How Discord Rebuilt Its Voice Engine for Sub-50ms Latency

Episode notes

In this episode of The CTO Podcast, Lucas and Luna dive into how Discord achieved sub-50 millisecond voice latency across millions of concurrent users. They break down the specific architectural changes Discord made: switching from Opus to a custom codec called Siren, rewriting their audio processing pipeline in Rust, and deploying edge relays in over 300 locations worldwide. The discussion covers why Discord chose to build its own transport protocol over WebRTC, how they handle packet loss with forward error correction, and the trade-offs between CPU usage and bandwidth. Lucas explains the key metric that guided their redesign - the 99th percentile one-way voice latency - and how they optimized for it without sacrificing audio quality. Luna challenges whether the effort was worth it given Discord's core use case for gamers, and Lucas argues that voice latency is the defining feature for real-time communication. The episode includes a brief donation segment near the end, seamlessly woven into the conversation about open-source tools and community support. Perfect for CTOs, engineering leaders, and anyone building real-time audio applications.

• How Airbnb Rebuilt Search for 8 Million Listings42 / 100
• How GitLab Built a Single Codebase for One Million CI Pipelines45 / 100
• How Slack Rebuilt Its Search Index for 10 Million Daily Queries37 / 100
• How Notion Rebuilt Its Sync Engine for Offline-First
• How Notion Rebuilt Its Block Engine for Hybrid Local-Sync