Re: [Trans-devel] Looking for new features in Transcriber 1.5.1

[Claude B. <Cla...@li...>]

Le 06/11/2014 17:25, Ivan Vanney a écrit :
> Hi dear, im new in this list. I work for a company which uses Transcriber 1.5.1 for transcription. 
> We use specific rules like for example, we dont transcribe overlapping but apply a [tag] instead, 
> but 100% of accuracy is impossible with human transcribers and sometimes a transcribers doesnt 
> apply the tag and instead transcribes.
> Is there a way the program may recognize when there are two noises or voices automatically?, we 
> would hire a developer to carry this out.
> Thank you,kindest regards.
>
Hi,

Transcriber wasn't designed for performing automatic audio processing, but instead to help human 
annotation (in order to develop automatic analysis systems using statistical machine learning) - 
however nothing prevents from using an automatic pre-processing and feed it as initial annotation.

Concerning overlapping speech or noises, it is an interesting but quite difficult problem, 
especially if you can't rely on multiple microphones for beamforming and source separation. It is 
far from a solved research problem even if some solution do exists - look eg. at 
http://bass-db.gforge.inria.fr/fasst/ . I doubt that currently an automatic detection would perform 
better than a human one, but it may help focusing on some sections: I am interested if ever you have 
a feedback on that!

Best regards,
   Claude Barras