Marieke Lycke
AI in the archive,
hype or hope?
October 24, 2019
CONTEXT
0
VRT’s brands
VRT archive
VRT’s brands
VRT’s brands
VRT’s brands
VRT’s brands
VRT’s brands
VRT’s brands
Archive ALL published content
VRT’s archive
VRT’s archive
• The whole archive department: 61 FTE
• Transformation plan: - 25%
• Video and audio: 33 FTE
• 6.000.000 objects (video, audio, music, article, photo)
• Video: 1.7M items or 9300 TB on tape
• 140h or 3300GB additional video’s archived/day
• Audio: 500k items or 5275 TB on disk
• 150h or 26GB additional audio archived/day
• And more to come with growing online media segment…
reduced
- 5 FTE
VRT’s archive
• The whole archive department: 61 FTE
• Transformation plan: - 25%
• Video and audio: 33 FTE
• 6.000.000 objects (video, audio, music, article, photo)
• Video: 1.7M items or 9300 TB on tape
• 140h or 3300GB additional video’s archived/day
• Audio: 500k items or 5275 TB on disk
• 150h or 26GB additional audio archived/day
• And more to come with growing online media segment…
reduced
- 5 FTE
LESS PEOPLE
MORE CONTENT
VRT’s archive
• The whole archive department: 61 FTE
• Transformation plan: - 25%
• Video and audio: 33 FTE
• 6.000.000 objects (video, audio, music, article, photo)
• Video: 1.7M items or 9300 TB on tape
• 140h or 3300GB additional video’s archived/day
• Audio: 500k items or 5275 TB on disk
• 150h or 26GB additional audio archived/day
• And more to come with growing online media segment…
reduced
- 5 FTE
LESS PEOPLE
MORE CONTENT
➔ How should we do that?
VRT’s archive reduced
VRT’s archive reduced
AI and technology
will solve it.
CONCLUSION
3
CONCLUSION
3It is complex and AI can’t solve everything
Strategy
Automated metadata
Manually annotated
metadata
Production metadata
Crowd
generated metadata
Strategy
Automated metadata
Manually annotated
metadata
Production metadata
Crowd
generated metadata
1
A creative process…
A creative process…
Call sheets Subtitles Scripts
Research
documents
Program
guide
information
Playlist
…
Text for
website
Editorial
information
Presentation
text
A creative process…
A creative process…
Opportunities
• Radio shows
Opportunities
• Radio shows use Pluxbox as their editorial system
Opportunities
• Radio shows use Pluxbox as their editorial system
Opportunities
• Radio shows use Pluxbox as their editorial system
News
Text (presentation or item)
Music
Opportunities
• Radio shows use Pluxbox as their editorial system
Opportunities
• Radio shows use Pluxbox as their editorial system
Opportunities
• Radio shows use Pluxbox as their editorial system
Estimated timecodes
Opportunities
• Radio shows use Pluxbox as their editorial system
Clickable time codes
Opportunities
• Radio shows use Pluxbox as their editorial system
Opportunities
• Radio shows use Pluxbox as their editorial system
+
Metadata is made at the source
Time and cost efficient
“Trusted” metadata
More programmes are archived in more detail
Not everyone wants to work with the application
Every user uses the application slightly differently → good agreements
Last minute changes are not (always) updated, so you still have incorrect information.
Time codes are only estimated (work in progress)
-
Opportunities
• Sport editors use Gracenote and add additional information in a
sportservice for their website
Opportunities
Sport
service
Opportunities
•Technological: Sport services
+
Metadata is made at the source
Time and cost efficient
“Trusted” metadata
More programmes are archived in more detail
Timecoded metadata
Not everyone wants to work with the application
Every user uses the application slightly different → good agreements
Last minute changes are not (always) updated, so you still have incorrect information.
Time codes are only estimated (work in progress)
-
Quality of time codes varies
Production metadata, in general
Metadata is made at the source
Time and cost efficient
“Trusted” metadata
More programmes with detailed metadata
Used data is not always meant for archiving purposes
Often more noisy data
Not consistent, depends on program/user
Less control of metadata quality
Human interactions or agreements necessary
+
-
ARTIFICIAL INTELLIGENCE
2
Market research
And others …
Other broadcasters
Text analysis: categorisation,
Named entity recognition, …
Text analysis: categorisation,
Named entity recognition, …
Archive use cases
Speech to text
( + text analysis)
Speaker labeling Face recognition
Speech to text
( + text analysis)
Speech to text
• Dutch is a small language region
• Northern versus Southern Dutch (Flemish)
Speech to text
• Evaluation
• Word Error Rates
Word Error Rate
Phonexia Speechmatics Google Newsbridge
Average 32,9% 32,5% 37,03% 23,64%
Testing an
improved version
Speech to text
• Evaluation
• Word Error Rates
Difficulties
• Seperating music-advertisement-speech parts
• Names
• Other languages
Word Error Rate
Phonexia Speechmatics Google Newsbridge
Average 32,9% 32,5% 37,03% 23,64%
➔ Can be improved by updating dictionary
Speech to text: use cases
Timecoded
metadata?
