EASAIER Status Report

March-April 2007

May 2007

Workpackage 1 – Project Management QMUL (Josh Reiss, QMUL)

QMUL has continued to work closely with all partners this period in supporting them in achieving their targets for the first 12 months of the EASAIER project. Particular support was needed in workpackages 2 and 3. In March, Josh Reiss went to Budapest to discuss issues with the workpackage leader. This resulted in further defining of the roles of partners NICE, QMUL and ALL in the overall speech and music retrieval system.

QMUL has also organized a preliminary agenda for the12 month meeting and review of EASAIER. It was concluded that the meeting would be over two days with the first day dedicated to coordination issues and run through of partner presentations and the second day being the actual review. The project coordinator will also review all partner presentations and make recommendations for revision where necessary.

The API has been updated to include data on cross media retrieval. The Software Modules document has also been updated. Chris Landone is currently working on a metadata document to support the development of the music ontology and retrieval system.

QMUL has also recruited Rouhua Zhou as a full time researcher for EASAIER.

Workpackage 2- Media Semantics and Ontologies

Since the meeting of the ontology group in January, DERI has released a testable version of the basic ontology to all partners in April 2007 and are working on the integration of the proposed modular ontology and music ontology from QMUL. Top level ontology and speech related ontology modules have been drafted and ontology relevant use cases are currently under evaluation.

DERI continues to work on a demo and presentation for the 12 month meeting as outlined below:

  • One demo will show the ontology editor they are developing. It will show how to build ontology, how to modify it, and how to populate it with data.
  • A second demo will demonstrate the benefits of having a semantic knowledge management system. It will show some example queries addressed to an RDF store, and how the reasoning works.

Workpackage 3 – Retrieval System

Activity:

Speech Recognition and Retrieval

ALL has started to design and build a multi layer language independent linguistic model which will work on top of the phoneme level recognition. The process uses language specific dictionaries with statistical data and morphological analysis to match reasonable word sequences to the recognized phoneme sequence. The output of the whole process is as set of probable word sequences or sentences. The indexing works on these words, and uses the phoneme level indexing only for words out of the dictionary. The retrieval will be performed on the index files. ALL is now focusing on the testing and performance tuning of the interaction of the phoneme level recognition layer and the linguistic layer.

ALL met with the project coordinator in March where they demonstrated the results.

Test corpuses

ALL have converted an English corpus, DARPA’s TIMIT speech database to their corpus format and have started to evaluate their algorithms in English. The have also converted a part of the US Supreme Court corpus for testing purposes.

In consultation with DIT it was realized that the noise on the broadcast quality recordings was actually emitted by a different source. As no methods are currently available for separating the noise, it was decided that noisy and multiple sound source parts of the corpus should be removed for later processing. ALL has developed special software to aid in the manual preparation of the corpus.

Workpackage 4 - Sound Object Representation

NICE have completed two research activities as part of their advanced pre-processing module for the speaker retrieval system: Outlier exclusion and Music Detection. Outlier Exclusion includes testing techniques for out detection and will test their impact on the Speaker recognition performance. Music Detection includes developing a unique feature for speech/music discrimination and compares the results. In addition NICE have worked on improving the performance of the speaker retrieval system. This would dramatically improve the real time, memory storage and accuracy of sea recognition performance.

Workpackage 5 – Enriched Access Tools DIT (Dan Barry)

In this period DIT has coordinated the drafting of the deliverable D5.1 Prototype of looping and marking modules. This was successfully submitted.

A number of enriched access tools are now ready for testing. These include:

  • Time Scale Modification - This has been submitted to SILOGIC for integration. It will be demonstrated at the annual review.
  • Sound Source Separation - This has also been submitted to SILOGIC. This will also be demonstrated at the annual review.
  • Noise Reduction - The work on the real-time version is not ready however a prototype off line version has been successfully implemented.

QMUL are working on a version for Video time stretching which should be ready for demonstration in the summer 2007.

Workpackage 6- Intelligent Interfaces (Luc Barthelemy and David Cher)

Web application interface

SILOGIC have successfully integrated some of the enhanced features to the EASAIER user interface. This will be demonstrated at the 12 month meeting. Draft mock ups have been circulated to all partners for comment.

Infrastructure

SILOGIC have set up a bi-processor machine that will be used to host development services including a database (DB2), BOCA storage, feature extractors, web services, ftp etc.

EASAIER QT- Application

Silogic have worked on the client QT application and released a first version of this to all partners in April for comment.

Workpackage 7- Evaluation and Benchmarking RSAMD (Celia Duffey and Joe Harrop

RSAMD have drafted a preliminary timetable for evaluation and testing of the EASAIER system. They have also started to plan a first meeting of the Expert User Advisory board which will be held in September 2007.

RSAMD have provided supplementary test bed material which has been uploaded onto the EASAIER ftp site located at QMUL.

Workpackage 8 - Dissemination and Exploitation

QMUL has drafted a business plan and has assembled documentation and references from all partners with respect to publications and dissemination. The Business plan will be further developed over the course of next year.