Optimization of the FARGAN Model for Speech Compression : Exploring Different Frame Partitions

Sammanfattning

The field of speech compression is actively changing with the currently evolving field of Artificial Intelligence (AI), with multiple AI speech compression models being developed at a fast pace. With the goal to raise end-to-end quality of speech transmission and storage while maintaining low bitrate, low model size and low computational complexity, it is important that many different optimizations are explored in order to develop an efficient model. In this Master’s thesis project, a state-of-the-art speech synthesis model Framewise Autoregressive Generative Adversarial Network (FARGAN) was adjusted and evaluated, in order to explore new versions of the model and thereby advancing research in the area. A number of adjustments were tested and the three most interesting ones were evaluated using three objective evaluation models: Perceptual Evaluation of Speech Quality (PESQ), WARP-Q and Perceptual Objective Listening Quality Analysis (POLQA) as well as one subjective evaluation method: Multiple Stimuli with Hidden Reference and Anchor (MUSHRA). The three adjustments made were changes to the size of the subframes which the FARGAN model synthesizes speech from, which in turn affected the number of weights of the model. This kind of adjustment does not seem to have been explored in previous work. The results showed that we were able to recreate the original FARGAN model in terms of quality, and that our new, adjusted models produced lower quality speech than the original one. However, depending on how an application might weigh importance of quality, model size and computational complexity, it could be worth exploring two of our new, adjusted models, one of them being smaller in model size and the other being less computationally expensive during inference.