Google has published the release of the Lyra 1.3 audio codec, aimed at achieving high voice quality in a limited amount of transmitted information. Speech quality at 3.2 kbps, 6 kbps, and 9.2 kbps bitrates using the Lyra codec roughly corresponds to 10 kbps, 13 kbps, and 14 kbps bit rates using the Opus codec. To solve this problem, in addition to the usual methods of sound compression and signal conversion, Lyra uses a speech model based on a machine learning system that allows you to recreate the missing information based on typical speech characteristics. The reference code implementation is written in C++ and distributed under the Apache 2.0 license.
In contrast to the radically redesigned release of Lyra 1.2 proposed in October, translated to a new neural network architecture, version 1.3 optimized the machine learning model without architectural changes. The new version uses 32-bit integers instead of 8-bit floating-point numbers to store weights and perform arithmetic, resulting in a 43% reduction in model size and a 20% speedup of model performance when tested on a Pixel 6 Pro smartphone. At the same time, the speech quality was maintained at the same level, but the format of the transmitted data has changed and is not compatible with previous releases.
Source: opennet.ru