Transfer Learning for Audio Processing: A Review of Recent Advances

Transfer learning (TL) is a vital method that should be utilized in order to generalize models that were established for one area or work to other settings or activities. For instance, an audio model that was trained on one language may be used to recognize speech in another language with very little or even no additional re-training data required. TL, which is typically studied under the category of "model adaptation," has a significant connection to multi-task learning (also known as cross-lingual vs. multilingual learning). TL is made substantially easier and more successful when high-level abstract characteristics are learnt by deep models. This 'transfer' can be accomplished not just between data distributions and data types, but also between model structures (for example, shallow nets and deep nets), or even across model types (for example, Bayesian models and neural models). This review article gives an overview of some recent noteworthy research that has been undertaken in this field, most notably for speech and language processing. The research that is covered in this review article was conducted in this area. In addition, we emphasize the possibilities of this fascinating field of research and show some discoveries from our team.