The general rule is that the GPL only covers 'distribution' of a program or work. 

So if you try to distribute modifications to the original work, you must provide users with your changes.  Or if you distribute the original work in object or executable form, you must provide the source.  If you never distribute the original GPL'ed work, GPL does not apply.

The GPL in section 2 says:

[...] If identifiable sections of that work are not derived from the Program, and can be reasonably considered independent and separate works in themselves, then this License, and its terms, do not apply to those sections when you distribute them as separate works. [...]

Thus, it is not the intent of this section to claim rights or contest your rights to work written entirely by you; rather, the intent is to exercise the right to control the distribution of derivative or collective works based on the Program.

(my emphasis added) 

Therefore, my reading of the GPL is that if someone develops a hardware (or proprietary software-based) speech recognition engine that uses VoxForge Acoustic Models ('AM'), GPL would not prevent him from telling his users to link to the VoxForge site and get their AM there.  I am not a lawyer, and I think that this question is important enough to put to the FSF to confirm my interpretation.

Having said that, the reality of the situation is that Acoustic Models, like software, are never static.  There is always room for improvement.  A commercial venture would most likely need to make improvements to transcriptions, clean up the audio, add their own domain specific audio to address specific customer needs.  Acoustic Models would have to be localized to improve their recognition rates for specific regions.  And these modifications, if 'distributed' to their customers, would need to be published pursuant to the GPL. 

That is how I can see that the community would win in this situation.

Ken

--- (Edited on 12/17/2006 10:21 pm [GMT-0500] by kmaclean) ---

Autoreply from SFS:

This message has been automatically generated in response to the a
licensing question you sent to the Free Software Foundation, with subject:
       "VoxForge Licensing Question".

There is no need to reply to this message right now.  Your request has
been assigned an ID of [gnu.org #324354].

Please include the string:
        [gnu.org #324354]
in the subject line of all future correspondence about this issue.  To do
so, you may reply to this message.


This address is answered primarily by volunteers, overseen by one staff
member of FSF.  We have very limited resources to answers requests. 

--- (Edited on 12/17/2006 10:32 pm [GMT-0500] by kmaclean) ---

Can you provide a link where everybody can follow this issue at FSF?

Or will you post FSF's clarification as a follow-up to this thread when it comes in? 

--- (Edited on 12/22/2006 10:06 am [GMT-0600] by Visitor) ---

I just sent an email to FSF and I posted the auto-reply e-mail they sent back.  I am not sure how this process works at FSF ... if there is a link I will post it, if not I will post their e-mail reply here.

Ken 

 

--- (Edited on 12/22/2006 12:37 pm [GMT-0500] by kmaclean) ---

Yoni Rabkin via RT <[email protected]>	Fri, Feb 23, 2007 at 2:31 AM
Reply-To: [email protected] To: [email protected]
Hello, Please accept our apologies for the delay in getting back to you. We rely on volunteer effort and often have difficulties keeping up, and thank you for your efforts. > I am the admin of the VoxForge web site (www.voxforge.org). We are > collecting GPL speech audio to be used in the creation of Acoustic Models > for use with Free and Open Source Speech Recognition Engines. I had a user > ask a question<http://www.voxforge.org/home/forums/message-boards/general-discussion/gpl-applicable-to-derived-hardware>(see > below) and just wanted to make sure my reply (see below) made sense > from a GPL perspective. Any comments would be greatly appreciated. > thanks, > Ken > -- > http://www.voxforge.org > User Question: > I wonder, how the GPL would be applied to the following situation: > A huge number of users like you and me collect a speech corpus that allows > to build acoustic models of similar quality compared to commercial systems. > Now a hardware developer comes up with some speech recognition tool that is > designed to work exactly with VoxForge's model. He doesn't distribute the > model with his device, but provides a link to voxforge.org where customers > may download the model. > - Is this situation covered by GPL? > If not, I'm reluctant to spend time on something that others may turn into > profit without returning anything to the community. > If yes: > - Does the GPL require the device developer to distribute his device > together with the model sources, OR > - Does the GPL reuqire him to distribute his device together with the model > sources PLUS any sources and hardware description of his own device? > The latter situation is most desirable, because everybody who benefits from > the acoustic models has to provide some improvement in return. But I doubt > GPL includes anything except for directly derived code (or models). Your > opinion? > --- (Edited on 12/17/2006 4:10 pm [GMT-0600] by Visitor) --- [snip] Please excuse me for snipping away your answer, but I feel that it would be simpler for me to explain the matter directly than via an interpretation of your answer. Note that I have only a layman's knowledge of speech recognition and would be happy for corrections. Finally, my answer concerns software only, not hardware. Here is how I see it: (i) An acoustic-model is the object code form of the speech corpus. The speech corpus being the GPL'd source code. Distribution of the acoustic-model is therefore governed by section 3 of the GPL. (ii) A work which relies on the GPL'd acoustic-model in order to perform is its primary functionality is a derivative of the acoustic-model. As such, the software using the acoustic-model needs to be licensed under the GNU GPL as well. I believe this follows from section 0 of the GPL: "The act of running the Program is not restricted, and the output from the Program is covered only if its contents constitute a work based on the Program (independent of having been made by running the Program). Whether that is true depends on what the Program does." I am treating the acoustic-model as a GPL'd library being linked to (and enables the software to do its primary task), an act which brings the entire work under the GPL. This is as opposed to, for example, playing a GPL'd Ogg file in a proprietary audio player, an act which does not violate the license of the Ogg file. I need your help to determine if the above paragraph is actually true, since this is entirely dependant on how acoustic-models are used by software. Finally, the matter of what constitutes a derivative of a work can only be ultimately decided by a court of law, and is a subtle matter. I hope my belated answer is of help. -- I am not a lawyer, the above is not legal advice   Regards, Yoni Rabkin

--- (Edited on 2/23/2007 11:22 am [GMT-0500] by kmaclean) ---

Hi Yoni,

thanks for getting back to me on this - I know your group is very busy and I appreciate your explanation greatly.  My comments follow:

>I am treating the acoustic-model as a GPL'd library being linked to (and
>enables the software to do its primary task), an act which brings the
>entire work under the GPL. This is as opposed to, for example, playing a
>GPL'd Ogg file in a proprietary audio player, an act which does not
>violate the license of the Ogg file.

