VoxForge
VoxForge
Hi ubanov,
thanks,
Added to Trac as ticket #435.
The format of the prompts looks good. It would be better to have more than 36 prompts (around 100-200 to start is good to get some variety).
I'm a little confused by your phoneme usage statistics... The number of prompts from the training column show 1330, which seems about right for the prompts in the text file, but I don't understand where the NumPhon for the Spanish column came from.
thanks,
Ken
hi kmaclean,
If 36 phrases are not enought, I will try to write more. I'm going to try to send they, this night.
Last week I take an spanish dictionary (containing almost all the words that spanish language have). With this file I run the program that I wrote (festival2phone) in order to obtain all the phonemes of spanish, and an initial statistic about the prob. of appearing each phoneme (I know, that some words are used more than others, and then this statistical is partial, but it's a first point to have some data). The data in the "NumPhon for the Spanish column" is this the global prob of appearing each phoneme in spanish (in order to compare with the number of times that appear in the training text).
I don´t use any program in order to take statistics about the appearing of each phoneme (I use the festival2phone program, some unix commands, like sort, uniq, awk, and staroffice).
Those files are in the last msg that I have sended in last days, but if you want I could resend again.
Regards,
Ivan
Hi again,
Another thing, thinking about voice donor process, I have one idea... let's see if you think it could be viable.
In english there are some programs that do voice recognition, but in spanish there is NO program that can do anything. (This happens in a lot of languages).
If a possible spanish voice donor, see the number of voice records that you have, it seems that it's very dificult to reach the primer 160 hours objetive.
Then a person, like me, could forgot the donor, or he could think about preparing training data for sphinx2 or sphinx3 or julian, in order to have basic functions of recognition for his own voice... the process of creating this it's not easy.
Could it be possible to have a application in voxforge that help the people in all this training process. Let's say that in the initial process you have the read applet. Then a person could record his voice (100 phrases could be about 10 minutes of voice), and voxforge will retain all this voice under the GNU license. Then let's say that Voxforge processes the voice of one person and prepare with HTK, the HMM files that this person could use it in order to use any of the voice recognition engines (the selected by voxforge). This could be a little gift that voxforge gives to a voice donor... and it could help convincing a person to donor his voice.
If this process is something that could be automatically done, it could be possible to write some messages in some linux "freak" forums, explaining the easy method of having voice recognition for their desktop...
What do you thing about this?!? could it be a way in order to persuade to donor his voice.
Regards.
Hi Ubanov,
>... it seems that it's very dificult to reach the
primer 160 hours objetive.
It's actually 140 hours, but still a high number...(Note that this is only for a command and control system).
> Then a person could record his voice (100 phrases could be about 10
>minutes of voice) [...] Then let's say that Voxforge processes the voice of
>one person and prepare with HTK, the HMM files that this person could
>use it in order to use any of the voice recognition engines (the selected
>by voxforge). This could be a little gift that voxforge gives to a voice
>donor... and it could help convincing a person to donor his voice.
Interesting, though I am not sure I understand the advantage of creating speaker dependent acoustic models with only a single person's speech... You would need more speech than 10 minutes to create a good user-specific acoustic model.
Allowing people to "adapt" the generic VoxForge acoustic model to their voice would probably make more sense...
In addition, since people are continuously adding speech to the English VoxForge corpus, and acoustic models are re-trained every night, users are getting acoustic models that are trained with their voice... whenever I get around to validating and submitting their submission to the Subversion Speech Corpus repository.
This validation of submitted audio is the current bottleneck that I am working to resolve... We need a web-enabled way to allow users to validate the audio (and correct the prompts, and readme entries...) in submissions and mark them for inclusion (or exclusion) in the nightly build of the acoustic model.
thanks,
Ken
P.S. please start new threads on new topics - because it gets difficult to follow things with a bunch of different topics being discussed in the same thread. In addition, other people, who might want to contribute to this discussion, might not be following this particular thread. :)
Hi again,
I thing that will be better to adapt the generic VoxForge acoustic model to their voice, but today there is nothing to adapt, because we don't have enough voice. Then the only way to make a simple voice recognition program in spanish is to make a speaker dependent acoustic model, isn't it?
I don't know how many hours are needed to make a simple command recognition program...
In order to the people to use the recognition, it's needed the existence of some voice aplication that could control a linux desktop... make a simple program could be easy (a program that handle a little grammar: open program, close program, and so on).
May be I'm going to copy my previous message in other thread, in order to see if any possible user give their opinion about this... if it could be interesting let's work on it... if not... Please, copy your previous reply, and then I will copy this one if you want.
Regards.
