Comments (7)
mravanelli
commented on May 12, 2024
Hi,
are you using TIMIT? Which is the decoding graph you are using?
Mirco
…On Sat, 9 Feb 2019 at 12:34, Caoang327 ***@***.***> wrote:
Hi. Thanks for your work on such an amazing package.
I have trained and tested tri-phone model and it worked pretty well. Right
now I hope to train a DNN model using mono phoneme model per frame. I
change the cfg file as you mentioned and set the output number of MLP as
48. I finish the train and validating phrase. But when I try to decode it,
an error occurred saying that "48 mismatched with 144".
So I wonder how could I set the training and decode the output of Mono
phone DNN? I guess 144 may refer to the 144 states, but I don't understand
why I need these states for mono phoneme label per frame. And if we don't
need to decode for mono phoneme case, how could I evaluate on test data?
Thanks
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#48>, or mute the
thread
<https://github.com/notifications/unsubscribe-auth/AQGs1vmfPC87D1_lkHJhLt73GJNG1qFDks5vLwa-gaJpZM4aypVa>
.
from pytorch-kaldi.
Caoang327
commented on May 12, 2024
Hi, I am using TIMIT. The setting is:
lab_folder=/home/ubuntu/kaldi/egs/timit/s5/exp/mono_ali
lab_opts=ali-to-phones --per-frame=true
lab_count_file=none
lab_data_folder=/home/ubuntu/kaldi/egs/timit/s5/data/train/
lab_graph=/home/ubuntu/kaldi/egs/timit/s5/exp/mono/graph
Mono is generated by kaldi s5 run.sh
from pytorch-kaldi.
mravanelli
commented on May 12, 2024
Hi,
this is a weird setup (usually the decoding is done with context-dependent
graphs). Is there a particular reason why you are using it?
However, since you are directly using monophone labels you should use:
lab_opts=ali-to-pdf
you also have to set the number of output to 144. 144 is the number of
state (each of the 48 phonemes are described by 3 states).
Please, let me know if this solves the problem...
Mirco
…On Sat, 9 Feb 2019 at 13:07, Caoang327 ***@***.***> wrote:
Hi, I am using TIMIT. The setting is:
lab_folder=/home/ubuntu/kaldi/egs/timit/s5/exp/mono_ali
lab_opts=ali-to-phones --per-frame=true
lab_count_file=none
lab_data_folder=/home/ubuntu/kaldi/egs/timit/s5/data/train/
lab_graph=/home/ubuntu/kaldi/egs/timit/s5/exp/mono/graph
Mono is generated by kaldi s5 run.sh
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#48 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AQGs1kcavdKCHuYVl6xkOnYCo-SlLU4tks5vLw5EgaJpZM4aypVa>
.
from pytorch-kaldi.
Caoang327
commented on May 12, 2024
Hi. I hope to translate the output of DNN by wireless communication so I hope the output number to be small enough. And I'm curious about the performance between mono phone and CTC based strategy.
I'm curious about your "directly using monophone labels". Does it mean I could also set lab_folder=/home/ubuntu/kaldi/egs/timit/s5/exp/dnn4_pretrain-dbn_dnn_ali and use lab_opts=ali-to-phones --per-frame=true? If it is feasible, how should I set graph? exp/mono/graph or exp/tri3/graph?
If I set the number of output to 144, how should I set the loss_mono=cost_nll(out_dnn3,lab_mono)? I guess lab_mono may be 48 dim. So how could I write this part?
Thank you very much
from pytorch-kaldi.
mravanelli
commented on May 12, 2024
To decode with monopone you should use a monophone graph and I think the
right way is what I explained in the previous mail. Let me know if it works.
Which of the available examples are you talking about? In the available
examples the monophones are onyl use to regulariza the DNN with the
so-called monophone reguarization, but the decoding is done with
context-dependent models..
Mirco
…On Sat, 9 Feb 2019 at 13:32, Caoang327 ***@***.***> wrote:
Hi. I hope to translate the output of DNN by wireless communication so I
hope the output number to be small enough. And I'm curious about the
performance between mono phone and CTC based strategy.
I'm curious about your "directly using monophone labels". Does it mean I
could also set
lab_folder=/home/ubuntu/kaldi/egs/timit/s5/exp/dnn4_pretrain-dbn_dnn_ali
and use lab_opts=ali-to-phones --per-frame=true? If it is feasible, how
should I set graph? exp/mono/graph or exp/tri3/graph?
If I set the number of output to 144, how should I set the
loss_mono=cost_nll(out_dnn3,lab_mono)? I guess lab_mono may be 48 dim. So
how could I write this part?
Thank you very much
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#48 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AQGs1tVqlXqajk6gaE4-wpfDx1DE1g98ks5vLxRNgaJpZM4aypVa>
.
from pytorch-kaldi.
Caoang327
commented on May 12, 2024
Thank you. I will try it.
from pytorch-kaldi.
mravanelli
commented on May 12, 2024
you are welcome!
…On Sat, 9 Feb 2019 at 14:03, Caoang327 ***@***.***> wrote:
Thank you. I will try it.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#48 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AQGs1osEimKUIOg77IBXeUuW-qLBp9FXks5vLxt6gaJpZM4aypVa>
.
from pytorch-kaldi.
Related Issues (20)
- How to setup parameters in "cfg/TIMIT_baselines/TIMIT_liGRU_fmllr.cfg"? HOT 1
- Do bidirectional layers share the input-to-hidden weights? HOT 2
- Can we resume training from the epoch we got interruption HOT 4
- input shape of nns HOT 3
- Question about the Dimension of wx.0.weight in my mlp model HOT 1
- The loss curve of train and dev is reasonable but why the Test Error keeps 53% or so? HOT 8
- Support for torch.nn.Transformer Class? HOT 1
- KaldiFatalError during decoding phase
- No WER stdout when decoding
- Does pytorch-kaldi support chain model training? HOT 1
- Word transcription of TIMIT dataset HOT 1
- No Decoding Output HOT 20
- How to train/decode on reverberant speech? HOT 1
- x-vector DNN model
- Unable to run forwarding step on test set
- Before switch to SpeechBrain, how to use trained model in pytorch
- Use final_architecture1.pkl for live test HOT 4
- err_te is 1
- using different features instead of FMLLR
- res.res
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 pytorch-kaldi.