Live subtitling with speech recognition Pilot research project and training at the University of...
Transcript of Live subtitling with speech recognition Pilot research project and training at the University of...
Live subtitling with speech recognitionPilot research project and training at the University of Antwerp and ArtesisUniversity College.
I. Research: Tijs Delbeke (research assistant), Mariëlle Leijten, Aline Remael & Luuk Van Waes (supervisors)
II.Training: Veerle Haverhals (Artesis/VTM)
2
Today’s programme
I. Research at UA-AHA (Oct. 2008-Jan.2009)1. Observational research 2. Experimental research (data to be processed)
II. Training: research & practical at UA-AHA1. MA dissertations (UA & Artesis)2. Within the MA in translation/interpreting at
Artesis3. Course structure & content at Artesis
3
Purpose of the Research
Short term:• Create a classification of different types of
reduction, error (production), delay and their interaction (delay = dependent variable)
Longer term:• Identify the ‘ideal reduction rate’• Identify the ideal respeaker-profile• Improve live-subtitling procedures
4
Two stages in research: both with ‘Inputlog’
Observational research Experimental research
‘Real live’ footage Recorded ‘as live’ footage
Sports programs Talk show
Observational Experimentally controlled
5
Participants
• 12 live subtitlers • Flemish Public Television (VRT)• 8 men, 4 women• Various experience levels (1-7 years)
6
I. 1. Observational Research
1. Live subtitling process: a schematic overview2. Corpus 3. Reduction4. Delay5. Error production
7
1.1. Production of live subtitles: overview
spoken > respeaking > speech > subtitle tv comment recognition
(1) (2) (3) x x+t
reduction correction error production
delay
8
1.2. First corpus• Flemish Public Television (VRT)• 15 hours of sports programs• Transcriptions & broadcast subtitles• Time stamps• Character & word counts• Audio recordings• Detailed logging data (inputlog)
- Speech input- Keystrokes- Mouse movements
9
1.3. Reduction
• Verbatim vs. reduced/summarized/edited/condensedContinuumLargely program dependentReduction crucial:
- Slower readers- Speech recognition constraints
- Quantitative analysis- Qualitative analysis
10
Reduction
Quantitative analysis
• -30% (football) • -45% (tennis) • -60 % (cycling)
Reduction table, example
11
Reduction (2)
Qualitative analysis
• Causes of reduction• Reduction classification
- Literature: only vaguely- 3 main classes - 30 categories
12
Reduction (3)
Qualitative analysis
- Reduction to prevent delay (49%)- Forced Reduction (22%)- Time-induced reduction (15%)
13
Reduction (4)
Qualitative analysis
• Prevention of delay- Deletion of redundant info
Repetition, obvious element, hesitation, interjection, …
- SubstitutionNames, metaphors, idioms, …
SUBTITLE SPOKEN COMMENT
But they can forget about that, I think. But they can forget about that, I think. They can forget about that
14
Reduction (5)
Qualitative analysis• Forced reduction
Erroneous grammatical construction, too difficult for respeaker/speech recognizer, meaning unclear,…
• Time-induced reductionComplicated interaction, sudden event, prepared title
coming up, not relevant anymore,…
SUBTITLE SPOKEN COMMENT
Cercle very dangerous using that combination.
Iachtchouk. De Smet. Passes back. Van Mol. De Sutter. Crosses. Yes. Cercle Brugge very dangerous using that combination.
15
1.4. DelayFactors
• Block mode vs. scrolling mode• Additional corrector vs. self correction• Reduction degree (mutual process)
Delay table, example
• 6 sec : cycling (-30% red.)• 11 sec : football & tennis (-45 & -60% red.)
