeech into the same scale used by WaveGlow"?

[TakoYuxin]

@TakoYuxin Can you please share how did you "convert the mel spectrogram predicted by FastSpeech into the same scale used by WaveGlow"?

Here is what I added after FastSpeech predicted mel_postnet but this didn't work.

waveglow_npy = audio._db_to_amp(audio._denormalize(mel_postnet) + hp.ref_level_db)
waveglow_npy = torch.from_numpy(waveglow_npy)
waveglow_npy = torch.log(torch.clamp(waveglow_npy, min=1e-5) * 1)
torch.save(waveglow_npy, os.path.join(mode+ '.pt'))

you can find FastSpeech audio processing functions in audio.py and WaveGlow audio processing functions in NVIDIA/tacotron2/layers.py, audio_processing.py

[tingyang01]

Thank you @TakoYuxin .
I've got some result but the quality of voice is not good.
My result is in epoch 1000.
Should I do more the training of FastSpeech model?
Could you explain for me?

[TakoYuxin]

I haven't got any good results either >_< I trained the model for 198k steps but the generated voice was so not clear that I could barely understand what it was saying and a few words were skipped. I don't really know exactly how to fix this but continuing to train for more epochs sounds like a plan lol. We should probably wait for the author's answer.

[enamoria]

Anyone got good result with waveglow? I tried Tako's denormalize and got some audible result but still can't get anything better. Not even close to Grinffin Lim

[xcmyz]

The newest repo have audio example synthesized by waveglow in result.

[aguazul]

I am also having this issue - I trained to 164,000 steps but the wav file was just silent background noise.

I also tried running synthesis.py on earlier training steps (2000, 9000, 130,000) and the only one that sounded remotely like speech was 2000. By 9000 steps it was just an empty silent wav file.

I saw that @xcmyz said that batch size needed to be 32 or more, but if I set it to 32 I get an out of memory error despite running on a good GPU. I reduced to 16 batch size and it runs without error.

I am not sure why the audio wav files are basically silent noise.