Fraud detection is the current buzzword. Many insurance companies are turning to anti-fraud teams to reduce the number of fraudulent claims and buying into the concept of voiced based lie detection or voice risk analysis.
By assessing the emotional, cognitive and physiological patterns in the spoken word and accurately measuring the frequency patterns of vocal segments of conversations it is possible to detect fraudulent claims. It is difficult for a fraudster, because he or she does not have emotions about the event because it’s not a genuine event. All they have is the hub of the story, without reactions or emotions. Furthermore, a fraudster will often speed up and rush out their stories and are also likely to become increasingly aggressive. They particularly like to treat young females at call centres with aggressive tactics. The project will consider the current industrial solutions to voice lie detection and build upon this standard to improve successful detection of lies and fraudulent claims. The developer will benefit from having access to novel emotion detection technologies and working with Affective Media. The project is suitable to a student will good programming abilities and an interest or knowledge of speech analysis.owledge of speech analysis.

Emotional intelligence is the ability to use your emotions to help you solve problems and live a more effective life. There is a lot of interest currently in emotional intelligence from commerce as a solution to management and team-working, however more than that emotional intelligence can reduce stress and improve personal wellbeing. The project will look at how we can use emotional intelligence to improve the state of mind of users, by for example interaction with virtual characters, emotive sounds and images, empathetic questions and biofeedback. Suitable to those with interest in psychology.
 

Emotional technologies are in their infancy. Currently there is considerable interest both academically and commercially in developing emotionally intelligent and responsive systems. Unfortunately unlike speech recognition there are no standard databases of emotive speech examples on which to train and compare system performance. Affective Media has compiled its own database of emotive speech however there are problems in attempting to induce people into emotional states or using actors to create emotions. This project will allow the student flexibility in designing an environment in which to make human participants highly emotional and thus allowing recording of good quality emotive examples. In addition to requiring good programming skills the project could be suitable to those with interest in psychology and social behaviour.
 

Affective Media has found that regional accent has an affect on the emotional characteristics of speech. In attempting to train an automated emotion recognition systems it is necessary to find baselines of emotion for the human population. However some accents such as ‘the Midlands’ can be detected a more downbeat and bored whereas others such as ‘Glasgow’ detected as more aggressive. The project will consider the link between the perceived emotional state of regional accents and their affect on achieving automated emotion recognition systems. In addition to requiring good programming skills the project could be suitable to those with interest in psychology and speech analysis.
 

We have presented a range of applications for emotionally intelligent systems including gaming, mobile comms, robotics and toys etc. If however you have an interesting and novel idea for emotional technology then please feel free to contact Affective Media. You may want to develop wedding rings which when one is rubbed the other heats up to allow married couples to show each other affection every when miles apart. You may want to develop mobile phones which you can squeeze if angry and the recipients mobile turns red. We may well be able to configure your idea into a project to satisfy your interests.
 

In the last few years we have seen the rise of ontologies in informatics, particularly in the realm of bioinformatics. We are now at the point where we should be able to move beyond the description of biological entities using ontological terms to statistical analysis of groups of biological entities according to their ontological attributes.

The project will be to create software, and associated database tables, that will cluster entities according to their attributes where the attributes are terms or nodes in one or more ontologies (the ontologies may be arranged as trees or as directed acyclic graphs). The algorithm should be able to find clusters in the space of all ontologies or in the space defined by arbitrary subsets of ontologies. The algorithm should return distance or probability measures that describe the significance of the clusters. The algorithm should calculate the weights given to probabilities or distances between nodes as roughly proportional to the distance from the root.

The student will have access to expert molecular biologists for the purposes of testing and validation of the software.

Languages: Java, Perl, or C

Public pathway databases are growing rapidly and this profusion of data is allowing researchers to develop analytical tools that can be used to generative comparative and predictive data.

