EE 516 Lecture 1 Geoffrey Zweig Microsoft Research

EE 516 Lecture 1 Geoffrey Zweig Microsoft Research 4/2/2009

Our Topics From JHU 2002 Super. SID Final Presentation – Reynolds et al. Introducing today!

Topic Coverage By Day • Data Representations and Models (4/23) – Vector Quantization – Gaussian Mixtures – The EM Algorithm • Speaker Identification (5/7) • Language Identification (5/7) • Hidden Markov Models (5/14) – Dynamic Programming • Building a Speech Recognizer (5/14)

Language Identification – Why Do it? • Multi-lingual society – Applications should be able to deal with anyone • Businesses – Automated help systems – Reservations, account access, etc. – Travel • Airport Kiosks • Train stations • Government – Funds research to identify languages – Runs evaluations in it

How Do You Do it? English Acoustic Model French Acoustic Model Output Likeliest … Tamil Acoustic Model Gaussian Mixture Models - 4/23

How Do You Do It? (2) “p ih n s” – probably English… “k r p s t” – probably Czech… Simple HMMs – 5/14 After Zissman 1996 Language Models – 4/30

How Do You Do It (3) Same methods multiple times Acero et al. , Chapter 4 4/23 After Zissman 1996

How Do You Do It? (4) Run a complete speech recognizer in each language And we will see several other ways, and combinations! After Zissman 1996

Gauging Progress – The NIST Evaluations • National Institute of Standards and Technology • Has sponsored benchmark tests in multiple language processing areas for over a decade – – – – Topic Detection & Tracking Content Extraction Video Analysis Speech Recognition Language Identification Speaker Identification Machine Translation http: //www. itl. nist. gov/iad/mig/tests/ • Coordination with site funding by Defense Advanced Research Projects Agency (DARPA) • Along with business interest, the driving force in advancing the State-of-the-Art

For Example, Progress in Speech Recognition

Language Identification - How Well Can It Be Done – Who Salutes? Organization Location Beijing Naphoo Technology Company+ Brno University of Technology Georgia Institute of Technology Groupe des Ecoles des Telecommunication, Ecole Nationale Superieure des Telecommunications IBM IKERLAN Technological Research Center Institut de Recherche en Informatique de Toulouse Institute for Infocomm Research Institute of Acoustics, Chinese Academy of Sciences+ China Czech Republic USA Institut National de Recherche sur les Transports et Leur Securite France International Computer Science Institute (USA) Laboratoire d'Informatique pour la Mecanique et les Sciences de l'Ingenieur MIT Lincoln Laboratory Nanyang Technological University Politecnico di Torino Spescom Datavoice Telefonica I & D TNO Human Factors Tsinghua University Universidad Autnoma de Madrid University of the Basque Country University of Stellenbosch University of Science and Technology of China+ USA From NIST 2007 LRE Website France USA Spain France Singapore China France USA Singapore Italy South Africa Spain The Netherlands China Spain South Africa China

How Well Can it Be Done – What Languages? From NIST 2007 LRE Website

How Well Can It Be Done? – Testing Conditions • 26 languages and dialects • Telephone speech • Multiple duration conditions – 3, 10, 30 seconds • Detection Error Tradeoff (DET) Curves used to measure performance

How Well Can it Be Done – Some Numbers From NIST 2007 LRE Website

Language Identification Project • Build a language ID system with the Call Friend Data set • Implement several of the main techniques • Set up a demo on your laptop that will recognize someone’s language

Flavors of Speaker Recognition Our Focus! From JHU 2002 Super. SID Final Presentation – Reynolds et al.

Speaker Recognition – Why Do It? • Personal Applications – Voice-print passwords – Voicemail transcription – who left that message? • Business Applications – Calling your bank • Government – Is that Osama calling from Pakistan? – Prison call monitoring – Automated parolee calling – is he where you think?

How Do You Do It? The most basic approach: Gaussian Mixture Models - 4/23 More recently: Support vector machines operating on GMMs (!)

How Do You Do It? (2) Also use high-level information! From JHU 2002 Super. SID Final Presentation – Reynolds et al.

How Well Can It Be Done – Who Salutes? From NIST 2008 SRE Presentation, Martin & Greenberg

More Salutes From NIST 2008 SRE Presentation, Martin & Greenberg

From Europe From NIST 2008 SRE Presentation, Martin & Greenberg

More From Europe From NIST 2008 SRE Presentation, Martin & Greenberg

U. S. Entries From NIST 2008 SRE Presentation, Martin & Greenberg

How Well Can It Be Done – Testing Conditions • Conditions for different amounts of data – 10 sec. – 3 -5 minutes – 8 minutes – Separate channel and summed channel conditions • English-speakers, non-English speakers, multilingual speakers

How Well Can It Be Done?

Speaker Verification Project • Implement a Speaker-ID system – Template based – GMM based – SVM based – Vector space model • Demonstrate it: – NIST data, e. g. 2001 Evaluation – Your own voice – implement on laptop

Speech Recognition Project • Implement an HMM based recognition system • Use, e. g. , Phonebook isolated word data set or Aurora digit set • Write features with existing front-end • Build your own HMM trainer/decoder • Set it up on your laptop for online word recognition (? !)

Highlights of Syllabus • Required Texts: – Huang, Acero, Hon: Spoken Language Processing – Deng and O’Shaughnessy, Speech Processing – EE 516 Reader, at Professional Copy ‘n Print, 4200 University Way • Grading: – Projects: 50% – Final Exam: 30% – Homework 20% • Projects: – Small team or individual • Teams are self-forming – Presentation times TBD – Read ahead & pick an area!!! • Talk to relevant instructor – Suggest deciding no later than 4/30 • Office Hours at end of class and by appointment • Please sign in on email list!