Hi,

trying to help..

does this project still needs readers for the french corpus ?

if yes, is that kind of french data usable ? : http://www.litteratureaudio.com/

for example :

audio : http://dl.free.fr/vf37Mz9L4

text: http://dl.free.fr/otmWOHbYr

la morte, guy de maupassant, female reader, 15 minute

We can always use more readers. If we have speech from many different speakers our acoustic models will become more speaker independent. This is good if an application should recognise many different speakers. So yes, more readers are always useful.

Any recording is only useful if it is either in the public domain, or available under the GPL (or a licence that would allow us to relicense it under the GPL).

The text of this recording is probably in the public domain (I assume that this particular text was written by Guy de Maupassant. If this is the case, then the copyright should be expired. Copyright normally expires 70 years after the death of the author.

That does not mean that the recording itself is in the public domain. The person who recorded it also has copyright on recording and only if she died more than 70 years ago, this recording will be copyright free (assuming that there is no background music or something like that).

So this recording is only usable for us if the person who recorded it licensed it under the GPL or a compatible license. In case of doubt it is always best practice to ask the speaker herself.

Furthermore there are technical issues. A recording should not be compressed in a lossy manner (so no MP3 for example). This recording seems to be compressed in a lossless manner (using flac) which is good. Of course the quality of the recording (good or bad microphone, lots of background noise, string sound artefacts etc) is also very important. We are mostly interested in relatively good quality recordings. With a bad quality microphone speech recognition will not work very well anyway (perhaps in the future, but not now).

Lastly, longer recordings have to be segmented into shorter pieces, because this makes an incredible difference in the time it takes to make an acoustic model. Ken has written a script that makes it possible to do this semiautomatically, but I guess it is still necessary that a native speaker actually segments the audio, so no mistakes are made.

Because of all this extra work and copyright issues, it is generally speaking more helpful and quicker to submit a couple of hours of your own speech through the speech submission app (or recording it line by line using e.g. audacity).

Most of this information is already available in the forums and in the FAQ, so if you have any related questions please check those sources first.

Hi Robin!

Two years ago, you submitted robin-20070821_de1.zip. I have built a speech model using these prompts (with my own voice, not your voice). Maybe you are interested to build a German speech model with your own German prompts? You would need to install the svn version of simon: $ svn co https://speech2text.svn.sourceforge.net/svnroot/speech2text/

Greetings,
Ralf