Google published Audio -Code Issue Lyra 1.3 , aimed at achieving a high quality of voice transmission in conditions of limited volume of transmitted information. The quality of speech on bitrates is 3.2 KBPS, 6 KBPS and 9.2 KBPS when using the Lyra codca approximately corresponds to the 10 KBPS, 13 KBPS and 14 KBPS beatings when using the OPUS codec. To solve the task, in addition to the usual methods of compressing sound and converting signals, a speech model based on the machine learning system is used in Lyra, which allows to recreate the missing information based on typical speech characteristics. The reference implementation of the code is written on C ++ and spreads under the license Apache 2.0.
, unlike the radically redesigned Lyra 1.2, which was proposed in October, transferred to the new architecture of the neural network, in version 1.3, the model of machine learning is optimized without architectural changes. In the new version for storing weights and performing arithmetic operations instead of 32-bit of the swimming, 8-bit integers are involved, which led to a reduction in the size of the model by 43% and the acceleration of the model by 20% when testing on the Pixel 6 Pro smartphone. At the same time, the quality of speech was preserved at the same level, but the format of the transmitted data has changed and is not compatible with past issues.