All work
Speech
Halcyon Audio
2024
Streaming speech-to-text tuned for noisy studio chatter, running under 200 ms end to end on a single consumer GPU.
Scroll to resolve
190 ms
End-to-end latency
Client
Halcyon Audio
Role
ML engineer, speech
Stack
Whisper fine-tune · CTranslate2 · WebRTC
Year
2024
The problem
Live captions for podcasts lagged three seconds and fell apart whenever two hosts talked at once.
What I built
I fine-tuned a speech model on 600 hours of overlapping studio audio and rebuilt the serving path around streaming chunks instead of full sentences.
The result
Captions now land in 190 ms, and word error rate on crosstalk dropped by a third.