I am currently working on blind source separation using sparse hypotheses and convolutive mixtures. For my project, I have compared three different methods and calculated the Signal-to-Distortion Ratio (SDR), Signal-to-Interference Ratio (SIR), and Signal-to-Artifact Ratio (SAR). Now, I would like to know what I can do next and if there are any other metrics or methods available for further comparison. Thank you...
Some more details would be helpful in better answering your question. However, here are a few suggestions anyways:
You could experiment with different configurations of the algorithms you have implemented, such as varying the sparsity levels, filter lengths, or other hyperparameters. And then test the same metrics that you have tested so far.
Test using some other evaluation metrics such as:
- Perceptual Evaluation of Speech Quality (PESQ): This metric evaluates the quality of speech signals, taking into account human perception. Keep in mind that it only really works with speech signals and nothing else.
- Short-Time Objective Intelligibility (STOI): This measures the intelligibility of speech signals, which can be particularly relevant for applications such as hearing aids or telecommunication systems which is what I assume this project is for. Again this one only really works well with speech signals.
- Global Normalized Source to Distortion Ratio (GNSDR): This is a variant of the SDR metric that takes into account the energy of the original sources. You might find it more useful than SDR for certain cases.
You could test them on real life datasets like we generally do in ML. You might find these helpful:
- Common Voice This is an open-source dataset of human voices that is designed to help train speech-to-text engines. It contains thousands of hours of audio from people speaking in various languages.
- Voxceleb This dataset contains over 1 million short audio clips of celebrities speaking. It is often used for speaker recognition and verification tasks but can also be handy for testing speech separation algorithms.
- UrbanSound8K This dataset contains 8,732 labeled sound clips of urban sounds, such as sirens, car horns, and jackhammers. It is often used for audio classification tasks. You could mix it with the other two to really test the robustness of your algorithms.