Regional Meeting for Central and Eastern Europe organized by International “eAccessibility in Television Broadcasting in Central and Eastern Europe” (HRT Academy, Zagreb, Croatia) – 3-4 December 2013) Subtitling (captioning) Gion Linder Head of subtitling SWISS TXT Chairman Eurovision Access Services Experts Group gion.linder@swisstxt.ch Zagreb, Croatia, 4 December 2013 Topics • • • • • • • • • • • • • What are subtitles? Target groups Legal issues Technology Input technologies Content access Reception of subtitles Subtitling market Quality Profession subtitler/respeaker Subtitles on all devices Added values International 2 What are subtitles? Subtitles are is a textual version of the dialog or commentary … [in audiovisual media]. (Wikipedia) • Intra- / interlingual subtitles • Closed / open subtitles 3 Target groups Hearing impaired • • Deaf Hard of hearing • • • • (WHO: 5% or 360’000’000) age-related profound Late deafened Deaf children Foreign language comprehension • Learners 4 Legal issues (1/2) There would be no such service without legislation (except the US) Fields: • Only broadcast or online as well? • Subtitles, signed programs, clean audio 5 Legal issues (2/2) • European directives • • • EU encourages member states National legislation Voluntary service 6 Technology • • Server Transmission system • • • DVB, Teletext or online Exchange format: STL EBU TT Equipment providers • • • • Screen (Sysmedia) FAB Softel Cavena (Scantitling) 7 Input technologies (1/10) Keyboard • • Ordinary Keyboard Stenotype • • Short key systems • • e.g. Norway Fast typers • • only in the UK and US (and Italy) e.g. Austria Dual keyboards 8 Input technologies (2/10) Voice recognition • Respeaking • Dragon and IBM • • Respeaking with correction • • Available in English, French, German, Spanish, Italian and Dutch France (correct sentences vs. delay) Automatic voice recognition 9 Input technologies (3/10) 10 Input technologies (4/10) Automatic subtitling = recognition + creation of subtitles • for live programs, with subsequent correction, alignment service • Problems that might occur • Challenges 11 Input technologies (5/10) Automatic subtitling NER value from different providers Prepared Spontaneous German 87.5 81.1 90.3 83.1 94.9 82.9 French 93.6 91.5 94.4 Italian 93.0 92.3 90.8 86.9 86.6 85.0 • analyzed with NERstar 12 Input technologies (6/10) Automatic subtitling - Difficulties • • • • • Spoken vs. written language Different from standard language Background music/noise Big variation within a program Wrong recognized words 13 Input technologies (7/10) Savas project • • • EU funded, 2 years 6 languages: German, French, Italian, Spanish, Portuguese and Basque Goal: live subtitling of news programs 14 Input technologies (8/10) 15 Input technologies (9/10) Automatic subtitling - Conclusions • • • • • Not all kind of programs suit Standard language Without spontaneous speech Homogenous program format Noise and background sound standardized 16 Input technologies (10/10) Automatic subtitling • • • • Possible area of use Local programs Automatic recognition with postponed correction Subtitles available on the internet and HbbTV 17 Content access • Prerecorded programs • • • • • Access to content as early as possible Convert to low res (+ meta data) Distribution Attach to data base or copy on MXF Live programs • • • Access to news desk Access to prerecorded videos Near live: access to audio at least 18 Reception of subtitles (1/5) • Display on screen • Common sense • • • • 2 lines, on the bottom Double height Only bright colors Duration chps or wpm 19 Reception of subtitles (2/5) • Processing live subtitles • • Scrolling vs. blocks Time spent on images Hearing Hard-of-Hearing Deaf Blocks 33.3% 33.2% 31.7% Scrolling 11.7% 11.4% 14.3% Source: Pablo Romero-Fresco 20 Reception of subtitles (3/5) 21 Reception of subtitles (4/5) 22 Reception of subtitles (5/5) 23 Subtitling market • Prerecorded programs • • • • • Hard competition Easy to enter the market Live programs Follow the value chain Costs for a broadcaster • Different ways to fulfill: With or without an own staff 24 Quality (1/7) • Definition of quality • Pretty easy for prerecorded programs • Correctly, regarding content and grammar Word order and key words Verbatim as long as possible Description of non visible sounds Linguistic nuances, ambiguities, irony In sync and frame accurate • Easy language? • • • • • 25 Quality (2/7) • Definition of quality • For live programs • • • Verbatim vs. slight synthesis Scrolling vs. block Real live vs. postponed editing (delay!) 26 Quality (3/7) • Quality assessment • For live programs • • • Word error rate NER model Assessment in the UK • • So far: WER New: NER model • • • 90 minutes per year per channel News, entertainment and sports Supervision by Roehampton University 27 Quality (4/7) 28 Quality (5/7) • • NER model: error examples Serious he’s having problems with the cheques instead of he’s having problems with the Czechs • Normal he’s a buy you a bull asset instead of he’s a valuable asset • Minor imon brown has been appointed new chairman of Rolls S Royce instead of Simon Brown has been appointed new chairman of Rolls Royce 29 Quality (6/7) NERstar tool • All you need: • • video, subtitle, transcript Calculates the NERvalue and the delay 30 Quality (7/7) NER model • British Ofcom plans: • • 6 videos, 5 minutes each for news, entertainment and sports Supervising Roehampton university 31 Profession subtitler • • • • Ordinary subtitler Respeaker In the future only corrector? Precarious work situation 32 Subtitles on all devices • • • From linear to non linear New format EBU TT On all devices? • • • • On On On On the own webpage? HbbTV? TV service providers like Zattoo? mobile devices? 33 Added values • Subtitles may improve • • • the search in the own archives the search for video search engines Subtitles may be useful • for recognition software 34 International • • Awareness EBU: Eurovision Access services experts group • • • • Knowledge exchange Interfere when necessary ITU: Focus Group Other stakeholders 35 Thank you! gion.linder@swisstxt.ch Geneva, Switzerland, 24 October 2013 36