Comments (5)
snsun
commented on August 13, 2024
Hi，
No, I haven't met this problem when I implemented PIT training. I think,
maybe you need to check your code again. The most possible reason is that
you PIT loss function implementation has bug.
Best
Xinyao(Alvin) Sun <[email protected]> 于2018年11月28日周三 上午10:33写道：
… Hi, the issue is not for your code, I am doing some research on speech
separation and I implemented basic PIT model (CNN, RNN), however, I always
face an issue that two masks trend to have equal values during the
training. It results in separation totally not work. I just wonder whether
you have ever faced a similar issue when you implemented the PIT algorithm?
Thanks
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#21>, or mute the
thread
<https://github.com/notifications/unsubscribe-auth/AMS4KSCl07nimScsNNzr-XZr8QGJRe1eks5uzfYFgaJpZM4Y2vJD>
.
from pit-speech-separation.
Lucklyric
commented on August 13, 2024
Thanks for your reply.
In fact that, I am quite confused about what's the problem.
The main part of PIT is not complicated. I also tried to replace the PIT part with your code and got the same issue.
The main difference I am sure is the way to mix data. I am using
Voxceleb dataset http://www.robots.ox.ac.uk/~vgg/data/voxceleb/. The way I created input data is using librosa's STFT and I only use a log of magnitude as the input of the network. I directly add two speaker's log magnitude as the mix data.
I do not have too many experiences on audio analysis before, do you think the way I created training data may cause some potential issues?
Thanks
from pit-speech-separation.
snsun
commented on August 13, 2024
Instead of using log magnitude, we used magnitude directly. If you use
log-magnitude, you can not multiply mask by the noisy log-magnitude
directly. I recommended you to use magnitude.
Best
Xinyao(Alvin) Sun <[email protected]> 于2018年11月28日周三 上午11:47写道：
… Thanks for your reply.
In fact that, I am quite confused about what's the problem.
The main part of PIT is not complicated. I also tried to replace the PIT
part with your code and got the same issue.
The main difference I am sure is the way to mix data. I am using
Voxceleb dataset http://www.robots.ox.ac.uk/~vgg/data/voxceleb/. The way
I created input data is using librosa's STFT and I only use a log of
magnitude as the input of the network. I directly add two speaker's log
magnitude as the mix data.
I do not have too many experiences on audio analysis before, do you think
the way I created training data may cause some potential issues?
Thanks
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#21 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AMS4KaE1Q8Nt50H38cDiC3IWfZ_RrGFNks5uzgddgaJpZM4Y2vJD>
.
from pit-speech-separation.
Lucklyric
commented on August 13, 2024
Thanks for your suggestion.
The reason why I did log is that the raw amplitude is distributed very unbalance. I did try using raw amplitude and it did not work. Because I do not have WSJ dataset, could you tell me what kind of distribution of your raw amplitude data? For mine it looks like this,
Sorry for disturbing you a lot, I just wanna to figure out the reason that casuses my implemente does work.
Thanks,
from pit-speech-separation.
snsun
commented on August 13, 2024
I didn't observe the distribution of WSJ data. But I believe the
distributions should be similar between different datasets.
From the info you have provided, I don't think I can figure out the reason.
How about your training and development sets loss?
Best
Xinyao(Alvin) Sun <[email protected]> 于2018年11月28日周三 下午2:41写道：
… Thanks for your suggestion.
The reason why I did log is that the raw amplitude is distributed very
unbalance. I did try using raw amplitude and it did not work. Because I do
not have WSJ dataset, could you tell me what kind of distribution of your
raw amplitude data? For mine it looks like this,
[image: image]
<https://user-images.githubusercontent.com/7803070/49133382-1e7b8980-f29d-11e8-873f-64b028c00d0c.png>
Sorry for disturbing you a lot, I just wanna to figure out the reason that
casuses my implemente does work.
Thanks,
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#21 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AMS4KdtttmjqN1MyYanYPA7I39ecBMTLks5uzjAJgaJpZM4Y2vJD>
.
from pit-speech-separation.
Related Issues (20)
- there is a small error HOT 1
- hi,can you offer me some data for train,i am new to the LSTM HOT 5
- Can I add you wechat to discuss this work?
- How to convert feats_mapping.lst to tfrecords? HOT 4
- a littile HOT 2
- tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimensions must be equal, but are 1488 and 992 for 'model/blstm/stack_bidirectional_rnn/cell_0/bidirectional_rnn/fw/fw/while/fw/basic_lstm_cell/MatMul_2' (op: 'MatMul') with input shapes: [25,1488], [992,1984]. HOT 2
- A problem when preparing my own data for TFRecords format. HOT 3
- Hello, I have a question... HOT 1
- Is this repo for training or training+speech separation? HOT 1
- Contact
- Dataset structure
- run.sh HOT 1
- what‘s meta-frame's meaning? HOT 1
- I have some problems with executing code. HOT 1
- How to separate the target speech? HOT 5
- dataset
- list index out of range
- wv version of wsj0 dataset
- Code clarity: Permutation for minimum loss HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from pit-speech-separation.