40 likes | 53 Views
This tutorial explores the considerations for designing speech recognition systems for different applications such as phone name dialing, taxi booking, pizza ordering, email dictation, and web navigation. It also provides a scenario-based analysis for optimizing outcomes based on speaker dependency, vocabulary size, and language models. Additionally, it includes an exercise on using Bayes' rule and a context-free grammar for generating an infinite language pattern.
E N D
Tutorial 2 September 8, 2005
If you were a system designer what kind of speech recognition system would you use for the following? Consider whether you would use • Speaker dependent or independent • Large or limited vocabulary • Language grammar or statistical language model and carefully consider what would lead to the best outcomes. • Phone name dialing • Taxi booking • Pizza ordering • E-mail dictation • Web navigation (go back, visit .. )
Suppose you are given a speech recognizer for a language consisting of only four words: tire, dire, tile, dial. With the following channel model for phone realizations. • P(ay -> ay) = 1.0 • P(t->t) = 0.8p(t->d) = 0.2 • P(d->d) = 0.9p(d->t) = 0.1 • P(r->r) = 0.6p(r->l) = 0.4 • P(l->l) = 0.6p(l->r)=0.4 Using Bayes rule and the following unigram language model, calculate the most probable true input for each of the outputs :/tayl/ and /dayl/ p(tire)=0.4p(dire)=0.2p(dial)=0.3p(tile)=0.1
3. Consider an infinite language that only accepts sentences of the form : the mouse died the cat died the dog died . . . the mouse that the cat bit died the cat that the mouse bit died the dog that the mouse bit died . . . the mouse that the cat that the dog chased bit died the cat that the mouse that the dog chased bit died the dog that the cat that the mouse chased bit died . . . etc any noun can substitute for mouse, cat or dog any verb could substitute for chased, bit or died. Write a context free grammar that accepts exactly this infinite language