Eigen
Problem?
Voice agents look good with clean test audio and break when real calls introduce street noise, side talk, and mumbles. Transcription errors rise, interruptions fire at the wrong time, and the rest of the agent works with the wrong context.
What did I build?
I helped ship Eigen, a voice isolation model that runs in under 10 ms. I worked across the model runtime, developer platform and SDK layer so developers could add it to existing stack. The SDK is available for direct Python audio processing and common Voice AI integrations including Pipecat and Livekit.
What changed?
Eigen now processes 100K+ minutes of audio every day and reduces word error rate by 57% on the industry leading ASRs. The benchmark evaluation were performed through custom fork of Open ASR Leaderboard's methodology.