This is a PyTorch implementation of our work "FloWaveNet : A Generative Flow for Raw Audio".
For the purpose of parallel sampling, we propose FloWaveNet, a flow-based generative model for raw audio synthesis. FloWaveNet can generate audio samples as fast as ClariNet and Parallel WaveNet, while its training procedure is really easy and stable. Our generated audio samples are available at http://bit.ly/2zpsElV. Also, our implementation of ClariNet (Gaussian WaveNet and Gaussian IAF) is available at https://github.com/ksw0306/ClariNet
PyTorch 0.4.1 & python 3.6 & Librosa
- LJSpeech : https://keithito.com/LJ-Speech-Dataset/
python preprocessing.py --in_dir ljspeech --out_dir DATASETS/ljspeech
python train.py --model_name flowavenet --batch_size 8 --n_block 8 --n_flow 6 --n_layer 2 --causal no
--load_step CHECKPOINT # of Pre-trained model
--temp TEMPERATURE / z ~ N(0, 1 * TEMPERATURE)
ex) python synthesize.py --model_name flowavenet --n_block 8 --n_flow 6 --n_layer 2 --causal no --load_step 100000 --temp 0.7 --num_samples 10
Sample Link : http://bit.ly/2zpsElV
Our implementation of ClariNet (Gaussian WaveNet, Gaussian IAF) : https://github.com/ksw0306/ClariNet
-
Results 1 : Model Comparisons (WaveNet (MoL, Gaussian), ClariNet and FloWaveNet)
-
Results 2 : Temperature effect on Audio Quality Trade-off (Temperature T : 0.0 ~ 1.0, Model : Gaussian IAF and FloWaveNet)
-
Results 3 : Analysis of ClariNet Loss Terms (Loss functions : 1. KLD + Frame Loss 2. Only KL 3. Only Frame)
-
Results 4 : Context Block and Long term Dependency (FloWaveNet : 8 Context Blocks, FloWaveNet_small : 6 Context Blocks)
-
Results 5 : Causality of WaveNet Dilated Convolutions (FloWaveNet : Non-causal WaveNet Affine Coupling Layers, FloWaveNet_causal : Causal WaveNet Affine Coupling Layers)
- WaveNet vocoder : https://github.com/r9y9/wavenet_vocoder
- glow-pytorch : https://github.com/rosinality/glow-pytorch