Lyra is Google's neural speech codec, built to deliver intelligible voice at extremely low bitrates — around 3 kbps and below, far past the point where classical codecs collapse into noise. Instead of hand-designed signal processing, it uses a machine-learned model: the encoder extracts a compact set of speech features and a neural generator reconstructs natural-sounding speech from them at the decoder. Its target use is keeping calls connected on very poor or congested networks where even Opus's lowest modes struggle. Lyra's later versions are built on SoundStream-style techniques, and it represents, alongside EnCodec and SoundStream, the shift from traditional to neural audio coding for the hardest low-bitrate cases.
The book · Volume 5 of 7
Audio for Video: A Complete Guide to Sound in Video Products: From Jitter Buffers and Lip Sync to Atmos, Quality Metrics, and Neural Codecs
Lyra shows that intelligible voice can survive at around 3 kbps, which changes what is possible on poor networks. Volume 5 places Lyra among neural audio codecs: SoundStream and EnCodec, how they compare with Opus at low bitrates, quality metrics for judging them and what they cost to run. Written by Nikolay Sapunov, CEO at Fora Soft.
Considering Lyra or another neural codec for voice on very weak networks?
We will test Lyra against Opus on your network conditions and devices and give you a go or no-go with quality and CPU numbers. 250+ video products shipped since 2005.