Long
fragment?
A lot of
speech?
Metadata?
ADDED
VALUE
SPEECH
RECOGNITION
PROOF
One speaker
at a time
No
background
noise
No
dialects or
regiolects
Good
audio
quality
(One
language)
Speech to text: use cases
• Old material, where we have no metadata
• Long talkshows f.e. programme with one guest and long dialogues
Speech to text
Solution for old material → Higher findability of material
Time coded metadata → Better searchability within one fragment
Time efficient
No commitment necessary from programme makers
It creates a lot of metadata, which makes it hard to get an overview
Not all programmes are suitable for speech to text
Quality of metadata? Speech to text makes mistakes
Will we need to train the models?
Will the processing cost outweigh the added value?
+
-
Speech to text
Metadata by Artificial intelligence, in general
+
Time efficient
Starts from content, no human interactions needed
Time coded metadata
AI creates a lot of metadata, and
is not that good (yet) in extracting the essentials (while humans are)
difficult to visualize it in a clean overview
Training of algorithms also takes time
You need good training material
Quality of metdata? AI makes mistakes
Will the processing cost outweigh the added value?
-
CONCLUSION
3
• Technology doesn’t compensate the reduction (yet)
• Most gain for programmes where we didn’t have time for.
• It ‘s an investment that will need time and effort
• There is not one solution or ideal workflow.
• Each techniques has it’s advantages and disadvantages
• Look use case per use case.
• Importance of search: both methods introduce a lot of additional but “noisy” metadata
Conclusion
Production metadata
Quick to implement
Gain from start
Trusted metadata
Artificial intelligence
Timecoded metadata
Starts from content
Solution for old material,
Addition to production metadataPreferred
Conclusion
It’s a hype
But also hope
• It can support archivist with their mission to make material find- and searchable.
• It will take time and effort to make it suitable for archive.
• A selection of good use cases is key, taking available production metadata into account.
AI in the archive,
hype or hope?
➔ We want to experiment and build expertise
Contact
marieke.lycke@vrt.be

LYCKE Artificial intelligence, hype or hope?

  • 1.
    Marieke Lycke AI inthe archive, hype or hope? October 24, 2019
  • 2.
  • 3.
  • 4.
  • 5.
  • 6.
  • 7.
  • 8.
  • 9.
  • 10.
    VRT’s brands Archive ALLpublished content
  • 11.
  • 12.
    VRT’s archive • Thewhole archive department: 61 FTE • Transformation plan: - 25% • Video and audio: 33 FTE • 6.000.000 objects (video, audio, music, article, photo) • Video: 1.7M items or 9300 TB on tape • 140h or 3300GB additional video’s archived/day • Audio: 500k items or 5275 TB on disk • 150h or 26GB additional audio archived/day • And more to come with growing online media segment… reduced - 5 FTE
  • 13.
    VRT’s archive • Thewhole archive department: 61 FTE • Transformation plan: - 25% • Video and audio: 33 FTE • 6.000.000 objects (video, audio, music, article, photo) • Video: 1.7M items or 9300 TB on tape • 140h or 3300GB additional video’s archived/day • Audio: 500k items or 5275 TB on disk • 150h or 26GB additional audio archived/day • And more to come with growing online media segment… reduced - 5 FTE LESS PEOPLE MORE CONTENT
  • 14.
    VRT’s archive • Thewhole archive department: 61 FTE • Transformation plan: - 25% • Video and audio: 33 FTE • 6.000.000 objects (video, audio, music, article, photo) • Video: 1.7M items or 9300 TB on tape • 140h or 3300GB additional video’s archived/day • Audio: 500k items or 5275 TB on disk • 150h or 26GB additional audio archived/day • And more to come with growing online media segment… reduced - 5 FTE LESS PEOPLE MORE CONTENT ➔ How should we do that?
  • 15.
  • 16.
    VRT’s archive reduced AIand technology will solve it.
  • 17.
  • 18.
    CONCLUSION 3It is complexand AI can’t solve everything
  • 19.
  • 20.
  • 21.
  • 22.
  • 23.
    A creative process… Callsheets Subtitles Scripts Research documents Program guide information Playlist … Text for website Editorial information Presentation text
  • 24.
  • 25.
  • 26.
  • 27.
