Google's Lyra v0.0.2 Speech Codec Gets Rid Of The Binary Blob

Earlier this year Google announced the Lyra voice codec that could work with AV1 video for video chats over 56kbps modems. Google is today shipping its newest Lyra version.

Lyra is a very low bit-rate codec for speech compression designed for WebRTC usage and other online chat purposes. Lyra can operate at 3kbps thanks to its machine learning and other design features. Following the February announcement, in April was the initial open-source code for Lyra. That initial v0.0.1 release has now been succeeded by version 0.0.2.

With Lyra 0.0.2, the project is now fully open-source. Formerly nearly all of the code was published except it still relied on a binary-only shared module, sparse_matmul.so. That sparse matrix multiplication library was co-developed by Google and DeepMind but initially not open-source and just shipped as a pre-built binary. That's now been addressed with their sparse_matmul implementation now being part of the source tree.

Lyra 0.0.2 is now fully open-source without needing to use any pre-compiled dynamic libraries, which also makes it possible to port Lyra now to more platforms. This also now allows Lyra to be built with the GCC code compiler on Linux.

More details on the new Lyra 0.0.2 speech codec release and source downloads can be obtained via GitHub. The Lyra code is available under an Apache 2.0 license.