Scale your database traffic with Read & Write split using MySQL Router
Presentation 17 may morning case study 2 sarahhaye aziz
1. RSI Archive: our experience working with
Speech to Text and Semantic Analysis
Sarah-Haye Aziz and Lorenzo Vassallo
May 17, 2013
2. 2
Come al solito, anche la recente
inaugurazione dell'ultima monumentale
opera di quell'eccezionale scultore che
Giacomo Manzù, vale a dire la nuova porta
del Duomo di Rotterdam avuto il sapore
di un avvenimento straordinario di
risonanza internazionale e per lavori in
corso Fabio Bonetti è riuscito ad avvicinare
l'insigne maestro bergamasco, a buon
diritto ritenuto ormai uno dei più alti
interpreti del nostro tempo, artista fra i più
grandi del secolo e non solo per la misura
del suo talento ma anche per il rigore
morale di cui è sempre stato esempio in
anni di sorta, tormentata, ispirata
attività.
Credits: Giacomo Manzù, Fabio Bonetti
Geographic Therms: Rotterdam
Themes: arte, cultura, intrattenimento
Errors
è
as
Audio Transcription
ha
Categorization
3. 3
Outline
1. Why an automatic indexing system?
2. The project timeline
3. Two paths: system and archivists workflow overview
4. Does it work? We learned that...
5. Next steps
6. Some advices
4. 4
Why an automatic indexing system?
RSI has a consolidated cataloguing system (CMM)
with a well-defined human workflow from 2008
RSI has plenty undocumented historical material
and no capacity to document it.
Increase (plus) the documented material adding
an automation but not substituting (vs) the archivist.
Not vs but plus!
5. 5
Archivists and Technicians Synergy
Project timeline
DeploymentDeploymentTuningTuningAnalysis & StartupAnalysis & Startup
Workflow DesignWorkflow Design
Language ModelLanguage Model
Tv & Radio
Programmes Choice
Tv & Radio
Programmes Choice
Workflow ReviewWorkflow Review
Transcription TestTranscription Test
System TestSystem Test
6. 6
Documenting a material: two paths
Ingestion
Catalogue
Publishing
Transcription Engine
Audio + Key frames
Semantic Engine
Audio
and Video
Key frames
Archivist
Documentation
+
Refinements
Speech
Transcription
Text +
Sequences
Categorization
Text + Sequences
Credits
SIA
Themes +
Geographical therms
Human
audio listening
and
transcription
+
Archivist
documentation
7. 7
The two paths for the archivist
Start ?
Invoke
Indexing
Human Task
on Catalogue
Detailed documentation
Manual creation of
logical sequences
Automated
Transcription and
Categorisation
Detailed documentation
Automatic creation of
logical sequences
Publish
Doc
Level
Basic Human
Limited set of
documented metadata
High Human
with Automation
Limited set of
documented metadata
Automatic creation of
logical sequences
?
?
Human Task
on Catalogue
Yes
No
Doc
Level
High Human
Basic Human
with Automation
9. 9
Does it work? Yes! But…
Differences between Radio and TV
Background Music/Noise does not help the transcription.
Based only on silences and
without key frames, the system
creates too many sequences.
Key frames help to locate a
change of context.
Speech rhythm and pauses are different between and .