WaveNODE: A Continuous Normalizing Flow for Speech Synthesis
Abstract
In recent years, various flow-based generative models have been proposed to generate high-fidelity waveforms in real-time. However, these models require either a well-trained teacher network or a number of flow steps making them memory-inefficient. In this paper, we propose a novel generative model called WaveNODE which exploits a continuous normalizing flow for speech synthesis. Unlike the conventional models, WaveNODE places no constraint on the function used for flow operation, thus allowing the usage of more flexible and complex functions. Moreover, WaveNODE can be optimized to maximize the likelihood without requiring any teacher network or auxiliary loss terms. We experimentally show that WaveNODE achieves comparable performance in terms of mean opinion score (MOS) and conditional log-likelihood (CLL) with fewer parameters compared to the existing models.
Audio Examples
Example 1
Ground Truth
Griffin-Lim
WaveNet
WaveGlow
WaveGlow (17.1M)
FloWaveNet
FloWaveNet (18.6M)
WaveNODE
Example 2
Ground Truth
Griffin-Lim
WaveNet
WaveGlow
WaveGlow (17.1M)
FloWaveNet
FloWaveNet (18.6M)
WaveNODE
Example 3
Ground Truth
Griffin-Lim
WaveNet
WaveGlow
WaveGlow (17.1M)
FloWaveNet
FloWaveNet (18.6M)
WaveNODE
Example 4
Ground Truth
Griffin-Lim
WaveNet
WaveGlow
WaveGlow (17.1M)
FloWaveNet
FloWaveNet (18.6M)
WaveNODE
Ablation Studies
1. Test Tolerance
Example 1
test_tol: 1E-5
test_tol: 1E-4
test_tol: 1E-3
test_tol: 5E-3
test_tol: 1E-2
Example 2
test_tol: 1E-5
test_tol: 1E-4
test_tol: 1E-3
test_tol: 5E-3
test_tol: 1E-2
Example 3
test_tol: 1E-5
test_tol: 1E-4
test_tol: 1E-3
test_tol: 5E-3
test_tol: 1E-2
Example 4
test_tol: 1E-5
test_tol: 1E-4
test_tol: 1E-3
test_tol: 5E-3
test_tol: 1E-2
2. Norm Layer
Example 1
Actnorm
MBN
None
Example 2
Actnorm
MBN
None
Example 3
Actnorm
MBN
None
Example 4
Actnorm
MBN
None
3. Batch Size
Example 1
batch_size: 20
batch_size: 4
Example 2
batch_size: 20
batch_size: 4
Example 3
batch_size: 20
batch_size: 4
Example 4
batch_size: 20
batch_size: 4
4. Dilation
Example 1
dilation: 3i
dilation: 2i
Example 2
dilation: 3i
dilation: 2i
Example 3
dilation: 3i
dilation: 2i
Example 4
dilation: 3i
dilation: 2i