The project is to create an algorithm and associated software that could be used to find similar networks given one network where some fraction of the nodes are described with terms from one or more tree or DAG ontologies. The algorithm should be able to evaluate similarities using all ontologies or an arbitrary subset of ontologies. The algorithm should also be able find similarities and return meaningful statistical measures of similarity.

The student will have access to expert molecular biologists for the purposes of testing and validation of the software.

Languages: Java, C, or Perl

Many public databases contain sequence information as well as reliable human annotation of protein sequence, and much of this annotation is in the form of ontology terms. One should be able to automatically match protein sequences to Hidden Markov Models constructed from public domain “motif” databases and correlate these matches to the corresponding annotations. The aim is to assign motifs, or groups of motifs, to positions on various ontological graphs and to use these assignments to automatically classify protein sequences.

Interesting problems should arise as motifs will be assigned to different positions on the ontological graphs – the solution may be through the use of measures of statistical significance.

Languages: Java or Perl

At present, Guitarmaster cannot determine the difference between a note which is actually played on the instrument by the user, and a harmonic (or “upper partial”) of that frequency, also present at significant amplitudes in the sound spectrum. This is not a problem when transcribing single-note melodies, but it becomes a problem when transcribing chords.

The current approach is to simply ignore all frequencies which are multiples of lower frequencies in the spectrum, that is, to ignore all frequencies which might be harmonics of lower notes in the chord, and therefore which might not have been played by the user. This has the effect of reducing the number of notes detected as having been played in a chord, usually down to the key constituent notes of a chord (root, fifth, third etc).

In order to produce realistic voicings of these chords, Guitarmaster then applies a rulebase to “put back” some of the notes which were deleted as potential harmonics, but which were probably played by the user when they played the chord (e.g. in a chord of G major, the octave of the low G would probably be contained in the chord voicing actually played, but would have been deleted by Guitarmaster as the first harmonic of low G).

In certain cases, the deletion of harmonics results in an incorrect transcription of the chord name – for example if the dominant seventh degree of the scale occurs in the second or third octave, it may be deleted as a harmonic. Since cannot be inferred that that note was played by the user, the name of the chord is therefore incorrectly rendered as G major instead of G7.

The task of this project is to devise a means of reliably determining whether a note is in fact a “played” note, i.e. a “fundamental”, or a harmonic of a played note. This will greatly increase the accuracy of chord transcription by the Guitarmaster software.

The way Guitarmaster is currently configured, the sharp “attack” (amplitude peak) at the start of a played note (which is characteristic of the guitar’s sound envelope), is used as a means of detecting the point in time where a note or chord “starts”. The decay of the note’s amplitude below a certain level in turn determines where that note “ends”.

This approach works perfectly well for single-note melodies played clearly or for strummed sequences of chords. However, it does not work well for so-called “finger-style” playing, that is, where the characteristic “attack” is not so pronounced, and where one note continues to sound (decay) while another new note is played over the top of the first one, and so on and so forth. An example might be an arpeggiated chord where the notes are left to ring.

We would like to investigate the possibility of developing an alternative “frequency tracking” approach to the detection of note durations, so that this was not dependent solely on the attack characteristic of the envelope, and so that note attacks and decays could overlap with one another. The current system does not allow for overlapping attacks and decays.

IIn some respects, the ability to accurately transcribe pre-recorded guitar music is contingent on the successful implementation of 2.2 above. However, there are two other issues fundamental to this development, namely the detection of the “tuning” of the recorded music (how far it deviates from concert pitch - A = 440 Hz), and detection of the tempo of the recorded piece. The automated detection of these two parameters is the key task of this third project.

It should be a relatively simple task to modify Guitarmaster to work with monophonic instruments such as brass and woodwind instruments, or to work with bass guitar. Essentially, this involves certain aspects of the work required under point 2.2, that is, finding another means of detecting the start and end of notes other than using the characteristic “attack” of the guitar sound envelope. This would be an interesting and highly feasible research project, providing a very high probability of success.

