FastSpeech 2 about tensorflowtts HOT 12 CLOSED

@loretoparisi yes, i have a plan :))). But because it's a new paper so i won't public it soon when i finished implementation :D.

from tensorflowtts.

Hi, I am reanding Fastspeech2 paper. could you help to explain how to perform quantize F0 and energy with log-scale bins for f0 and uniform bins for energy? @dathudeptrai

from tensorflowtts.

@dathudeptrai @superhg2012 Check this :
Energy quantize at Uniform bins

# Extract RMS energy
# y is raw wavefom
S = librosa.magphase(librosa.core.stft(y, n_fft=1024, hop_length=256))[0]
e = librosa.feature.rms(S=S)
bins = np.linspace(e.min(), e.max(), num=256)
e_quantize = np.digitize(e, bins)
e_quantize = torch.from_numpy(e_quantize-1).to(torch.device("cuda"))   # e_quantize-1 is to convert 1 to 256 --> 0 to 255
one_hot_e = F.one_hot(e_quantize.long(), 256).float()
one_hot_e.shape # torch.Size([1, 654, 256])

For Pitch

# Extract Pitch/f0 from raw waveform using PyWORLD
y = y.astype(np.float64)
f0, timeaxis = pw.harvest(y, 22050,f0_ceil=8000.0, frame_period=11.6) # For hop size 256 frame period is 11.6 ms
f0[f0 == 0] = 1 # Because log(0) is -ve infinite
log_f0 = np.log(f0)
bins = np.linspace(log_f0.min(), log_f0.max(), num=256)
p_quantize = np.digitize(log_f0, bins)
p_quantize = torch.from_numpy(p_quantize-1).to(torch.device("cuda"))
one_hot_p = F.one_hot(p_quantize.long(), 256).float()
one_hot_p.shape # torch.Size([654, 256])

from tensorflowtts.

@superhg2012 i haven't read paper in detail yet. This is what i think, when you compute F0 (f0 value in range 0 - 100 for example) then we have 10 bins (f0 from 0 - 9 in bin 1, f0 from 10-19 in bin 2...) then we will learn 10 f0 embeddings for each bin :))). The procedure is the same for energy. But note that before we separate bins we need apply function F to F0 and energy to rescale it, seems paper use log function :D

from tensorflowtts.

Is the latest released FastSpeech V3 = FastSpeech 2?

from tensorflowtts.

@ZDisket nope :)). it's fastspeech V3 :(. I'm trying to find free time to implement fastspeech v2, fastspeech v2 is not hard to implement, almost layers already implemented on this framework

from tensorflowtts.

@superhg2012 i haven't read paper in detail yet. This is what i think, when you compute F0 (f0 value in range 0 - 100 for example) then we have 10 bins (f0 from 0 - 9 in bin 1, f0 from 10-19 in bin 2...) then we will learn 10 f0 embeddings for each bin :))). The procedure is the same for energy. But note that before we separate bins we need apply function F to F0 and energy to rescale it, seems paper use log function :D

In FastSpeech2 paper, they quantize F0 and energy of each frame to 256 possible values.

from tensorflowtts.

@superhg2012 yes, after you apply log function to F0, u can quantize it to any posible values. 256 is just 1 option :D. Example after apply log, min f0 is 0 and max f0 is 20, so you can calculate the range of each bins ((20 - 0) / 256) = 0.078125. So all f0 values in range 0, 0.078125 will quantize to 0, f0 values in range 0.078125, 0.078125*2 will quantize to 1, ... then we just need to learn 256 f0 embeddings.

from tensorflowtts.

@dathudeptrai @rishikksh20 get it! thanks very much

from tensorflowtts.

@rishikksh20 @superhg2012 note that the min and max should be compute on all traing set :))).

from tensorflowtts.

Yes

from tensorflowtts.

@dathudeptrai you mean the min and max should be computed from all dataset and then use it to compute bins and then apply bin to each wave?

from tensorflowtts.