Teams Voice Bridge
Microsoft’s real-time Teams media stack is built for Windows: native DLLs, a .NET SDK, and no Linux runtime. Instead of adding a Windows sidecar to a Linux system, I reverse-engineered the media protocol and rebuilt the bridge in Python. The bot can join a live meeting, receive audio, and speak back from Linux.
- Role
- Engineer
- Period
- 2025 to 2026
- Status
- Production
How the system fits together.
The calls that shaped it.
- 01
Stay Linux-native
Official paths meant the .NET media SDK, a Windows sidecar, or a third-party meeting bot — so I chose the harder route: understand the Teams wire protocol well enough to implement it directly.
- 02
Take Microsoft’s binaries apart
Decompiled the native and .NET DLLs in Ghidra and ILSpy, live-debugged the Teams client with x64dbg and Frida, and read decrypted traffic in Wireshark — wiring the decompiler into an LLM over MCP to interrogate thousands of lines and get through dead ends faster.
- 03
Rebuild it on Linux
From-scratch Python implementations of Microsoft’s binary framing and SOAP-over-the-wire protocols, an ICE agent to negotiate the UDP media path, SRTP key derivation, and a fix for two hidden four-second timers that were silently killing every connection.
- 04
It joins, and it talks
The bridge joins a real meeting, plays audio people can hear, and captures returning speech byte-for-byte after decryption — and for the codec path I drove Microsoft’s native decoder from Python, validated against captured frames.
- 05
Disciplined, and never blocked
Versioned every experiment, kept real captured calls as test fixtures, and refused to ship guesses for parts I couldn’t yet name — while shipping a simpler working voice agent alongside so nothing waited on the hard path.
The interesting work isn't the stack. It's the boundaries.