Hi all. Excuse me for my long post. I'm new to Sphinx. My aim is to make Sphinx understanding Italian Language. Obviously I don't expect this is a simple task. So, to understand Sphinx guidlines, I decided to begin with only three words: uno (one), due (two) and tre (three). So I began training Sphinx with SphinxTrain. I registered five wavefile in which I say these three words. Then I did created the files "unoduetre_train.fileids" "unoduetre.filler" unoduetre.phone" "unoduetre_train.transcription" unoduetre.dic". I made feats and run scripts_pl/runAll.pl. So my first result is a set of files describing (I hope) my continuous acoustic model made by three words. First question: after training I have 9 directories with model parameters and 1 directory with model architecture, so which are the files necessary to define the model? My answer is: the last file created in model architecture (in my example unoduetre.1000.mdef) and the last files created
in model parameter directory unoduetre.cd_cont_1000_8*).
Now I'm trying to use this acoustic model in Sphinx 4. To avoid interferences between my model and the default Sphinx models I've compiled only my model modifying build.xml. I've modified wavfile demo to make it using only my model but I missed something, cause it doesn't recognize the right words. I've used a simple wordlist grammar with the three words "uno","due","tre". So my second (and last) question: to make Sphinx4 understand and recognize these three simple words I need something else? Or is it only configuration issue?
Thank you in advance for your help.
greetings
Matteo Villari
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have the same problem. I want to make model for Slovenian language.I for training the model I'm using SphinxTrain package. So I will be very glad too if someone explains how to handle SphinxTrain. To be more specific, I have a problem with bw command.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I solved most of my problems. Errors during training were caused by a simple thing: I didn't leave a final blank row in files required for training (!).
Regards
Matteo Villari
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have trained an acoustic model based on the AN4 dataset,then modified an4 performence test to use my acoustic model. the test passed successfuly giving %100 recognition. that encourged me to recored 68 short utternces (about 4 min in total) containing only 4 words spoken in different orders. I was successful in training a acoustic model, but when this model is used it gives strang results. It keeps predicting the same word for all utternces, more starngly than that, when I turn the isLooping property of the simpelword grammer off and on it predicts a different word. Do you have any idea whats going on ?!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
5 wavefiles are far from enough to train a new acoustic model.....
Ok. This is what you usually need to do in building a recognizer for a new language disregard which sphinxes are you using.
1, Collect data for a particular language. If you are interested, I could tell you more. You usually need to use quite a lot of data to train a model in the first place.
3, Given the model, you would need to plug them in the decoder. Depends on the decoders of the choise. Different things need to be done. Sphinx 4, for example, require conversions from Sphinx 3's model to a specific format of S4(essentially a jar package).
thank you very much for your reply.
If you have time I would be very interested in understanding the basics (in the beginning) of Sphinx.
I read once again the document you have suggested me but I'm again a little bit confused. I made AN4 train and tried to live decode with Sphinx 4.
it do not worked very well but my English is not perfect so i don't mind for now cause i have a new problem in training. When I run the script 02.ci_schmm\slave_convg.pl I have the following error:
FATAL_ERROR: C:\new\SphinxTrain\src\libs\libio\corpus.c, line 256: input string too long. Truncated.
So I can't continue my training.
Now I have ten files (i know that are not enaugh even for few words, but i am interested in understanding the basics) in which I say the three words mixed for each file. I've searched in the forums but I didn't find solutions.
thank you a lot for your help.
