Google has released version 1.3 of the Lyra audio codec, aimed at achieving high-quality voice transmission under conditions of limited data bandwidth. The speech quality at bitrates of 3.2 kbps, 6 kbps, and 9.2 kbps when using the Lyra codec is roughly equivalent to bitrates of 10 kbps, 13 kbps, and 14 kbps with the Opus codec. To achieve this goal, Lyra utilizes a speech model based on a machine learning system that allows for the reconstruction of missing information based on typical speech characteristics, alongside conventional audio compression and signal transformation methods. The reference implementation is written in C++ and distributed under the Apache 2.0 license.
Unlike the radically redesigned Lyra 1.2 released in October, which was transitioned to a new neural network architecture, version 1.3 features an optimization of the machine learning model without architectural changes. In this new version, 8-bit integers are used for weight storage and arithmetic operations instead of 32-bit floating-point numbers, resulting in a 43% reduction in model size and a 20% speedup in model performance during testing on the Pixel 6 Pro smartphone. Speech quality has been maintained at the same level, but the data format has changed and is not compatible with previous releases.
Source: opennet.ru
