1. From Word Spotting to OOV Modeling
OOV OOV spotting OOV OOV OOV
[f r ah m] OOV spotting [t uw] OOV OOV
[f r ah m] [w er d] spotting [t uw] OOV [m aa d el ih ng]
From Word Spotting to OOV Modeling
Paul Fitzpatrick (6345g11)
Goal
To automatically extract filler vocabulary for word- spotting
Why?
So language model has something to work with
May improve recognition accuracy on keywords
Gives earlier payoff in domain-specific training
Scenario
Start with small lexicon (e.g words)
Start with weak language model
Bootstrap by clustering filler vocabulary from large collection of untranscribed data

2. Methodology Run recognizer Extract OOV fragments Identify competition
Identify rarely-used additions
Remove from lexicon
Add to lexicon
Update lexicon, baseforms
Hypothesized transcript
N-Best hypotheses
Update Language Model

3. Results Initial lexicon
, phone, room, office, address
Top 10 OOV clusters found (ranked by frequency)
1. n ah m b er 6. p l iy z
2. w eh r ih z ae ng k y uw
3. w ah t ih z 8. n ow
4. t eh l m iy 9. hh aw ax b aw
5. k ix n y uw g r uw p
Example sentence hypothesis
(w ah t ih z) (ih t er z uw) room (n ah m b er)
What is Victor Zue’s room number?