>I need your help to determine if the above paragraph is actually true,
>since this is entirely dependant on how acoustic-models are used by
>software.

Basically, without an Acoustic Model a Speech Recognition Engine will not work. 

An Acoustic Model contains the statistical representations of the sounds that make up the words in a pronunciation dictionary (which can contain over a hundred thousand words and their associated pronunciations). These statistical representations are obtained by analyzing a great deal of speech audio (hundreds of hours) and their corresponding text transcriptions through a process called 'Acoustic Model Training'.  When trying to recognize someone's speech, the Speech Recognition Engine ("SRE") takes the speech audio to be recognized, breaks it down to individual sounds, and 'searches' the Acoustic Model for the same sounds.  When a sequence of sounds corresponding to a word or sentence is identified, the SRE returns the word or sentence to the calling application (this is an oversimplification of the process, but gives you an idea of how it is done). 

So an Acoustic Model is much more integral to the functioning of a Speech Recognition Engine that an Ogg file would be for a proprietary audio player.  Your analysis makes more sense than my previous stab at it.

Therefore any Speech Recognition Engine that uses the VoxForge Acoustic Model would have to operate under the GPL license or an FSF approved Open Source license that can use  GPL libraries.

thanks for your help,

Ken

I'm not a lawyer either, and the above is not legal advice

--- (Edited on 2/27/2007 11:24 am [GMT-0500] by kmaclean) ---

Hi Yoni,

More thoughts about licensing and the VoxForge Acoustic Models:

I think there might be a distinction to be made between a 'Software Library' and an 'Acoustic Model'. 

I may be wrong, but my understanding is that a Library performs some type of action that is required by the calling programs.  It forms an integral part of the software.  It can be statically or dynamically linked to the calling software.

An Acoustic Model is more like data (i.e. statistical representations of sounds) required by the calling software.  There is no requirement for linking in the traditional sense (either statically at compilation time or dynamically at run-time).  An Acoustic Model does not perform any action for the Speech Recognition Engine software.

So although an Acoustic Model is much more integral to the functioning of a Speech Recognition Engine that an Ogg file would be for a proprietary audio player, it is less integral than a Library is to a software program.

Ken

I'm not a lawyer, and the above is not legal advice

--- (Edited on 2/27/2007 11:25 am [GMT-0500] by kmaclean) ---

Yoni Rabkin via RT <[email protected]>	Tue, Feb 27, 2007 at 10:44 AM
Reply-To: [email protected] To: [email protected]

Hi,

I'm still convinced that the acoustic model is a library as far as the
GPL is concerned. Allow me to recapitulate:

To the best of my understanding, the primary purpose of a speech
recognition engine is not to read the acoustic model as data. Instead,
the speech recognition engine uses the acoustic model in order to
achieve its primary purpose. Moreover, I think that the output of the
speech recognition engine is derived from the acoustic model.

If the acoustic model was "just data", switching it should not have a
profound effect on the speech recognition engine's primary
functionality.

To contrast, an audio player plays any media file just as well. Just as
a compiler would compile any program just as well. If an audio player
needed *at least* a single instance of some data X in order to perform
its primary function (playing music), I would say that the data X is as
essential as any of the other libraries the audio player requires.

I'll stress again that without legal precedence, we can continue to
argue and convince each other, but ultimately only a court can decide
such a subtle matter.

Since the speech corpus is licensed under the GPL, I am taking the
stance which I view has the best chance of protecting the wishes of the
copyright holder/s. If the speech corpus was licensed under the LGPL, I
would probably have chosen a different route.

I'm just a volunteer, and this is uncharted territory, so this is the
best I can do. But I look at it like this: once upon a time, long ago,
licensing almost *any* software under the GNU GPL was uncharted
territory, and things seem to have been working out well.

--- (Edited on 2/27/2007 11:26 am [GMT-0500] by kmaclean) ---

--- (Edited on 2/27/2007 11:28 am [GMT-0500] by kmaclean) ---

Yoni,

let me comment on two paragraphs of your preivous post.

>To the best of my understanding, the primary purpose of a speech
>recognition engine is not to read the acoustic model as data. Instead,
>the speech recognition engine uses the acoustic model in order to
>achieve its primary purpose.

 Yes, but there is not "the" acoustic model (like "the" library version against which a program has to be compiled)

>If the acoustic model was "just data", switching it should not have a
>profound effect on the speech recognition engine's primary
>functionality.

There can be several similar models with no profound difference in recognition rate. A user might even be able to train his own acoustic model with better (single user) performance than the voxforge model.

 Given these comments, does it still make sense  to say, that the voxforge acoustic model is like a library to some software that allows to use this model, but can instead work with other acoustic models as well?

--- (Edited on 4/22/2007 6:21 pm [GMT-0500] by Visitor) ---