aug29_applications_share

aug29_applications_share - Animals in the News Next...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Animals in the News Next Wednesday, 05 September 2007 43rd anniversary #1 spot on Billboard Singles Chart INDV101 `Language' August 29, 2007 Applications of Linguistics House of the Rising Sun INDV101 Fall 2007 Room change Section 19 8:00am meeting time TA: Jaime Parchment NEW ROOM! Chavez 316 Logistics Homework 1 due Friday, Sept 7 Worksheet D2L quiz, covers Winkler chs 1, 4 & Course docs Not turnitin.com Next week No class Monday (holiday) Begin Phonetics/Phonology on Weds. INDV101 Fall 2007 INDV101 Fall 2007 1 Review What is a `language'? (first pass) language' Linguistic varieties Mutual intelligibility Language vs dialect Geographic dialect, sociolect Idiolect Speech style or register INDV101 Fall 2007 By the end of this lecture You should have some idea of what these are speech recognition speech synthesis machine translation INDV101 Fall 2007 Points to Consider Our understanding of how humans perform language tasks is limited Our ability to make computers perform language tasks is limited in many of the same ways Points to Consider Sometimes we try to make computers perform language tasks like humans do Sometimes we abandon those efforts and take an entirely different approach INDV101 Fall 2007 INDV101 Fall 2007 2 Points to Consider One important non-commercial aspect nonof computational linguistics is testing and influencing linguistic theory Theory informs practice Practice tests theory Applications of Linguistics Humorous Anecdote 1.0 INDV101 Fall 2007 INDV101 Fall 2007 Speech Recognition You talk, computer writes what you say Commercial software Training Reading word lists Why Proofread? Error rate Best are around 93% accurate INDV101 Fall 2007 INDV101 Fall 2007 3 Why Proofread? Why Proofread? INDV101 Fall 2007 INDV101 Fall 2007 Automatic Speech Recognition "Speech to Text" processor Text" Input Speech Sounds Output Text or Action Automatic Speech Recognition PROBLEM: Different people sound different Lack of Invariance Problem There is no single invariant feature that allows us to recognize different voices INDV101 Fall 2007 INDV101 Fall 2007 4 Lack of Invariance Question How do people deal with it? Frightfully Brief and Woefully Incomplete Introduction to Acoustic Analysis Sound = pressure waves Waveform 7 Answer We don't have a complete understanding don' of how humans do it. 0 3 0 T im e ( s ) 1 . 4 8 INDV101 Fall 2007 INDV101 Fall 2007 Comparing sounds Comparing sounds INDV101 Fall 2007 INDV101 Fall 2007 5 Comparing sounds Single Speaker Recognition Dr. Martinez is just one person Why the confusion? INDV101 Fall 2007 INDV101 Fall 2007 Compare these Waveforms Compare these Waveforms INDV101 Fall 2007 INDV101 Fall 2007 6 Dealing With New Voices Speaker normalization The process of "adjusting" your perception to adjusting" accommodate a new speaker One Strategy for ASR Limit functionality to just one speaker Train on just that speaker Record word lists to learn what sound combinations look like for that speaker INDV101 Fall 2007 INDV101 Fall 2007 Phonotactics Rules of allowable sound sequences in a given language English does not allow /ps/ at the /ps/ beginning of words Greek does English has flaps only between vowels Spanish uses flaps more freely INDV101 Fall 2007 Word Lists for Training /i/ /p/ /t/ /g/ /pik/ pik/ peek /tin/ teen /gik/ gik/ geek /e/ /pet/ pet /ten/ ten /ges/ ges/ guess /l/ /plm/ plum /tls/ atlas /glv/ glove INDV101 Fall 2007 7 Guessing Game Conditional probability Given my best guesses for the first sound, how likely are my best guesses for the second sound? Guessing Game m n b a o d j t p i I n m e I s f No flaps before consonants in English only between vowels INDV101 Fall 2007 INDV101 Fall 2007 Guessing Game m n b m a o a d j j t p p i I i n m n e I I s f s Dealing with Many Speakers Strategies for making the problem easier? Limit the lexicon Martinez my penis INDV101 Fall 2007 INDV101 Fall 2007 8 Examples of Limited Lexicons Systems that expect numbers Everything is interpreted as a number Expectation of cooperation Examples of Limited Lexicons Systems that expect city names Interesting training case: shouting INDV101 Fall 2007 INDV101 Fall 2007 Speech Recognition Relies on Context Acoustic context adjacent sounds Speech Synthesis Creating understandable speech Is this problem easier or harder than Speech Recognition? Lexical context legitimate word? Pragmatic context make sense? INDV101 Fall 2007 INDV101 Fall 2007 9 Speech Synthesis Creating understandable speech Is this problem easier or harder than Speech Recognition? I think it's easier. it' We understand bad speech better than machines understand clear speech. Speech Synthesis Examples 1980s Current INDV101 Fall 2007 INDV101 Fall 2007 Speech Synthesis Difficult aspects of speech synthesis parallel difficulties in speech recognition. The natural speech signal is extremely complex Hard to decipher Hard to reproduce Machine Translation Input: Text in one language Output: Text in another language Easier or harder? Hard comparison to make No sound to worry about, but meaning is critical INDV101 Fall 2007 INDV101 Fall 2007 10 Machine Translation 1960 "Ten years away" away" 1970 "Ten years away" away" Current Computer Aided Translation Machine Translation Rule-based translation Rule "Logico-Semantic" approach Logico- Semantic" Corpus-based translation Corpus- INDV101 Fall 2007 INDV101 Fall 2007 Computational Modeling Doesn't produce anything marketable Doesn' Intended to test theories Doesn't directly say what a human brain Doesn' can do Does place limits on the problem, particularly questions of learnability Summary Efforts to make computers "do language" better parallel language" our understanding of how humans "do language". language" Sometimes it's best to mimic a system (human brain) that it' works (rule-based and cognitive approaches). (rule Sometimes it's best to take advantage of the strengths of it' the machine (statistical and corpus-based approaches). corpus Efforts to directly model human language use test our theories of how language works, particularly boundaries of learnability. learnability. INDV101 Fall 2007 INDV101 Fall 2007 11 INDV101 Fall 2007 12 ...
View Full Document

This note was uploaded on 02/25/2010 for the course INDV 101 taught by Professor Walker during the Fall '07 term at Arizona.

Ask a homework question - tutors are online