Enhancing Speech Compression by Intra-Inter Broad Attention: The Case of IBACodec

Abstract

This study investigates the improvements in speech compression efficiency, particularly at low bitrates, by addressing the challenges of redundancy removal and enhancing context awareness. The research proposes IBACodec, a novel codec that leverages advanced attention mechanisms, such as the intra-inter broad transformer, and a dual-branch conformer for efficient redundancy elimination. Five core hypotheses were tested: the impact of context awareness, the effectiveness of dual-branch modeling, a comparative analysis with existing codecs, subjective evaluations, and objective metric performance. Results demonstrate that IBACodec outperforms traditional codecs like SoundStream and Opus in compression efficiency and quality at lower bitrates. Furthermore, subjective assessments reveal superior performance at low bitrates, while objective metrics such as ViSQOL, LLR, and CEP also confirm the codec's advantages. This research highlights the potential of IBACodec as a leading solution in speech compression, emphasizing the role of advanced machine learning techniques in enhancing codec performance.