Arjun Pankajakshan, Ph.D Email & Phone Number

United Kingdom

United States

• Audio Deepfake Detection• Speaker verification/ identification

My research area is Computational Audio Analysis (Machine Listening) applied to urban, everyday, healthcare, bio-acoustic, and nature sounds. In particular, my work is focused on improving sound recognition using audio sequence encoding and audio sequence modelling-based methods. Key contributions are:• Developed CRNN-based Multi-task Learning models, incorporating sound activity and onset detection as auxiliary tasks, which improved sound event detection performance in UrbanSED and DCASE datasets.• Formulated self-attention mechanisms for sound event sequences, enabling analysis at the audio frame, segment, and event levels.• Developed a time-restricted self-attention-based SED model, showcasing the benefits of local attention context in modelling temporal relationships in sound event sequences.• Designed and implemented Transformer Encoder-based sound event detection models with adaptive and dynamic context selection, and a source-separation functionalities within the multi-head attention unit.

London, England, United Kingdom

• Developed a real-time bird audio detection system capable of classifying a large set of bird species, experimenting with up to 200 classes.• Utilised the Scaper audio library to generate synthetic audio samples to add more variations to the dataset.• Employed pre-trained audio models like PANNs, VGGish, and L3-Net for fine-tuning the bird audio model.

London, England, United Kingdom

• Replicated VITS-based TTS model results on LJ Speech and VCTK datasets.• Preprocessed the Spotify podcast dataset for TTS model development, involving data cleansing, segmentation, tokenization, normalization, formatting, silence removal, and audio feature extraction.• Explored challenges and variations in developing and evaluating a TTS model using the Spotify podcast dataset, including investigating different text and audio durations.• Explored spectrogram-based objective evaluation methods for TTS models.

India

• Contributed to bird data (image and video) collection and annotation.• Developed a fully convolutional model for bird audio detection that achieved comparable performance to the winning entry of the BAD challenge 2017.• Developed a Faster RCNN-CNN based two-stage image model to enhance bird-part localization in images. The first-stage model is exclusively trained using bird-part images, contributing to the refinement of the localization model.• Developed a pre-processing technique for temporally localizing bird activity frames in long-duration recorded video data by leveraging a pretrained bird image model.

Doctor Of Philosophy - Phd, Computational Audio Analysis, Deep Learning

Master Of Technology (Mtech), Communication Engineering & Signal Processing

Engineer’S Degree, Electronics And Communications Engineering, First Class

Matriculation, Science