wavernn
所属分类:数值算法/人工智能
开发工具:HTML
文件大小:20585KB
下载次数:0
上传日期:2018-08-19 07:38:29
上 传 者:
sh-1993
说明: 基于wavernn的波形发生器和TensorFlow CuDNN GRU使用演示。
(wavernn,WaveRNN-based waveform generator & demo of TensorFlow CuDNN-GRU usage.)
文件列表:
cudnn_gru.html (287991, 2018-08-19)
cudnn_gru.ipynb (15751, 2018-08-19)
img (0, 2018-08-19)
img\gru_small.png (13048, 2018-08-19)
img\pipeline.png (99408, 2018-08-19)
models (0, 2018-08-19)
models\frozen.pb (20825136, 2018-08-19)
outputs (0, 2018-08-19)
outputs\audio.wav (619258, 2018-08-19)
requirements.txt (76, 2018-08-19)
run_wavernn.py (3345, 2018-08-19)
samples (0, 2018-08-19)
samples\LJ016-0277.wav (664230, 2018-08-19)
samples\LJ046-0252.wav (851622, 2018-08-19)
spectrogram.png (61353, 2018-08-19)
# wavernn-demo
Since this work was part of a private Insight fellowship project, the model
structure and training code could not be open-sourced.
![Model](https://github.com/austinmoehle/wavernn/blob/master/img/pipeline.png)
This repository contains a demo of a WaveRNN-based waveform generator trained
on 13000 spoken sentences; inference can be run on CPU or GPU using the frozen
graph.
`run_wavernn.py` takes an input WAV file, applies an FFT to produce an 80-band
mel spectrogram, then uses the spectrogram to generate 16 kHz audio with a
frozen WaveRNN model.
`cudnn_gru.ipynb` demonstrates usage of the (poorly documented) CuDNN GRU cell
in TensorFlow, and some of the tricks and workarounds needed to get the network
up and running. See [this link](https://github.com/austinmoehle/wavernn/blob/master/http://htmlpreview.github.io/?https://github.com/austinmoehle/wavernn/blob/master/cudnn_gru.html)
for an HTML-rendered version.
## Setup
Create a virtualenv, activate then run
```
pip install -r requirements.txt
```
## Run
Choose an input WAV file and run, e.g.
```
python run_wavernn.py samples/LJ016-0277.wav
```
## CuDNN GRU in TensorFlow - Code and Demonstration
For this model, Nvidia's highly optimized CuDNN GRU trained __7x__ faster
than a slow TensorFlow while-loop with explicit TF operations. Unfortunately,
I couldn't get the `CudnnGRU` module to work out-of-the-box so I resorted to
some workarounds to make the cell usable. The Jupyter
[notebook](https://github.com/austinmoehle/wavernn/blob/master/cudnn_gru.ipynb) contains code snippets from the original model,
along with an example of how to replicate the CuDNN GRU cell in native
TensorFlow.
Requires __tensorflow-gpu__.
近期下载者:
相关文件:
收藏者: