Date: 9 Dec 2014
Lecture by Dr. Frank Melchior, Head of Audio Research, BBC R&D
The audio research group at BBC R&D has for the past four years been developing the next generation of audio for broadcast. Working with academic partners in the BBC Audio Research Partnership, the group has created novel audience experiences based on new forms of audio content representations. This work has led to the launch of large-scale funded projects involving a significant number of researchers and international industrial partners.
Dr. Melchior will outline a number of public trials around object-based audio and binaural experiences which demonstrates his vision of the future of broadcast. His presentation will concentrate on the audience experience, but will also detail the technological challenges from a content provider perspective. In addition to the latest research results from the BBC and its academic partners, the talk will highlight four projects in depth:
- An interactive web-based football broadcast, which includes the ability to adjust the level of commentary versus background.
- An immersive radio drama, produced in stereo, surround and immersive audio versions based on a single mix.
- An object-based binaural production system, used to create and deliver a binaural experience to the audience, receiving very positive feedback.
- A variable-length documentary, whose duration can be adjusted without compromising on the quality of the storytelling.
Date: 10 Dec 2014
Dr. Daniel Wolff, City University London
Music similarity estimation is a key topic in Music Information
Retrieval. In scenarios such as music exploration or recommendation,
user satisfaction depends on the agreement between the user and the
system on which music is more and which is less similar. The perceived
similarity is specific to the individual user and influenced by a number
of factors such as cultural background, age and education. We will
discuss how to adapt similarity models to the relative similarity data
collected from users, using machine learning techniques or metric
At this point, there are few similarity datasets available for training
and evaluation of such systems. We will present the “Spot the Odd
Song Out” game, which collects relative similarity judgements of users
on triplets of songs: Players are they are asked to choose one song as
the “odd song out”. This data is annotated with user attributes such
as age, location and spoken language. The game is designed as
multi-player and rewards blind agreement of players. Based on the
CASimIR API, it has been extended to multiple question types and
scenarios including annotations of tempo, rhythm and further
classification. Game URL: http://goo.gl/6sNcmm
the event is completely free, but it would be great if you could register so we can manage the room/coffee more effectively: https://www.eventbrite.com/e/aes-midlands-lecture-spot-the-odd-song-out-a-system-for-music-similarity-estimation-tickets-14134182721
Daniel Wolff recently finished his PhD on “Similarity Model Adaptation
and Analysis using Relative Human Ratings” at the Music Informatics
Group of City University London, now researching in the Digital Music
Lab project. Apart from modelling music similarity, his past research
includes feature extraction from audio with a focus on periodic
patterns, as well as computational bioacoustics with a focus on birdsong
recognition. He is furthermore an active musician in the City University
Experimental music Ensemble. Author’s homepage: http://www.soi.city.ac.uk/~abdz038