neural-homomorphic-vocoder 0.0.8

Pytorch implementation of neural homomorphic vocoder

neural-homomorphic-vocoder

A neural vocoder based on source-filter model called neural homomorphic vocoder

Install

pip install neural-homomorphic-vocoder

Usage

Usage for NeuralHomomorphicVocoder class

• Input
  • x: mel-filterbank
  • cf0: continuous f0
  • uv: u/v symbol

import torch
from nhv import NeuralHomomorphicVocoder

net = NeuralHomomorphicVocoder(
        fs=24000,             # sampling frequency
        fft_size=1024,        # size for impuluse responce of LTV
        hop_size=256,         # hop size in each mel-filterbank frame
        in_channels=80,       # input channels (i.e., dimension of mel-filterbank)
        conv_channels=256,    # channel size of LTV filter
        ccep_size=222,        # output ccep size of LTV filter      
        out_channels=1,       # output size of network
        ccep_size=222,        # output size of LTV filter
        kernel_size=3,        # kernel size of LTV filter
        dilation_size=1,      # dilation size of LTV filter
        group_size=8,         # group size of LTV filter
        fmin=80,              # min freq. for melspc 
        fmax=7600,            # max freq. for melspc (recommend to use full-band)
        roll_size=24,         # frame size to get median to estimate logspc from melspc
        look_ahead=32,        # # of look_ahead samples (if use_causal=True)
        n_ltv_layers=3,       # # layers for LTV ccep generator
        n_postfilter_layers=4,   # # layers for output postfilter 
        use_causal=False,        # use causal conv LTV filter
        use_reference_mag=False, # use reference logspc calculated from melspc
        use_tanh=False,       # apply tanh to output else linear
        use_uvmask=False,     # apply uv-based mask to harmonic
        use_weight_norm=True, # apply weight norm to conv1d layer
        conv_type="original"  # ltv generator network type ["original", "ddsconv"]
        postfilter_type=None, # postfilter network type ["None", "normal", "ddsconv"]
        ltv_postfilter_type="conv",  # ltv postfilter network type \
                                     # ["None", "normal", "ddsconv"]
        scaler_file=None      # path to .pkl for internal scaling of melspc
                              # (dict["mlfb"] = sklearn.preprocessing.StandardScaler)
)

B, T, D = 3, 100, in_channels   # batch_size, frame_size, n_mels
z = torch.randn(B, 1, T * hop_size)
x = torch.randn(B, T, D)
cf0 = torch.randn(B, T, 1)
uv = torch.randn(B, T, 1)
y = net(z, torch.cat([x, cf0, uv], dim=-1))   # z: (B, 1, T * hop_size), c: (B, D+2, T)
y = net._forward(z, cf0, uv)

Features

• (2021/05/21): Train using kan-bayashi/ParallelWaveGAN with continuous F1 and uv symbols
• (2021/05/24): Final FIR filter is implemented by 1D causal conv
• (2021/06/17): Implement depth-wise separable convolution

References

@article{liu20,
  title={Neural Homomorphic Vocoder},
  author={Z.~Liu and K.~Chen and K.~Yu},
  journal={Proc. Interspeech 2020},
  pages={240--244},
  year={2020}
}