rnnoise
所属分类:人工智能/神经网络/深度学习
开发工具:C
文件大小:3434KB
下载次数:3
上传日期:2020-09-13 13:57:21
上 传 者:
sh-1993
说明: rnnoise,一个基于递归神经网络的噪声抑制库
(rnnoise,A noise suppression library based on a recurrent neural network)
文件列表:
AUTHORS (37, 2020-09-13)
COPYING (1581, 2020-09-13)
Makefile (1583, 2020-09-13)
doc (0, 2020-09-13)
doc\Doxyfile.in (518, 2020-09-13)
examples (0, 2020-09-13)
examples\bertool.c (905, 2020-09-13)
examples\rnnoise_demo.c (5271, 2020-09-13)
include (0, 2020-09-13)
include\rnnoise.h (2366, 2020-09-13)
sample.wav (3564558, 2020-09-13)
src (0, 2020-09-13)
src\_kiss_fft_guts.h (6010, 2020-09-13)
src\arch.h (7628, 2020-09-13)
src\celt_lpc.c (7780, 2020-09-13)
src\celt_lpc.h (2040, 2020-09-13)
src\common.h (1232, 2020-09-13)
src\denoise.c (18520, 2020-09-13)
src\kiss_fft.c (16990, 2020-09-13)
src\kiss_fft.h (6352, 2020-09-13)
src\opus_types.h (4954, 2020-09-13)
src\pitch.c (14119, 2020-09-13)
src\pitch.h (4682, 2020-09-13)
src\rnn.c (6197, 2020-09-13)
src\rnn.h (2213, 2020-09-13)
src\rnn_data.c (1159, 2020-09-13)
src\rnn_data.h (545, 2020-09-13)
src\rnn_reader.c (5192, 2020-09-13)
src\rnn_train.py (2047, 2020-09-13)
src\rnn_weights.c (527208, 2020-09-13)
src\tansig_table.h (2304, 2020-09-13)
training (0, 2020-09-13)
training\bin2hdf5.py (299, 2020-09-13)
training\dump_rnn.py (3710, 2020-09-13)
training\rnn_train.py (4611, 2020-09-13)
# RNNoise
A noise suppression library based on a recurrent neural network.
## Build
Prerequisite
* macOS: `brew install libsndfile libsoxr sox`
* Debian/Ubuntu: `sudo apt install libsndfile1-dev libsoxr-dev libsox-dev`
To compile, just type:
```shell
make
```
Sample noisy file `sample.wav` was included, and you can run `make check` to generate the processed one, `clean.wav`.
## Test
While it is meant to be used as a library, a simple command-line tool is
provided as an example. It can be used as:
```shell
examples/rnnoise_demo sample.wav output.wav
```
## Training
### Audio feature extract
Build audio feature extraction tool
```shell
make src/denoise_training
```
Use the tool `denoise_training` to get the audio feature array from speech and noise audio clip
```shell
src/denoise_training signal.raw noise.raw count > training.f32
```
(note the matrix size and replace 500000 87 below)
### RNN model traning
Pick feature array to "training" dir and go through the training process.
```shell
cd training ; ./bin2hdf5.py ../src/training.f32 500000 87 training.h5
./rnn_train.py
./dump_rnn.py weights.hdf5 ../src/rnn_data.c ../src/rnn_data.h
```
Training process will generate the RNN model weight code file (default is `rnn_data.c`) and layer definition header file (default is `rnn_data.h`).
They can be used to refresh the `src/rnn_data.c`, `src/rnn_data.h` and rebuild the rnnoise library and/or examples.
## License
`rnnoise` is freely redistributable under the revised BSD license.
Use of this source code is governed by a BSD-style license that can be found in the `COPYING` file.
近期下载者:
相关文件:
收藏者: