The communication frameworks that power online voice and video calls (e.g., WebRTC) use codecs, compression techniques to efficiently transmit data. These codecs are designed to optimize quality while using less data and minimizing latency in real-time communication. Although modern video codecs have high performance, speech codecs can be less reliable. Recently, Google OSS’ed Lyra, a high-quality, very low-bitrate speech codec designed to enable voice communication on low-bandwidth networks. Lyra leverages generative models that synthesize a speech signal from features (log mel spectrograms) extracted every 40 ms and compressed for transmission at a bitrate of 3kbps.