I'm trying to create an acoustic model in brazilian portuguese, but I'm not able to train my model. Always the same problem
Running training
Configuration (e.g. ./etc/sphinx_train.cfg) not defined
Compilation failed in require at /usr/local/lib/sphinxtrain/scripts/000.comp_feat/slave_feat.pl line 51
BEGIN failed --compilation aborted at /usr/local/lib/sphinxtrain/scripts/000.comp_feat/slave_feat.pl line 51
anyone has any tips for me?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I only need to change the one on the myproject/etc/sphinx_train.cfg or do I also need to alter the one in /usr/local/lib/sphinxtrain/etc/sphinx_train.cfg ?? I believe I only need to alter the first one, but it seems odd.
~~edit~~ forget about it, just saw that the second one is [Read-Only]
Last edit: Guilherme Jordao 2018-01-10
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
ok, just updating here, when I train my model with the standard sphinx_train.cfg that is generated with sphinxtrain -t "myProject" setup, it runs, but with several errors, but that was expected because it wasn't with my especifications, but everytime I edit sphinx_train.cfg the file gets corrupted somehow or I mess up something, it's very odd.
I'll keep this updated
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
My texts had that, but I still have this error (this is the log file) :
INFO: ngram_model_trie.c(354): Trying to read LM in trie binary format
ERROR: "ngram_search.c", line 206: Language model/set does not contain , recognition will fail
FATAL: "batch.c", line 822: PocketSphinx decoder init failed
does it has something to do with pocketsphinx_batch??
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
can you take a look at my language model please??
I trained again yesterday and it fails the decoding process but it keeps training. I'm going to rebuild my lm, because I believe that it's the problem.
I tested the results with pocketsphinx_continuous using my lm.bin, my dic and my -hmm generated in model_paramethers and I got very bad results.
I'm going to use this mostly for hotword detection and I'm will transfer those files to my raspberry pi to how it will perform, but I'm going to be training another lm to use instead of this.
I found the "building a language model" tutorial a bit confusing, but I'll try again.
I tried to use SRILM but I had problems with the download.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It worked better than the last time, but there are a lot of errors yet.
37% correct and 63% errors
But I believe that this happened because I do not have enought training data, but that's fine by now. My question is: Is this going to work on Raspberry Pi3 to detect a hotword?
until the end of the day I'm going to upload the files and post the URL here
Last edit: Guilherme Jordao 2018-01-16
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It works ok with the model that come with pocketsphinx, but my model isn't good enought. I'll try to work with honk. thx so much for your help Nockolay, if I ever make this works, I'll post here.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm trying to create an acoustic model in brazilian portuguese, but I'm not able to train my model. Always the same problem
Running training
Configuration (e.g. ./etc/sphinx_train.cfg) not defined
Compilation failed in require at /usr/local/lib/sphinxtrain/scripts/000.comp_feat/slave_feat.pl line 51
BEGIN failed --compilation aborted at /usr/local/lib/sphinxtrain/scripts/000.comp_feat/slave_feat.pl line 51
anyone has any tips for me?
You probably miss setup step or you corrupted sphinx_train.cfg file while editing
I'll try again tomorrow. Maybe there is any relation with me using a VM to do this?
Definitely not. Such errors appear when you are not careful enough.
I only need to change the one on the myproject/etc/sphinx_train.cfg or do I also need to alter the one in /usr/local/lib/sphinxtrain/etc/sphinx_train.cfg ?? I believe I only need to alter the first one, but it seems odd.
~~edit~~ forget about it, just saw that the second one is [Read-Only]
Last edit: Guilherme Jordao 2018-01-10
ok, just updating here, when I train my model with the standard sphinx_train.cfg that is generated with sphinxtrain -t "myProject" setup, it runs, but with several errors, but that was expected because it wasn't with my especifications, but everytime I edit sphinx_train.cfg the file gets corrupted somehow or I mess up something, it's very odd.
I'll keep this updated
Use better editor.
ok, so all the process run fine until MODULE:DECODE.
I was reading here ( https://sourceforge.net/p/cmusphinx/discussion/help/thread/518d9ceb/?limit=25#ffde ) and it seems that the texts used to train the LM must have '
' and '' with them.My texts had that, but I still have this error (this is the log file) :
INFO: ngram_model_trie.c(354): Trying to read LM in trie binary format
ERROR: "ngram_search.c", line 206: Language model/set does not contain , recognition will fail
FATAL: "batch.c", line 822: PocketSphinx decoder init failed
does it has something to do with pocketsphinx_batch??
You can open lm arpa file with a text editor to check that required tags are really present.
Are you saying that every single word must be involved by tags?
like this:
-5.6113
amanda-0.1650The model on the web site doesn't have the tags on every word, just some.
No
can you take a look at my language model please??
I trained again yesterday and it fails the decoding process but it keeps training. I'm going to rebuild my lm, because I believe that it's the problem.
The LM is not quite correct, there are multiple things you ignored in tutorial, read it again and use srilm.
yeah, that's what I was afraid of...
I tested the results with pocketsphinx_continuous using my lm.bin, my dic and my -hmm generated in model_paramethers and I got very bad results.
I'm going to use this mostly for hotword detection and I'm will transfer those files to my raspberry pi to how it will perform, but I'm going to be training another lm to use instead of this.
I found the "building a language model" tutorial a bit confusing, but I'll try again.
I tried to use SRILM but I had problems with the download.
the model does not work on my raspberry pi.
I got an error, but i changed the feat.params file and edited the cmn to "current".
I finnaly was able to download and use SRILM to create a new LM, but it generated a LM file of almost 150MB. Was this supposed to happen?
Language models are usually large. You can prune them to reduce size.
update
It worked better than the last time, but there are a lot of errors yet.
37% correct and 63% errors
But I believe that this happened because I do not have enought training data, but that's fine by now. My question is: Is this going to work on Raspberry Pi3 to detect a hotword?
until the end of the day I'm going to upload the files and post the URL here
Last edit: Guilherme Jordao 2018-01-16
For simple hotword detection on Raspberry Pi it is better to use neural networks, something like
https://github.com/castorini/honk
I'm going to read about this, thx!
yeah, it won't work. I tryied to use thi code here >> https://stackoverflow.com/questions/34596576/how-can-i-detect-one-word-with-speech-recognition-in-python
It works ok with the model that come with pocketsphinx, but my model isn't good enought. I'll try to work with honk. thx so much for your help Nockolay, if I ever make this works, I'll post here.