Deep Neural Networks for End-to-End Optimized Speech Coding

Abstract

Modern compression algorithms are the result of years of research; industry standards such as MP3, JPEG, and G.722.1 required complex hand-engineered compression pipelines, often with much manual tuning involved on the part of the engineers who created them. Recently, deep neural networks have shown a sophisticated ability to learn directly from data, achieving incredible success over traditional hand-engineered features in many areas. Our aim is to extend these "deep learning" methods into the domain of compression. We present a novel deep neural network model and train it to optimize all the steps of a wideband speech-coding pipeline (compression, quantization, entropy coding, and decompression) end-to-end directly from raw speech data, no manual feature engineering necessary. In testing, our learned speech coder performs on par with or better than current standards at a variety of bitrates (~9kbps up to ~24kbps). It also runs in realtime on an Intel i7-4790K CPU.