.png){kind=icon}
返回
LMCodec2: Ultra-low bit rate codec with causal multiple transformers
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2024.109960.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
In recent years, the bandwidth constraints in satellite Internet of Things (IoT) applications have spurred the development of novel methods for compressing transmitted speech. For satellite voice communications, it is essential to achieve high-quality codecs with a bit rate below 1 kbps, particularly for channels such as Beidou-3, which often operate under such limitations. Neural network-based vocoders have emerged as a promising solution within the AI community, offering high-fidelity audio compression. In this paper, we propose LMCodec2, a causal speech codec designed to operate across a range of bit rates while delivering high-quality audio at extremely low bit rates, specifically tailored for satellite voice transmission. LMCodec2 utilizes a Transformer-based language model to predict tokens frame by frame, achieving a 25 % reduction in bit rate without compromising decoded audio quality. Our experimental evaluations demonstrate that LMCodec2 produces high-quality decoded audio at 0.76 kbps and 1.15 kbps. Notably, at 0.76 kbps, LMCodec2 achieves a MUSHRA (Multi-Stimulus Test with Hidden Reference and Anchor) score that surpasses Encodec's performance at 1.5 kbps. Audio demonstrations, including real-world self-recorded speech datasets, are available at https://dingweipeng.github.io/JACK. github.io. LMCodec2 provides a new way of thinking to addressing the challenges of bandwidth-limited satellite voice communications.
Keyword:
End-to-end codec
VQ-VAE
GAN
Transformer model
Huffman coding
期刊
C
IF:
4.9
论文数:
6.7K
被引数:
1.3W
机构
引用论文
.png){kind=icon}
分享收藏 没有更多内容
没有更多内容