16
1.5. Error production
• 6 fragments of 60 titlesQuantitativelyPure recognition:• Title: 72,22% (7 out of 10 titles correct)
After correction:• 84% corrected --> 93% titles correct.• 22% by respeaker vs. 78% by corrector• 12% with speech vs. 88% with keyboard and mouse
17
1.5. Error production (2)
Qualitatively• Classification model• Based on Karat (1999) & Leijten (2007)
18
1.5. Error production (3)1. Technical errors (71,6%)
- a. Erroneous Recognition» i. One word» ii. Multiple words» iii. Proper names (20,6%)» iv. Geographical names
- b. Erroneous Interpretation» i. Command as text» ii. Text as command» iii. Word as letter» iv. Letter as word» v. Abbreviation or acronyms as words
- c. Programming Errors» i. Grammatical error» ii. Background noise as text» iii. Crash
19
1.5. Error production (4)
2. Human errors (14,3%)- a. (Corrector)- b. Respeaker
» i. Misinterpretation» ii. Wrong word» iii. Additions or transformations» iv. Formal revision
3. Technical & Human errors (1,6%)- Slurred speech/mumbling or inaccurate recognition?
4. Other Errors (12,5%)
20
2. Experimental Research
• Infotainment talk show ‘Phara’• 3 excerpts (15 minutes)
21
2.1 Method: procedure• Backward Digit Span• Reading task• Verbatim subtitling (9 min)
Aim at 100% subtitling. Quantity > Quality.
• Summarized subtitling (15 min)Aim at 50 % subtitling. Quantity = Quality. (usual)
• Heavily reduced subtitling (15 min)Aim at 25 % subtitling. Quantity < Quality. (no errors)
• Concluding interview
22
2.2 Results
Quantitative analyses of 1 excerpt• Reduction• Error production• Relation reduction & error production
23
2.2 Results: Reduction (1)Subtitling percentage in function of reduction mode
0%
20%
40%
60%
80%
100%
1 2 3
Reduction Mode
Su
bti
tlin
g %
8,00
8,50
9,00
9,50
10,00
10,50
11,00
11,50
12,00
12,50
Tit
les
pe
r m
inu
te
Demanded subtitling %
Performed subtitling %
Subtitles (number)
24
2.2 Results: Reduction (2)• Fairly inaccurate execution of demanded
reduction mode- Subtitling percentage lower than demanded
• Verbatim (100%) 51%• Summarized (50%) 38%
Important: Theoretical OptimumStop words
RepetitionsHesitations …
- Subtitling percentage higher than demanded• Highly reduced (25%) 35%
25
2.2 Results: Reduction (3)
• Reduction mode affects number of broadcast subtitles Less reduction = more titles
• Reduction mode moderately affects subtitle length Longer titles for verbatim mode
26
2.2 Results: Error ProductionError rate
0,00%
5,00%
10,00%
15,00%
20,00%
25,00%
30,00%
1 2 3
reduction mode
Err
or
%
Title level
Word level
27
2.2 Results: Error Production (2)
Title level Word level
Verbatim
Summarized
Highly reduced96% 99,5%
Level
Red. Mode
95%
98%
73%
89%
Accuracy per reduction mode
28
2.3 Concluding remarks
• Indication of maximal performance (verbatim subtitling)
• Error in 3 out of 10 subtitles
• Indication ‘normal’ performance• Error in 1 out of 10 subtitles
• Subtitle production drops after 10 minutes • More reduction yields more accurate subtitling
29
II. Training: 1. MA dissertations
MA dissertations in support of ongoing research: error analyses, trial classifications, reception research, Dragon training, …
- UA (Master in multilingual business communication)- Artesis (Interpreting, 2007-2008)
30
II. Training: 2. Interpreting –general (1)
At Artesis:
- MA in Interpreting- European Master in Conference Interpreting
31
II. Training: 2. Interpreting - general (2)
At Artesis: MA in Interpreting= initiation in different types
Community InterpretingBusiness InterpretingIncludes consecutive interpreting, speech training, research topics, institutions, …
Option: Live subtitling with speech recognition (Dragon)
32
II. Training: 2. Interpreting – Live subtitling
Research training (beside MA theses)
- Within interpreting programme Artesis- Within AVT programme Artesis
Practical training- Within translation programme: subtitling (sem. 1)- Within interpreting programme Artesis: live ST (sem 2)
Veerle Haverhals: MA in interpreting and full time respeaker at VTM
33
II. Training: 2. Interpreting – Live subtitling: course topics practical training (1)
- Initiation to DRAGON: make a profile, try out all the functions, add terminology and test it.
- Working with codes, anticipating mistakes (e.g. TOX-Leterme)
- Test accuracy of the above with CRER (terminology added/or not, terminology without ‘TOX’): get acquainted with errors.
34
II. Training: 2. Interpreting – Live subtitling: course topics practical training (2)
- Live subtitling in Flanders & the Netherlands: programmes, challenges, speed, different speakers + examples
- Visit to VRT: live cycling session
- Introduction to “News production” at VTM, in preparation of internship at VTM
35
II. Training: 2. Interpreting – Live subtitling: course topics practical training (3)
Series of sessions to train respeaking (to be expanded)- Summarizing for deaf/hard of hearing (choice of words)
- The use of colours (or not)
- Multi-tasking in real time:corrections, colours
- Seek compromise: completeness/errors
36
II. Training: 2. Interpreting – Live subtitling: course topics practical training (4)
Special issues:
- Linguistic variation (or not)
- Onomatopeia (or not)
37
II. Training: 2. Interpreting – Live subtitling: course topics practical training (4)
One day internship at VTM- Watch news broadcast + question time- Live simulation of the one o’clock news
Preparation (cf. above):Learning to use the software(s), marking live passages, combining prepared with live, studying key codes, forwarding the subtitles, correcting and forwarding, …
.
38
Literature• Baaring, I. (2006). "Respeaking-based online subtitling in Denmark." InTRAlinea. SPecial issue: Respeaking.• Daelemans, W., A. Höthker, et al. (2004). "Automatic Sentence Simplification for Subtitling in Dutch and English."
Proceedings of the 4th International Conference on Language Resources and Evaluation: 1045-1048 • de Korte, T. (2006). "Live inter-lingual subtitling in the Netherlands." InTRAlinea. SPecial issue: Respeaking.• Den Boer, C. (2001) “Live interlingual subtitling.” Gambier & Gotlieb (2001)• Gambier, Y. and H. Gottlieb, Eds. (2001). (Multi) Media Translation. Concepts, Practises, and Research.• Jones, R. (2002). Conference Interpreting explained.• Karat, C. et al. (1999). “Patterns of entry and correction in large vocabulary continuous speech recognition
systems.” Paper presented at the CHI 99, Pittsburg.• Lambourne, A. (2006). "Subtitle respeaking." InTRAlinea. SPecial issue: Respeaking.• Lambourne, A., J. Hewitt, et al. (2004). "Speech-based Real-time Subtitling Services." International Journal of
Speech Technology 7: 269-279.• Leijten, M. (2007). “Writing and Speech Recognition: Observing Error and Correction Strategies of Professional
Writers.” Utrecht: LOT• MacArthur, C. A. (2006). The Effects of New Technologies on Writing Processes. Handbook of Writing Research. C.
A. MacArthur, S. Graham and J. Fitzgerald.• Mack, G. (2006). "Detto scritto: un fenomeno, tanti nomi." inTRAlinea. SPecial issue: Respeaking.• Ogata, J. and M. Goto (2005). "Speech Repair: Quick Error Correction Just by Using Selection Operation for Speech
Input Interfaces." Proceedings of Interspeech 2005: 133-136.• Remael, A. (2004). Vertaling in beeld: audiovisuele vertaling en ondertitels.• Robson, G. D. (2004). The closed captioning handbook. • Slembrouck, S. and M. Van Herrewege (2004). Teletekstondertiteling en tussentaal: de pragmatiek van het
alledaagse. Schatbewaarder van de taal. Johan Taeldeman. Liber amicorum. J. De Caluwe, G. De Schutter, M. Devos and J. Van Keymeulen.
• van der Veer, B. (2008) De tolk als respeaker: een kwestie van training.• Wald, M., Boulain, P., Bell, J., Doody, K. and Gerrard, J. (2007) “Correcting Automatic Speech Recognition Errors in
Real Time.” International Journal of Speech Technology
Thank you for your attention