Hi!, I'm

Sacha Arbonel

I build speech AI and video infra.

Recent writing

Introduces a dataset of Whisper non-speech hallucinations, explains why they happen, and shows BoH filtering with Silero VAD that cuts WER on noise.

27 Sep 2025

I work on ASR, WebRTC, and GPU runtimes.
I keep things simple and fast.

Want to ship faster?
Email me.
I reply in 24 hours.

Serving Whisper to 30+ customers at Stream 2k reqs/min at peak. p90 latency cut by 50%. 35-40x cheaper than OpenAI.

Multilingual noise dataset for false-speech.

Reinventing Speech-to-Text Transcriptions with Go and Whisper

"You have to build the transcriptions feature as good as Google and Zoom but we don't have a similar budget". That was a challenge we faced and solved with hundreds of coffee cups and incredible hair loss, our solution is cost-effective, open-source, and fits within our Go stack. Whisper.cpp is a lightweight implementation of Open AI’s Whisper speech-to-text model, that comes with Go bindings. This solution is leveraged by a lot of emerging startups and existing companies.