Audio paths in production telephony platforms, measured rather than guessed.
Voice over narrow, high-latency links whose capacity changes by the second.
Ideas tested against real encoders and modelled links before anyone builds them.
Every voice network loses packets. The question is what the listener hears.
Dysonsphere Limited is a UK engineering company working on the media layer of real-time voice. Our work starts where the network stops cooperating: packet loss on congested access links, bursts of loss during handovers, and satellite links that fade from broadband to a few hundred bits per second.
We publish what we find. Our papers explain each technique, measure it against real encoders and libraries, and come with reference implementations that engineers can use in their own systems.
See the publicationsAnywhere voice has to cross a network that is lossy, narrow or both.
High-volume SIP and WebRTC platforms where a small gain in loss resilience shows up on every call.
Crews and vehicles on mobile and satellite links, where coverage changes along the route.
Offshore and remote installations that depend on narrowband satellite terminals for voice.
Configurations with FEC at full bandwidthVersion-aware keep-FEC, across libopus generations
Speech-quality gain from keeping FEC onWideband PESQ at 5–20 % random loss, libopus 1.1
Lowest rung of the voice ladderCodec2 700C, 1.3 kbps on air with headers
Audio lost in a deep satellite fadevs 35 % with Opus alone, in simulation
Technical papers by Dysonsphere Limited, free to download as PDF.
How the keep-fec-enabled technique in FreeSWITCH mod_opus works, what a decade of libopus changes did to it, and a version-aware reference implementation verified against real encoders.
Download PDFSwitching mid-call between Opus and Codec2 on a fading satellite link, from 25 kbps down to 1.3 kbps, with a simulation using real encoders over a modelled GEO link.
Download PDFBoth documents © 2026 Dysonsphere Limited. All rights reserved. Product names are trademarks of their respective owners.
Questions about the papers, the reference code, or a voice-quality problem on your own network? Send a message and we will reply by email.
