Project 1 – Improving the performance of deep learning based speaker recognition systems

Speaker recognition can be applied to a number of person authentication applications, including credit card verification, over-the-phone secure access in call centres, suspect identification by voice. Recent studies have found that the current state-of-the-art speaker verification systems, a significant amount of speech is required for development, enrolment and verification. It was also found that when speaker recognition system is developed using one database (out-domain database) and evaluated using another database (in-domain database), it significantly affects the speaker recognition performance.

In order to ensure the wide spread deployment of speaker verification technology in many practical situations, three major challenges must be faced:

  1. The amount of development data required to design state-of-the-art speaker verification systems must be reduced as it hard to collect significant amount of data.
  2. The duration of speech required to train speaker models as well the duration of testing utterances that are needed to be spoken by users must be reduced significantly.
  3. The mismatch between development data and evaluation data must be compensated.

In recent times, deep learning based approaches have become state-of-the-art approach, and in this proposal, deep learning based speaker recognition systems will be investigated with short utterance development and evaluation and novel approaches will be proposed to mitigate the short utterance issues. Subsequently, deep learning based speaker recognition systems performance will be also analysed when deep learning speaker recognition system is trained using out-domain database and evaluated using in-domain databased, and novel domain mismatch compensation approaches will be proposed to improve the performance of speaker recognition systems.


  1. Deep learning based speaker recognition systems will be developed.
  2. A number of novel techniques will be proposed to improve the accuracy of deep learning based systems.

GitHub link -

Project 2 – Forecasting of a solar photovoltaic power generation using deep learning approaches

Nowadays home owners and medium and large size investors are trying to have their own solar power plant. The field of PV forecasting has also been evolving rapidly in the recent years. There are different types of PV forecasts for different uses and applications. The forecasting duration ranges from short term period of hourly forecasting to long term period of monthly forecasting. The small scale PV owners will get benefit from short term forecasting whereas large scale PV owners will get benefit from long term forecasting.

Several machine learning techniques such as ANN, SVM, multivariate linear regression (MLR) have been studied by researchers to predict the daily solar power generation. As weather and solar irradiance predictions are complex, the existing machine learning techniques failed to achieve higher accuracy.

In recent times, deep learning approaches have become the state-of-the-art approach in speech processing, computer vision and pattern recognition applications. However, these new approaches have not been applied to solar irradiance and solar PV power generation forecasting yet. These state- of-the-art approaches will be studied with solar forecasting and based on our study, the existing network will be modified to achieve state-of-the-art performance.

In Sri Lanka, at the moment we don’t have proper tool(s) to assess a proposed solar PV site using local parameters. Based on this research, a tool will be developed and will be used to assess any site properly.


  1. A state-of-the-art LSTM recurrent network based prediction model to accurately predict the solar PV power generation.
  2. The medium and large scale power generation companies can use this software package as a tool to assess any site and decide whether that place can be used to build a Solar PV plant.
  3. The small scale owners such as home owners can use this software package to predict when they will have power shortage so that they can plan the backup power generation.

GitHub link -