Google search by voice Daniel Bocho k (bochod1@fel.cvut.cz) - - PowerPoint PPT Presentation

IBM - CVUT Student Research Projects

Google search by voice

Daniel Bochořák (bochod1@fel.cvut.cz)

IBM - CVUT Student Research Projects 2

Introduction

• Google search by voice on Smartphone, PDA, …
• Speech recognizer:

– recognize one word(for example: hilton) – not so difficult – recognize the words in sentence (for example: hilton hotels) - difficult!

• Target:

– to improve the recognization of words in sentence (create data model) – use Google suggest

IBM - CVUT Student Research Projects 3

Google suggest

• For an input word returns the most frequently searched

sentences and the number of searching

• Remote Procedure Call (RPC) could be used
• http://labs.google.com/suggest
• For example.:

– input:

hilton

– output:

sentence number of searching

hilton hotels 16.700.000 results hilton honors 1.030.000 results hilton hotel 33.200.000 results

IBM - CVUT Student Research Projects 4

Data model

• Word (input), successors (Google suggest output)
• Probability of jump to successor

p = - log (a / b )

• For example:

– input:

hilton

– output

sentence probability

hilton hotels 0.744363 p = -log (16.700.000 / 92.700.000) = 0.744363 a – sucessor, number of searching b – input word, number of searching

IBM - CVUT Student Research Projects 5

Data model

• Input: Database of words
• Find the successors and their probability
• Creation of data model for speech

recognizer => better prediction of words in sentences