    Opportunities • Radio showsuse Pluxbox as their editorial system
  • 28.
    Opportunities • Radio showsuse Pluxbox as their editorial system
  • 29.
    Opportunities • Radio showsuse Pluxbox as their editorial system News Text (presentation or item) Music
  • 30.
    Opportunities • Radio showsuse Pluxbox as their editorial system
  • 31.
    Opportunities • Radio showsuse Pluxbox as their editorial system
  • 32.
    Opportunities • Radio showsuse Pluxbox as their editorial system Estimated timecodes
  • 33.
    Opportunities • Radio showsuse Pluxbox as their editorial system Clickable time codes
  • 34.
    Opportunities • Radio showsuse Pluxbox as their editorial system
  • 35.
    Opportunities • Radio showsuse Pluxbox as their editorial system + Metadata is made at the source Time and cost efficient “Trusted” metadata More programmes are archived in more detail Not everyone wants to work with the application Every user uses the application slightly differently → good agreements Last minute changes are not (always) updated, so you still have incorrect information. Time codes are only estimated (work in progress) -
  • 36.
    Opportunities • Sport editorsuse Gracenote and add additional information in a sportservice for their website
  • 37.
  • 38.
    Opportunities •Technological: Sport services + Metadatais made at the source Time and cost efficient “Trusted” metadata More programmes are archived in more detail Timecoded metadata Not everyone wants to work with the application Every user uses the application slightly different → good agreements Last minute changes are not (always) updated, so you still have incorrect information. Time codes are only estimated (work in progress) - Quality of time codes varies
  • 39.
    Production metadata, ingeneral Metadata is made at the source Time and cost efficient “Trusted” metadata More programmes with detailed metadata Used data is not always meant for archiving purposes Often more noisy data Not consistent, depends on program/user Less control of metadata quality Human interactions or agreements necessary + -
  • 40.
  • 41.
  • 42.
  • 43.
    Text analysis: categorisation, Namedentity recognition, … Text analysis: categorisation, Named entity recognition, … Archive use cases Speech to text ( + text analysis) Speaker labeling Face recognition Speech to text ( + text analysis)
  • 44.
    Speech to text •Dutch is a small language region • Northern versus Southern Dutch (Flemish)
  • 45.
    Speech to text •Evaluation • Word Error Rates Word Error Rate Phonexia Speechmatics Google Newsbridge Average 32,9% 32,5% 37,03% 23,64% Testing an improved version
  • 46.
    Speech to text •Evaluation • Word Error Rates Difficulties • Seperating music-advertisement-speech parts • Names • Other languages Word Error Rate Phonexia Speechmatics Google Newsbridge Average 32,9% 32,5% 37,03% 23,64% ➔ Can be improved by updating dictionary
  • 47.
    Speech to text:use cases Timecoded metadata? Long fragment? A lot of speech? Metadata? ADDED VALUE SPEECH RECOGNITION PROOF One speaker at a time No background noise No dialects or regiolects Good audio quality (One language)
  • 48.
    Speech to text:use cases • Old material, where we have no metadata • Long talkshows f.e. programme with one guest and long dialogues
  • 49.
    Speech to text Solutionfor old material → Higher findability of material Time coded metadata → Better searchability within one fragment Time efficient No commitment necessary from programme makers It creates a lot of metadata, which makes it hard to get an overview Not all programmes are suitable for speech to text Quality of metadata? Speech to text makes mistakes Will we need to train the models? Will the processing cost outweigh the added value? + -
  • 50.
  • 51.
    Metadata by Artificialintelligence, in general + Time efficient Starts from content, no human interactions needed Time coded metadata AI creates a lot of metadata, and is not that good (yet) in extracting the essentials (while humans are) difficult to visualize it in a clean overview Training of algorithms also takes time You need good training material Quality of metdata? AI makes mistakes Will the processing cost outweigh the added value? -
  • 52.
  • 53.
    • Technology doesn’tcompensate the reduction (yet) • Most gain for programmes where we didn’t have time for. • It ‘s an investment that will need time and effort • There is not one solution or ideal workflow. • Each techniques has it’s advantages and disadvantages • Look use case per use case. • Importance of search: both methods introduce a lot of additional but “noisy” metadata Conclusion Production metadata Quick to implement Gain from start Trusted metadata Artificial intelligence Timecoded metadata Starts from content Solution for old material, Addition to production metadataPreferred
  • 54.
    Conclusion It’s a hype Butalso hope • It can support archivist with their mission to make material find- and searchable. • It will take time and effort to make it suitable for archive. • A selection of good use cases is key, taking available production metadata into account. AI in the archive, hype or hope? ➔ We want to experiment and build expertise
  • 55.