On Sun, 24 Nov 2002 [EMAIL PROTECTED] wrote: > On 24 Nov, Michael Still wrote: > > This is interesting. Speech recognition is based on a large database of > > sample speech. For this database, the accent of the speaker matters a lot. > > > > If the Linux version is only trained with American voices, as it is only > > for sale in the states, then I would expect it it not work very well at > > all here. > > Actually, I believe it is on sale in the UK as well. And I know that > this version does require each person to `trial it' with their own voice > (which all good systems should do).
Most modern speech recognition is entirely probabilitic. If the model says that the word was "frog", then that equates to "frog" had the highest probability (it might still be a very low value). All you're doing during the training is tweaking the database to improve the probability of a match with your voice. However the records aren't of your voice. Then again, is they're old, they probably have those dinky British accents. > Basically, for the price, I'm willing to take a gamble. Fair enough. Let me know how you go. Mikal -- Michael Still ([EMAIL PROTECTED]) UTC +11 hours -- SLUG - Sydney Linux User's Group - http://slug.org.au/ More Info: http://lists.slug.org.au/listinfo/slug
