It is very important that your speaker and you understand the legal
status of the recorded database. It is very wise that the speaker
signs a statement before you start recording or at least talk to them
ensuring they understand what you want to do with the data and what
restrictions if any they require. Remember in recording their
voice you are potentially allowing anyone (who gets access to the
database) to fake that person's voice. The whole issue of
building a synthetic voice from recordings is still actually an
uninvestigated part of copyright but there are clear ways to ensure you
wont be caught out by a law suit, or a disgruntled subject later.
Explain what you going to do with the database. Get the speaker
to agree to the level use you may make of the recordings (and any
use of them). This will roughly be:
You must find out what the speaker agrees to before you start spending
your time recording. There is nothing worse than spending weeks
on building a good voice only to discover that you don't have rights
to do anything with it.
Also, don't lie to the speaker make it clear, what it means if their
voice is to be released for free. If you release the voice on the net
(as we do with our voices), anyone may use it. It could be used
anywhere, from reading porn stories to emergency broadcast systems.
Also note that effectively building a voice from a synthesizer means
that the person will no longer be able to use voice id systems as a
password protection (actually that depends on the type of voice id
system). However also reassure them that these extremes are very
unlikely and actually they will be contributing to world of speech
science and people will use their voice because they like it.
We (KAL and AWB) have already given up the idea that our voices are in
anyway ours and have recorded databases and made them public (even
though AWB has a funny accent). When recording others we ensure they
understand the consequences and get them to explicitly sign a license that
gives us (and/or our institution) the rights to do anything they wish,
but the intention is the voice will be released for free without
restriction. From our point of view, having no restrictions is by far
the easiest. We also give (non-exclusive) commercial rights to the
voice to the speaker themselves. This actually costs us nothing, and
given most of our recorded voices are for free the speaker could
re-release the free version and use it commercially (as can anyone
else) but its nice that the original license allows the speaker direct
commercial rights (none that I know of have actually done anything
with those rights).
There may be other factors though. Someone else may be paying for the
database so they need to be accommodated on any such license. Also a
database may already be recorded under some license and you wish to use
it to build a synthetic voice, make sure you have the rights to do this.
Its amazing how mainly people record speech databases and don't take
into account the fact that someone else may build a general TTS systems
from their voice. Its better that you check that have to deal with
problems later.
An example of the license we use at CMU is given in the festvox
distribution festvox/src/vox_files/speaker.licence.
Also note that there are legal aspects to other parts of a synthetic
voice the builder must also ensure they have rights to. Lexicons may
have various restrictions. The Oxford Advanced Learners' Dictionary
that we currently use for UK English voices is free for non-commercial
use only, thus effectively imposing the same restriction on the complete
voice even though the prosodic models and diphone databases are free.
Also be careful you check the rights when building models from existing
data. Some databases are free for research only and even data derived
from them (e.g. duration models) may not be further distributed. Check
at the start, question all pieces of the system to make sure you know
who owns what and what restrictions they impose. This process is worth
doing at the start of a project so things are always clear.
