CITISEN: A Deep Learning-Based Speech Signal-Processing Mobile Application

This study presents a deep learning-based speech signal-processing mobile application known as CITISEN. The CITISEN can perform three functions: enhancement (SE), model adaptation (MA), and background noise conversion (BNC), which allow to be used platform for utilizing evaluating SE models flexibly extend the address various environments users. For SE, downloads pretrained on cloud server then uses these effectively reduce components from prerecordings or instant recordings provided by When it encounters noisy signals with unknown speakers types, MA function allows improve performance effectively. A few audio files of unseen types are recorded uploaded adapt model. Finally, BNC, removes original using an mixes processed signal new noise. novel BNC evaluate under specific conditions, cover people’s tracks, provide entertainment. experimental results confirmed effectiveness MA, functions. Compared signals, enhanced achieved about 6% 33% improvements, respectively, in terms short-time objective intelligibility (STOI) perceptual evaluation quality (PESQ). With STOI PESQ could further improved approximately 11%, respectively. Note that method not limited ones described this replaced any method. experiment indicated converted backgrounds have close scene identification accuracy similar embeddings acoustic classification Therefore, proposed convert data augmentation when clean unavailable.