Greetings,
Matteo Villari
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi all. Excuse me for my long post. I'm new to Sphinx. My aim is to make Sphinx understanding Italian Language. Obviously I don't expect this is a simple task. So, to understand Sphinx guidlines, I decided to begin with only three words: uno (one), due (two) and tre (three). So I began training Sphinx with SphinxTrain. I registered five wavefile in which I say these three words. Then I did created the files "unoduetre_train.fileids" "unoduetre.filler" unoduetre.phone" "unoduetre_train.transcription" unoduetre.dic". I made feats and run scripts_pl/runAll.pl. So my first result is a set of files describing (I hope) my continuous acoustic model made by three words. First question: after training I have 9 directories with model parameters and 1 directory with model architecture, so which are the files necessary to define the model? My answer is: the last file created in model architecture (in my example unoduetre.1000.mdef) and the last files created
in model parameter directory unoduetre.cd_cont_1000_8*).
Now I'm trying to use this acoustic model in Sphinx 4. To avoid interferences between my model and the default Sphinx models I've compiled only my model modifying build.xml. I've modified wavfile demo to make it using only my model but I missed something, cause it doesn't recognize the right words. I've used a simple wordlist grammar with the three words "uno","due","tre". So my second (and last) question: to make Sphinx4 understand and recognize these three simple words I need something else? Or is it only configuration issue?
Thank you in advance for your help.
greetings
Matteo Villari
I have the same problem. I want to make model for Slovenian language.I for training the model I'm using SphinxTrain package. So I will be very glad too if someone explains how to handle SphinxTrain. To be more specific, I have a problem with bw command.
I solved most of my problems. Errors during training were caused by a simple thing: I didn't leave a final blank row in files required for training (!).
Regards
Matteo Villari
I have trained an acoustic model based on the AN4 dataset,then modified an4 performence test to use my acoustic model. the test passed successfuly giving %100 recognition. that encourged me to recored 68 short utternces (about 4 min in total) containing only 4 words spoken in different orders. I was successful in training a acoustic model, but when this model is used it gives strang results. It keeps predicting the same word for all utternces, more starngly than that, when I turn the isLooping property of the simpelword grammer off and on it predicts a different word. Do you have any idea whats going on ?!
5 wavefiles are far from enough to train a new acoustic model.....
Ok. This is what you usually need to do in building a recognizer for a new language disregard which sphinxes are you using.
1, Collect data for a particular language. If you are interested, I could tell you more. You usually need to use quite a lot of data to train a model in the first place.
2, Train a model using the data. There are several comprehensive guides on how to do it including
Rita's training documents
http://fife.speech.cs.cmu.edu/sphinxman/fr4.html
Tutorial created by Robust Group,
http://www.speech.cs.cmu.edu/sphinx/tutorial.html
A caveat/FAQ by me,
http://www.cs.cmu.edu/~archan/10CommonPitfalls_ST.html
3, Given the model, you would need to plug them in the decoder. Depends on the decoders of the choise. Different things need to be done. Sphinx 4, for example, require conversions from Sphinx 3's model to a specific format of S4(essentially a jar package).
http://cmusphinx.sourceforge.net/sphinx4/javadoc/edu/cmu/sphinx/linguist/acoustic/tiedstate/Sphinx3Loader.html
matjaz, I knewn you have problems with the bw, could you give me a more concrete ideas of what you don't understand at this point?
Arthur
Arthur
thank you very much for your reply.
If you have time I would be very interested in understanding the basics (in the beginning) of Sphinx.
I read once again the document you have suggested me but I'm again a little bit confused. I made AN4 train and tried to live decode with Sphinx 4.
it do not worked very well but my English is not perfect so i don't mind for now cause i have a new problem in training. When I run the script 02.ci_schmm\slave_convg.pl I have the following error:
FATAL_ERROR: C:\new\SphinxTrain\src\libs\libio\corpus.c, line 256: input string too long. Truncated.
So I can't continue my training.
Now I have ten files (i know that are not enaugh even for few words, but i am interested in understanding the basics) in which I say the three words mixed for each file. I've searched in the forums but I didn't find solutions.
thank you a lot for your help.
Greetings,
Matteo Villari
Hi. Actually, your answer to Matteo helped me for now, so thank you. I will send anoter message to this forum, if I have any further questions.
Bye,
Matjaz