Currently, Guitarmaster is a stand-alone application which produces its output as text (.txt) files (for guitar tablature) and as MIDI (.mid) files for music. Guitarmaster can then be used to launch a music notation application which will then load the MIDI file automatically in order to display it as notation.

However, if the user’s primary requirement is to display their guitar music as standard notation, the above procedure can become cumbersome. A more efficient solution might be to reconfigure Guitarmaster as a third-party “Plug-In” which could be accessed directly from a notation application such as Sibelius or Cubase VST. An interesting and self-contained research project would be represented by an investigation of the technical issues surrounding the conversion of Guitarmaster to a Sibelius or Cubase plug-in. This would involve researching the proprietary application-programme interface (API) used by these applications and determining how to encapsulate Guitarmaster within this framework.

It has been suggested that, if the issues noted under 2.2 above could be satisfactorily resolved, it might be possible to produce a product based on the Guitarmaster technology which could transcribe not just a vocal melody but the lyrics as well, so that a user could physically sing into a microphone and have the software produce a vocal lead sheet containing melody line and lyrics together. This is an advanced project, and would be dependent upon other issues referred to above already having been overcome. It would involve research into how well voice-recognition software operates with sung words, and how one might combine such software with the modules within Guitarmaster to create a hybrid which could achieve the desired resul

A local company, Verbalis, carries out German to English machine translation for a German client. The raw output is post-edited by another company engaged by the client. A substantial parallel corpus now exists of the raw and post-edited translations (and the original German). The aim of the project is to investigate methods of comparing raw and post-edited translations, to derive guidance for improving the raw translations.
The problem of comparing raw and post-edited translations can be couched as a paraphrasing task. A substantial literature on authomatic paraphrasing exists, which this project can rely on (e.g., Barzilay and Lee 2003; see also ACL 2003 Paraphrasing Workshop). Also the literature on statistical machine translation is relevant (starting with Brown et al. 1993; see also NIST MT activity). Existing software tools will be usable or adaptable for much of this activity.

Resources Required: A range of natural language evaluation software

Degree of Difficulty: medium to hard

Background Needed: Experience of natural language software; ideally an interest in machine translation

References:

ACL 2003 Paraphrasing Workshop: http://nlp.nagaokaut.ac.jp/IWP2003/

Barzilay, Regina and Lillian Lee (2003). Learning to Paraphrase: An Unsupervised Approach Using Multiple-Sequence Alignment, Proceedings of HLT-NAACL 2003, pp. 16-23.

Brown, P., S. Della Pietra, V. Della Pietra, and R. Mercer (1993). The Mathematics of Statistical Machine Translation: Parameter Estimation, Computational Linguistics, 19(2).

NIST MT activity: http://www.nist.gov/speech/tests/mt/index.htm

This environment should be based on the existing SeeTrack software.

Discipline:
Artificial intelligence, sensor fusion

Institution:
SeeByte Ltd

Type:
Engineering research

Supervisor:
Dr. Kelvin Hamilton

Project Description:
SeeByte is heavily involved with autonomous (no human intervention) underwater robotics for survey, intervention and docking. These tasks require mission management technology that is currently at or beyond the state of the art. SeeByte is currently developing suitable algorithms.
In order to perform initial evaluation of these mission management algorithms SeeByte needs to use a simulated environment, before moving to real vehicles in test tanks and, ultimately, the deep ocean.
SeeByte has technology for fusing multiple sensors into single environment, similar to a ‘world model’ as used in many artificial intelligence strategies.
You will be required to modify this technology to allow the use of mobile ‘agents’ and other virtual objects within this environment. Virtual sensors, currents, tides, and other underwater characteristics must be created.

