Technological research results at INA
The Research Area offers demonstrations of tools, processes or approaches in connection with INA's research activities. Here you can see, test or experiment with projects currently being conducted by INA. In an easy-to-use environment, the Research Area presents in a clear manner the experiments, methods of use and concepts implemented.
This area brings together some salient INA developments in the field of digital media. The Research Department spotlights our prototypes and demonstrations to show how our researchers and research engineers have contributed to the elaboration and use of these tools. The Research Area welcomes reactions, questions and comment, with a view towards improving the results or even initiating new collaborations.
To find out more about INA's research activities and themes, visit the company's site
Amalia.js video player is now open-sourceTry it !
Demonstrations and prototypes
The French National Audiovisual Institute (INA) provides the scientific and technological community a corpus of audiovisual documents from its collections, document sheets and metadata associated with these documents.
This corpus is intended for finalising, experimentation and evaluation of search and analysis tools for multimedia content, strictly as part of scientific research. To access the Corpus, you must have first subscribed and have an FTP client available allowing you to download the Corpus.
The Corpus is made available under the conditions specified in the Conditions of Use, to any legal entity having previously subscribed and accepted all the aforesaid Conditions of Use.
Subscription is restricted to research laboratories, innovative SMEs as well as all other legal entities having a scientific research department or activity.
Your request will be sent to INA for consideration. After your request has been approved, you will receive an e-mail confirming the address for the FTP server as well as the confidential login name and password assigned to your Organisation, allowing you to access the Corpus.
Videos in INA's archives contain a wide variety and large numbers of visual entities that are really worth annotating if they one day need to be found quickly. These visual entities, which may be logos, works of art, building, faces, etc., are difficult to annotate by humans, even expert humans like INA's research assistants. First of all because this is a very time-consuming task; then because certain entities are so rare that few people are able to identify them. So the purpose of DigInPix is to automatically identify as many visual entities as possible from a predefined list of entities grouped into "dictionaries". Albeit more difficult, the system can sometimes detect an entity occupying a very small portion of the image. On the other hand it is not suited to categorizing images or videos (sunsets, beach scenes, mountain scenes, dogs, etc.).
The synopses drafted by INA's archivists do not always include information that can pinpoint them on the medium (start and end times, or "TC in" and "TC out"). This demonstration illustrates an automatic time-slicing process for TV news programs and automatic time-stamping ("TC in" and "TC out") of the related synopses.
First of all, "on set" sequences are pinpointed by automatically identifying the voice of the presenter with software developed by the University of Maine's IT Laboratory (LIUM). Speech is automatically transcribed with the VoxSigma software package (developed by VOCAPIA Research). Each news item synopsis is then "positioned" on the timeline of the news program. To that end the text of the synopsis (title, summary and keywords) is compared with the portions of the transcription that temporally match studio scenes and sequences between them, using a comparison measure that takes into account all the words in the synopsis on the one hand and the portion of the transcription on the other. Each synopsis is then matched up with the time segment whose similarity between the synopsis and the transcription portion is the closest. A distinction is made between studio items, which are positioned within studio sequences, and other items such as reports (positioned between the studio sequences).
This study is the direct result of needs stated by INA's operational sectors. The two main uses are:
From the 1930s to the 1960s, all audio recordings of radio stations were kept on 78rpm discs (tapes only started to be used routinely from 1960). Over time these records (discs) have deteriorated through handling or natural causes. Some records having priceless historical value can no longer be read using conventional means (magnetic cells with a diamond stylus) and the surface of the disk cannot be mechanically restored. Some 20,000 of INA's 276,000 discs are in such a condition. The same applies to large amounts of archives of radio broadcasting organisations all over the world.
Scratched, broken or split discs cannot be played by mechanical means. This is also true of discs whose surface has altered and deteriorated.
Our process can read these discs optically, without a stylus and without any physical contact.
Research and Development Project Manager
HMI and experimentations Research Engineer
Write us : BOBINO@INA.FR
Famous voices haunt the French audiovisual heritage. Politicians, athletes, intellectuals, anchormen, our broadcast history resonates with heroic, tender, comical or dramatic accents.
Speech processing technologies allow the extraction of valuable information from the audio signal. Discrimination of speech and music zones, automatic segmentation in speaker turns and identification of a person using his voice characteristics or automatic speech recognition, there are nowadays many tools available.
The aim of the Speech Trax demo (for speech tracking) is to gather results of these multiple analysis in order to visualize them in an interface allowing an easy exploration. Choice was made to treat radio and television data from March 2014 for public channels France 2, France 5, France 24, France Inter, France Info and France Culture. This specific time period was chosen due to the important events that occurred (local elections, former president Sarkozy phone tapping, Ukraine invasion by Russian troops, Malaysian Airlines flight 370, etc.) In total more than 250 hours of broadcast news and magazines have been analyzed.
Collaborations with several French universities were led: the Laboratoire d’Informatique de l’Université du Maine (LIUM) and the Laboratoire d’Informatique d’Avignon (LIA) for the speaker segmentation and identification, the Institut de Recherche en Informatique de Toulouse (IRIT) for the speech/music discrimination and the society Vocapia Research for the speech recognition realized using the software VoxSigma.
Please fill in the form below