General Discussion

Nested
Whole-Word Recognition
User: Nick
Date: 4/29/2009 9:43 am
Views: 9418
Rating: 2

Hi,

I'm builing a whole-word recognition system (also referred to as isolated word recognition) and I'm not sure about the form of my training wav files:

I need to know, if I need wav snippets of each single word which I want to train (with silence before and after the word -> labels: sil "word" sil) or if it's also possible to use a sentence of words which I want to use (i.e. the labeling is for example of the form: sil "word" sp "word" sil).

This question is also interesting if someone wants to use the VoxFroge Speech Data for whole-word recognitioning!

I hope someone can help me with my question!

--- (Edited on 4/29/2009 9:43 am [GMT-0500] by Visitor) ---

Re: Whole-Word Recognition
User: tpavelka
Date: 4/30/2009 2:32 am
Views: 5441
Rating: 4

Hi Nick,

> I need to know, if I need wav snippets of each single word which

> I want to train (with silence before and after the word -> labels:

> sil "word" sil) or if it's also possible to use a sentence of words

> which I want to use (i.e. the labeling is for example of the form:

> sil "word" sp "word" sil).

You can do both, the question is whether to use whole word HMMs or phonetic unit (phoneme) HMMs. Whole word HMM are usefull if you want to do IWR and do not have enough training data. Otherwise, just use phoneme HMM (i.e. the standard described in all tutorials).

What is the language that you do recognition in? If it is English just use the official VoxForge acoustic model + HVite or Julius.

E.g. for HVite you just need a grammar and a dictionary and you can do IWR or grammar based recognition.

 

--- (Edited on 30.04.2009 09:32 [GMT+0200] by tpavelka) ---

PreviousNext