DurIAN

Implementation of DurIAN: Duration Informed Attention Network For Multimodal Synthesis

Blog

论文笔记：腾讯AI lab多模态语音合成模型DurIAN

Structure

English:

1. I use the same encoder as tacotron2
2. I remove attention module in decoder and use average pooling to implement "predicting r frames at once"
3. I remove position encoding and skip encoder in this implementation

Chinese:

1. 我用了和tacotron2相同的encoder结构，但是参数更小
2. 我去除了decoder中的attention模块，由于一步输出三帧，我对三个时间步的memory进行了相加求均值的操作，表现在代码中为average pooling，经过实验，相比与存在attention模块的decoder，这样的音质会受到很小的负面影响，但是训练速度有了极大的提高
3. 我舍弃了position encoding和skip encoder，这对合成效果的影响很小

Sample & Pretrained model

sample here, I use waveglow as vocoder, pretrained model here, batchsize is 32, step is 180k.

Usage

training:

1. pip install -r requirements.txt
2. download and extract LJSpeech dataset
3. put LJSpeech dataset in data
4. unzip alignments.zip
5. python3 preprocess.py
6. CUDA_VISIBLE_DEVICES=0 python3 train.py

testing:

1. Put Nvidia pretrained waveglow model in the waveglow/pretrained_model
2. CUDA_VISIBLE_DEVICES=0 python3 test.py --step [step-of-checkpoint]

testing using pretrained model:

1. put pretrained model in model_new
2. CUDA_VISIBLE_DEVICES=0 python3 test.py --step 180000

Notes

尽管DurIAN的生成速度比不上FastSpeech，但是DurIAN生成的样本音质好于FastSpeech，并且计算量也小于FastSpeech，在实际部署中，DurIAN的生成速度已经完全满足RTF要求。