Opening up the BBC's archive
A novel method of on-line programme discovery by automatically classifying the programme's mood using signal processing and machine learning techniques.
What we're doing
We have built a prototype mood classification system for discovery of iPlayer programme content. We have developed algorithms that analyse the audio and video of a programme and automatically create mood metadata for each programme.
Mood and happiness, Snippets Radio Waveform, Comma, Vistatv visualisation dashboard, World Service Archive, accessibility
Another busy week for IRFS, new projects, international conferences and more!
A bumper couple of weeks, with lots of project activity, talks and a whole-team hardware prototyping event keeping us busy.
Weeknotes from IRFS featuring iPlayer logs, elasticsearch, machine listening, W3C Audio API, FI-Content 2, ABC-IP
A new person joins the team, we do some hardware hacking and bring a wealth of radio content to Snippets.
More project info
Why it matters
We want to open up the BBC's archive and enable discovery of programmes unknown to the public. To enable content discovery we need high level metadata about these programmes that is useful and understandable to the public.
Due to the vast size of the BBC's archive, we are researching automatic techniques for high-level metadata generation for entertainment.
To develop automated methods to create high-level metadata for the BBC's archives
To determine which low level features are required to generate high-level metadata
To research what kind of high-level metadata are useful
How it works
We automatically generate mood metadata from programme content using signal processing techniques to analyse the low level audio and video, then classify in terms of high level mood using machine learning techniques.
We conducted a large study with 200 participants who watched short excerpts from TV programmes and assigned mood labels.
From the overall consensus, the moods were reduced to two principal dimensions, the first relating to the seriousness or light-heartedness of a programme (serious/humorous), the second describing the perceived pace (slow-paced/fast-paced).
Machine learning classified both mood dimensions to a high degree of accuracy, reaching more than 95% for programmes with clear moods.
Each classified programme is represented as a dot on the mood chart with serious/humorous along the x-axis and slow-paced/fast-paced along the y-axis, enabling any of the classified programmes to be selected by its mood.