Hi Ahel,

Glad to hear that you are interested in helping out with open source speech recognition!

My replies follow:

On Tue, Apr 5, 2011 at 5:50 PM, Ahel ibn Alquivr <[email protected]> wrote:

Hi!
I'm attending to apply at this idea related to Voxforge: briefly Peter (my mentor) has thought that everyone who install simon, train for himself the vocabulary recognition and so on; so it would be cool if simon would envoy semi-automatically the trained audio file that Voxforge is collecting.

When you say "envoy" do you mean send or upload to the VoxForge collection site? 

 

May I know what do you think about this idea?

I think it is a great idea, especially for the languages for which we have fewer submissions (i.e. non-English languages).  For English, it would be an excellent way to obtain more speech, and to improve the quality of submissions, since, as Peter mentions in one of the comment in the page you linked, Simon can help with filtering out poor recordings.

Some things to think about:

How much speech do we need?

Peter also mentions in the GSOC project description that we need lots more speech to create acoustic models for dictation applications (unfortunately, I have helped spread that misconception...).  Though more speech generally is better, there are diminishing returns after a point - and it is not clear where that point is.  See the following links:

• Nickolay (CMU Sphinx maintainer) post in this thread: VoxForge Acoustic Models w/ Sphinx 4):
• Nickolay's blog post here: How to create a speech recognition application for your needs.
• Arthur Chan (former CMU Sphinx maintainer) paper here: Do we have a true open source dictation machine?: 
• The CMU SPhinx site's speech corpus size estimates

I have guessed 140 hours of speech is required for a good command and control acoustic model - simply because that is what the main CMU Sphinx acoustic model uses...  

Not sure how much is required for dictation, though one would "assume" more would be required.  However, as Nickolay clearly argues in his posts, and the papers he links to, more is not necessarily better. 

Types of speech

There are many types of speech that can be collected - see this FAQ entry: What is a speech corpus or speech corpora?  Command and control can use acoustic models with 'read' speech, and work quite well.  

Dictation acoustic models *may* need to be trained with 'spontaneous speech' rather than 'read' speech (or a mix thereof) to be effective.  See this thread: how to get more voice samples?

Conclusion

I can see your application being very good at collecting lots of 'read' speech for the applications that people actually use command and control speech recognition for...  That in and of itself would be very valuable to the open source community.  

 

Also I'd like to discuss with you if and how ways would be better to implement, in your opinion, the technology in support of this idea.

The VoxForge speech submission applet uses Java-based Postlet code for its client uploader, and the php server based code described on the postlet site.

That should give you a pretty good idea of what is needed on the upload side. Please make sure the URL for site for uploading is configurable by the user - so that if we change providers, it is just a simple change in URL.

Do you have any other approaches that might work? 

How much speech do we need?

Peter also mentions in the GSOC project description that we need lots more speech to create acoustic models for dictation applications (unfortunately, I have helped spread that misconception...).  Though more speech generally is better, there are diminishing returns after a point - and it is not clear where that point is.

Also I'd like to discuss with you if and how ways would be better to implement, in your opinion, the technology in support of this idea.

The VoxForge speech submission applet uses Java-based Postlet code for its client uploader, and the php server based code described on the postlet site.

That should give you a pretty good idea of what is needed on the upload side. Please make sure the URL for site for uploading is configurable by the user - so that if we change providers, it is just a simple change in URL.

 

We already use it for our sample acquisition ourselves and having sscd and simon compatible would be a huge plus for us.
Would you consider setting up an sscd server for Voxforge?

depends - can it run on a regular web hoster, or does it need a specialized server instance?

The VoxForge front end (WebGUI CMS) is on a server in my basement, and the VoxForge speech submission applet uploads to php code on a 1&1 web hosting account.  This was done on the assumption that my home internet connection (10 mbit down/ 1 mbit up) was not fast enough to accommodate speech uploads (along with current web traffic).  If the ssc could throttle itself a bit, then maybe it could be installed on the VoxForge server. 

Well it has a server component so you need to have _some_ access. It doesn't really need root access but if it doesn't have one it can't bind to any of the lower ports (OS restriction).

I assume you mean megabit, right?

Well you can set up QoS to limit the impact of incoming traffic on your web traffic. It really shouldn't be noticeable.

 So you alone could accommodate 10 other people having the same internet connection you have (10 x 1mbit up = your 10mbit down).

Maybe you could have a look at ssc/sscd? Both have a manual but you can also ask me if you have any difficulties...

Well it has a server component so you need to have _some_ access. It doesn't really need root access but if it doesn't have one it can't bind to any of the lower ports (OS restriction).

my 1&1 account does not allow access to the lower ports.

 So you alone could accommodate 10 other people having the same internet connection you have (10 x 1mbit up = your 10mbit down).

I agree, in theory - but doesn't seem to work that way in practice for some reason...

Maybe you could have a look at ssc/sscd? Both have a manual but you can also ask me if you have any difficulties...

Will do, after April 23, I will have much more free time.

Noticed that all audio is stored in MySql (rather than pointers to static files), is there a reason for this?  What kind of security does sscd use - has it been used on the Internet or mostly in "behind the firewall" LAN configurations?