MSc Project Specification – 2015/2016
Project Specification:
Project Title: Automatic Activity Recognition from Acoustic Signal
Student Name: Wei Song (Ve)
Supervisor: PJ
Background. (Please include a general scene-setting overview of the project – targeted at the non- specialist)
During recent years, there has been a huge increase of the amount of various types of multimedia data (audio, speech, music, and video) available in digital format. This has created a large demand for development of automatic intelligent tools that could organise and search through this data, or extract knowledge from this data.
This project is concerned with recognition of human activities from a short piece of audio recording. The aim is to design and develop an activity recognition system, exploring different type of feature representation and acoustic modelling. The system should be developed using Python / Matlab and would also involve the use of the HTK toolkit.
Expected Outcomes. (Please include a specification for the expected outcomes of this project when undertaken by an average student. e.g. ‘The aim of this project is to design and ….’)
Represent the audio signal as a sequence of features, e.g., Mel-frequency cepstral coefficients (MFCCs).
(I) Develop and evaluate a conventional activity recognition system based on using the Gaussian Mixture Model (GMM) – perform the training of the model for each activity with corresponding data.
(II) Develop and evaluate a GMM-UBM system – build a ‘general’ GMM based on data from all activities and then employ maximum a-posteriori (MAP) adaptation using activity-specific data to obtain the model of each activity.
(III) Develop and evaluate a GMM-SVM system – this is based on representing an utterance of recording as a ‘supervector’ consisting of the means of the adapted GMM components and then using support vector machine (SVM) for classification.
(IV) Develop and evaluate an ‘i-vector’-based system – this is based on using the ‘supervector’ representation but then transforming this into an i-vector with reduced dimensionality for classification.
To design and perform experimental evaluations using leave-one-out procedure on a given corpus of audio recordings.
Fallback and Rebuild Position. (Students sometimes have difficulty in delivering the stated outcomes. Using bullet points, please list a suitable set of minimal target objectives.)
Develop and evaluate the system denoted as (I) and (II) above.
Basic analysis of the results and discussion.
Enhancement Position. (It is anticipated that many students will achieve the expected outcomes stated above. Using bullet points, please list a suitable set of achievable enhancement objectives.)
In addition to the Fallback Position, develop and evaluate the system denoted as (III) and (IV) above.
Perform a comprehensive literature review of recent research in the project area.
Thorough analysis of the results, discussion and consequent improvements to the
system.