±«Óãtv

Experiences team

This week, the Cross Platform Authentication (CPA) protocol specification has been published as an ETSI standard. You can read more about CPA in Chris’s blog post  and the standard itself is available . Chris and Michael from the EBU are now working on updating the documentation of our to point to the published standard.

Zillah has been over at the in New York along with Chris Pike from the North Lab and members of the VR production team from VRTOV, setting up their immersive installation: The Turning Forest. The piece showcases future possibilities of 3D sound in VR, so they’ve been building a giant soundproof tent for the best possible experience. Research into 360 video has continued with Henry investigating the possibility of streaming a live episode of Click Radio from the Radio Theatre in 360 video. And Andrew has been analysing the qualitative data from the transcripts from the 360 viewing user testing.

Development on the atomised news prototype continued, Lei and Tom started working on tracking analytics. Anthony has been working on the server infrastructure, deployment, and caching using the ±«Óãtv’s Cosmos platform on AWS. Chris and Chrissy implemented an API for retrieving the definitions and article data. After feedback from the user testing, Thomas has been refining the design for the initial launch and has been exploring designs for future versions.

Libby has built a Radiodan for a connected studio project. It's like the Archers Avoider but with specific pieces of a programme that you can avoid other pieces of the programme with. This helped us a lot with improving Radiodan's audio (we're now using a DAC) and also helped us debug the physical interface - (buttons and dials). 

Andrew’s also been doing research into Amazon X-ray type services, aiming to create some design mock-ups for the story explorer. Alan’s been parsing Casualty scripts from Final Draft XML to extract characters and is starting to extract some insights. Henry, Libby, Joanne and Calliope have been starting to think about quick prototypes that could be made to explore some research questions around set top boxes. They have been working on a questionnaire to circulate about people's behavior when they watch TV.

Data team

Tom has been making great progress with the Speech/Music discrimination project. He has generated some synthetic data to train and evaluate the classification of different combinations of audio types. He has also added smoothing to his Support Vector Machine classifier, which has improved the overall accuracy and also reduced the number of smaller segments returned by the algorithm.

Ben and Jana have come up against an interesting issue when running the CODAM finger-printer over ±«Óãtv Red Button content. It turns out the Red Button broadcasts video with a slightly “stretchy” (technical term) pixel aspect ratio - we think this is done in order to broadcast widescreen on limited bandwidth. The different aspect ratio caused a few issues, which means we need to pre-process Red Button content to convert to a 1:1 ratio. 

The data team is about to release a new version of Kaldi with an improved accuracy, which will go out to our colleagues in Rewind and ±«Óãtv News. They are also now looking at a task called “Long Audio Alignment” - where we have a transcript for some audio but no timestamps. The task is to best align the words to the audio. This has many uses, from improving subtitles accuracy to being able to generate new training data from the many transcripts the ±«Óãtv owns. We will be assessing open source tools and our own technology against a peer reviewed test set from the .

With some guidance from Nick H, Denise has started to use LastFM tags over to train the playlist classifier she is working on for Radio & Music (R&M). She has also discovered there may be more training data available from the R&M systems and is working towards integrating that into her system.

Discovery team

For the discovery team, the focus this sprint was mainly on the applications of our project and some multilingual work for our semantic tagger.

Last sprint we'd almost finished building a prototype around curating music articles based on sentiment analysis, and this sprint we got to test it in the lab with a selection of users. Olivier put some finishing touches on an algorithm to improve the quality of the content mix for the prototype, building a selection of items with a good mix of sources and people, and ensuring that only articles with strong sentimented language are represented. We took the finished prototype to two days of user testing, and are currently waiting on the analysis of the collated findings.

In the meantime, we continued work on the underlying platform built on all our editorial algorithms. We scaled up the number of AWS workers with use to cope with the growing number of sources they analyse. Work also started on a system to automatically construct a set of "smart topics" to generate a stream of content for any given interest.

Michel worked on a script to batch process subtitles, and in parallel also set up an API to allow audio to be submitted to our speech-to-text system via HTTP requests.

On the  side of things, there has been some progress on the multilingual side of our semantic tagger. Chris explored sources of ground truth data in order to evaluate multilingual tagging and classification for the systems we develop. He created Farsi and Spanish ground truth tagging dataset from World Service data and used them to test the tagger in these languages.

Lastly we would like to welcome Calliope Georgousi, a UX trainee, who’ll be with us for 3 months. 

Interesting links: