lyra is a neural audio codec designed to deliver intelligible, natural-sounding speech at extremely low bitrates, making real-time communication viable on constrained networks. It replaces hand-engineered codecs with learned models that capture speech characteristics more efficiently and reconstruct waveforms with a neural vocoder. The system targets mobile-class hardware, balancing latency and quality so it can run in real-time on phones. Its architecture is resilient to packet loss and jitter through framing strategies and error concealment, helping conversations remain understandable under adverse conditions. The codebase includes encoder and decoder components, along with tools for data preparation and evaluation. By pushing bitrates down to just a few kilobits per second while retaining quality, lyra expands access to voice calls where bandwidth is scarce.
Features
- Neural speech coding at very low bitrates suitable for real time
- Encoder–decoder pipeline with a neural vocoder back end
- Packet loss concealment and jitter tolerance for unstable networks
- Mobile-friendly performance and latency characteristics
- Tools for training data prep and quality evaluation
- Focus on naturalness and intelligibility for conversational speech
Project Samples
