Click here to register.

Edit Message

Visitor Name
Subject
Message

Re: participating in project

Hi Dan,

>If voice files specific to medicine are required to create data sets specific

>for medical dictation

The main non-software components used by speech recognition engines (such as Sphinx or Julius) are: a language model and an acoustic model

For dictation, in addition to the speech recognition engine, another software component would be required: a way continuously improve the acoustic model by updating the standard acoustic model with speech from the user - a process called 'adaption' (Nickolay talks about the importance to implementing adaptation for dictation here)

We would only be  addressing the speech audio requirements for creating acoustic models if we were to collect medical speech. 

Depending on the sentences you provide, these might be used for language model creation too. 

I am not sure if there are open source dictation systems that implement user adaptation of acoustic models: the EvalDictator dictation dialog manager might do this...

>could we set up a way in which individual physicians could donate their

>own speech and text on voxforge

Yes, we could start with a list of common medical words and phrases in a separate 'medical' section (similar to how each language has its own 'read' page).  We would target this to physicians, since they would presumably know how to correctly pronounce this specialized vocabulary.

Once we get a few hours of speech, we could start on a pronunciation dictionary (I would need your help for that) so that people could generate their own acoustic models based on the submitted medical speech.

Ken