Skills required:
C++, Simulation, AI

To apply contact:
Dr. Kelvin Hamilton
Development Manager (Robotics)

SeeByte Ltd.
Canaan Court
6A Canaan Lane
Edinburgh EH10 4SY
Scotland UK
Email: Kelvin.hamilton@seebyte.com
Web: www.seebyte.com
Tel: 0131 447 4200
Fax: 0131 447 4911

Discipline:
Artificial intelligence, sensor fusion

Institution:
SeeByte Ltd

Type:
Engineering research

Supervisor:
Dr. Kelvin Hamilton

Project Description:
SeeByte is heavily involved with autonomous (no human intervention) underwater robotics for survey, intervention and docking. These tasks require mission management technology that is currently at or beyond the state of the art. SeeByte is currently developing suitable algorithms.
In order to perform initial evaluation of these mission management algorithms SeeByte needs to use a simulated environment, before moving to real vehicles in test tanks and, ultimately, the deep ocean.
SeeByte has the use of a robotics submersible (see www.ece.eps.hw.ac.uk/research/oceans and follow the RAUVER link) that it wishes to use as an agent in a simulated environment. Making this model will involve data gathering trips in our wave tank and local reservoirs, before building the model from this date.

Skills required:
C++, Mathematical Modelling, AI

To apply contact:
Dr. Kelvin Hamilton
Development Manager (Robotics)

SeeByte Ltd.
Canaan Court
6A Canaan Lane
Edinburgh EH10 4SY
Scotland UK
Email: Kelvin.hamilton@seebyte.com
Web: www.seebyte.com
Tel: 0131 447 4200
Fax: 0131 447 4911

Affective Media are considered leaders in the development of emotion detection technologies. They are working with the mobile and animation industry to create emotion tracking systems for automated character animation, with the call centre and sales sectors for customer satisfaction and agent performance measurement, and with the automotive industry on improving car safety. They are one of the 2003 Scottish spin out company of the year award winners.
 

 
Cognia Corporation is a developer and distributor of information products and services that facilitate the use of biological and chemical information. Cognia's information solutions help pharmaceutical and biotechnology companies cost-effectively accelerate discovery and research processes. 

Cognia's business/product model was developed through experience and use with over one hundred Pharmaceutical, Biotechnology and Research Centers . Customers like Merck & Co., GlaxoSmithKline, Aventis, Schering-Plough Research Institute, and research groups at academic institutes such as Harvard Medical School , The Rockefeller University and The Whitehead Institute.
 
Our core product, Cognia Molecular™, enables industry and academic users alike to integrate, manage, and utilize biological and chemical information from many formats and sources in support of drug discovery and basic research. Cognia Molecular can be augmented with Cognia's value-added content products and services, as well as most datasets and systems our customers may already be using.
 
Cognia also distributes content products such as the gene regulation database products of BIOBASE GmbH and the mass spectral and anti-microbial chemical database products of John Wiley & Sons.
 

RoboSens Ltd of 75 Montpelier Park, Edinburgh, are the developers and sole publishers of Guitarmaster V 2.0 for Windows, a low-cost, software-only notation aid for electric guitarists.
 
The software allows the user to plug the output from an electric guitar directly into the sound card of a standard PC and produce music notation, guitar tablature and MIDI* files for chord sequences and single-note melodies directly from the audio signal produced by the guitar. No additional hardware is needed except for a simple connecting lead which RoboSens can supply.
 
The software is currently on sale on the Internet at www.guitar-master.co.uk, where further information on the product and the company can also be found. A demonstration version of the program can also be downloaded free of charge.
 
The software makes use of FFT and band-pass filtering techniques to convert the raw audio data from the time domain to the frequency domain.
 
RoboSens Ltd are currently seeking to further develop Guitarmaster technically, and are working in partnership with Nick Wright and Mike Clouser at the Informatics Department of Edinburgh University and the Edinburgh-Stanford Link. Students of informatics are being invited to collaborate productively on this innovative and interdisciplinary project by tackling small, self-contained research projects relating to the functionality of the Guitarmaster product. A number of these projects are discussed in the previous abstracts.
 w.