This project is no longer listed in the FindAPhD database
and may not be available.
a) Project aim, objectives and outline Querying, searching, and mining of time series form an integral research tasks in many branches of science and engineering. These tasks have also numerous industrial applications, ranging from stock market prediction to query by singing/humming. At the very center of the time series analysis is comparison of two or more time series. The comparison operation focuses on determining a distance between the analyzed sequences according to some well defined metric.
QBSH [1] allows a user to sing or hum a tune to a microphone. The extracted pitch curve (a time series) of the tune is compared with other tunes in the database (also in the format of time series), and returns similar songs in a ranking list. The main aim of this project is to study efficiency and effectiveness of time series comparison, with the emphasis on the application of query by singing/humming (QBSH).
The main objectives of this project refer to further study and in turn to improvement of integral elements of the QBSH search mechanism including:
o Multi-stage progressive filtering: Search for simple but fast filtering methods in the database, followed by more sophisticated, however, effective methods for detail comparison. This includes fine tuning search parameters for such multi-stage progressive filtering.
o Indexing for different comparison methods: Basic methods include linear scaling and dynamic time warping. We will adapt or modify already known solutions to make them suitable for QBSH search.
o Repeating pattern identification: In order to speed up search in a large volume (20000) of songs present in QBSH we will also explore regularities including repeated patterns.
o Distributed & parallel computation: We will also work on parallelization of the QBSH time series comparison. Our initial attempt shows the speedup factor is 66 for using a GPU with 384 cores.
b) Pattern of studies After taking the qualification exam (coinciding with the conclusion of MSc studies) the student will be enrolled on the 4 year PhD program. The student is expected to spend a minimum of 12 months at either of the participating institutions.
c) Facilities This project requires the use of standard research facilities including access to library resources, the internet, and a dedicated programming environment. This project provides an opportunity to study both the theoretical as well as more practical aspects of time series comparison. The main focus of the studies in Liverpool will be on the design and analysis of efficient sequence matching algorithms. The time spent at NTHU will be mostly devoted to the analysis of large-scale dataset for QBSH and QBSH prototyping using GPUs [1,2].
Funding Notes:
This project is a part of a 4-year dual PhD programme between National Tsing Hua University (NTHU) in Taiwan and the University of Liverpool in England. Students can start in either institution, after passing a ‘qualifying examination’ and must spend at least 12 months in each institution. When away from the ‘home’ institution they will receive a stipend of 10,000 TWD. There are no tuition fees for students starting at UoL and for students starting at NTHU no tuition fees will be charged while in Liverpool. Applications can be made on-line to either institution mentioning the dual PhD programme.
References:
[2] J.-S. Roger Jang and Hong-Ru Lee, A General Framework of Progressive Filtering and Its Application to Query by Singing/Humming, IEEE Trans. on Audio, Speech, and Lang. Proc., Vol. 16(2), pp. 350-358, 2008.
[3] QBSH prototype using GPU: http://mirlab.org/demo/miracle