hello，I want to know how to change the count of 80000000？because when I just want to try to train the model，the feature extracting costed a lot of time. Hope you can answer it，thank you very much！

Hello, may I ask what is the problem that the generated training_16k_v3.f32 file size is 0

Hi guys,

I would like to know if someone faced some issues about the first step of features extraction with the denoise file, I'm trying to retrain the model with 16khz audio but it looks like after running denoise.c, I can't get the proper shape when I run bin2hdf5, I even checked the values of the training.f32 file and I got high values or NaN. I didn't modify the code except for the count, I fixed it to 500000, I used Microsoft Dataset available here : https://github.com/microsoft/MS-SNSD.

Maybe my datasets didn't fit with the code, can I ask you what datasets did you use for training your model, it would help me a lot @YongyuG

Wondering if you achieved better results with 16 Khz models compared to the original 46Khz models ?

音频数据和噪音数据已放在指定的文件夹。我修改了denoise.c中的count数量为10000。随后进行了compile编译，得到了denoise_training文件，当我使用该文件处理数据时会报错
count ==== 0
Floating point exception (core dumped)
执行脚本为：
./denoise_training /home/rnnoise_16k-master/src/media/clean_dir /home/rnnoise_16k-master/src/media/noise_dir 10000 > training_16k_v3.f32

但我使用您提供的denoise_training_gao脚本处理则不会报错，但也不会在指定的count数量后停止。

在编译denoise.c文件时，出现一下提示，不知道是不是编译失败导致数据处理程序不可用？

期待您的解答！

我可以在使用模型的时候使用帧长度为128吗，我看现有的模型是使用的帧长度是160。如果在使用模型的时候使用128帧长不合适，那么需要改网络训练的帧长，重新训练吗？

When make the CMakeLists.txt in the working directory, there are many “multiple definition" errors. Because the main.c and the denoise.c both have a main function (and some other same functions). How did you solve this error? thanks.

Thanks for sharing this 16 k model. However, there are still exist some questions that make me confusing.

(1) As far as I know, pitch feather is important. The original code about pitch analysis is conducted in 48 kHz. In your example, only the macros PITCH_MIN_PERIOD, PITCH_MAX_PERIOD and PITCH_FRAME_SIZE are revised. So, have you ever tested wheter the code could extract pitch correctly with only small revise?

(2) Have you ever test this model with the performance of PESQ? Using the original RNNOISE code, the PESQ is good, however, using the weight with my own training, it is not very good.

Thanks very much.

I use denoise.c to extract features. And I changed count to 100000, so it printed 100000*75 on the screen. However when I run bin2hdf5.py , an error occurs: "cannot reshape array of size 7555881 into shape (100000,75)" . Why? looking forward to your reply, thank you.

could you provide source code for denoise_training_gao?
I want to retrain this mode with new data.

Epoch trained for about 20 times, but got stuck during the training process and didn't continue training. What is the likely reason

the second problem is :

if I want to train my data ,in the run.sh,I need prepared a speech.wav and a noise.wav ,but,in the run.sh,mixed.wav, need I prepare ? Or,it will auto generate?

1、src/compile.sh
2、../src/denoise_training ./data/clean ./data/noise ./mixed.wav > training_16k_v3.f32
然后报了Floating point exception (core dumped)的错误，我没改动代码

Hi guys,

I would like to know what changes do I need to do if I want to train a model with 8khz audio, I tried to change some parameters in the denoise.c like this

#define BLOCK_SIZE 8000

#define FRAME_SIZE (20<<FRAME_SIZE_SHIFT)
#define SAMPLE_RATE 8000


#define PITCH_MIN_PERIOD 10
#define PITCH_MAX_PERIOD 128
#define PITCH_FRAME_SIZE 160 

#if SMOOTH_BANDS
//#define NB_BANDS 22
#define NB_BANDS 14
#else
//#define NB_BANDS 21
#define NB_BANDS 13
#endif

After compiling, I got a matrix with shape 500000 x 63, the number of features (63) is for me normal because we have less samples, and even the training is doing well, I do not get high loss or what, it seems to learn something. But at the end of the day, I get sizzle audio and it doesn't seems to do anything with noise reduction.

Sorry for bothering you again @YongyuG, can you provide me some informations about the code, how did you manage to turn it in 16khz and according to you, is it possible or not to turn it in 8khz.

I want to use noisy audio files to test the noise reduction effect of our trained model. How should we operate? I didn't see the relevant description. The source code of 48k is to put the input file before noise reduction and the output file after noise reduction into rnnoise_demo,

.### /examples/rnnoise_demo < input.raw > output.raw

but the file of rnnoise_demo is not seen in the 16k code. I hope the author can answer it. thank you @YongyuG

我想直接读16bit 16k的pcm进行处理，但是显示"check gain"，处理后的音频爆音，查看了作者的读取wav方法，发现使用f32读取后，值应该是【-1,1】，为什么还要在乘36728。
想问一下怎么修改才可以直接处理pcm不出错呢，谢谢

可以分享一下数据集的处理代码吗，刚刚接触语音，不太会做，期望您的回复

如果你训练的时候需要生成denoise_training的可执行文件，不是只要运行./compile.sh就行呀。。。

Compiling according to your instructions shows an error as shown. I know little about compilation. Can you elaborate on your own compilation process for me, thank you very much.

我的噪声数据有18.5h,纯净语音数据10.5h，wav格式，当我在终端执行./denoise_training clean_speech.wav noise_background.wav 50000000 > output.f32，没有成功提取特征，也没有报错，为什么

how this 300 should be in 16K model even in 8K (because I want to change it to 8k )
thanks