Wake-word, finally shipping.
After eight months of false starts with Vosk, Moonshine, and OpenWakeWord, here's the architecture that won — and why it runs offline at 4% CPU.
Product updates, engineering deep-dives, customer stories, and the occasional rant about latency. Updated weekly.
After eight months of false starts with Vosk, Moonshine, and OpenWakeWord, here's the architecture that won — and why it runs offline at 4% CPU.
The four dimensions of a workspace — domain, style, tone, instructions — and how InkSpoke picks the right one without you noticing.
We ran 1,000 hours of customer dictation through three engines. The winner isn't the one with the highest WER.
Tom's hands gave out in 2024. He's writing more code in 2026 than he ever did with two working wrists. Here's how he set up InkSpoke.
A walkthrough of running InkSpoke fully offline — including which models to download, how to verify in Little Snitch, and the latency you'll see.
From IBM's 'Voice Writer' to Dragon NaturallySpeaking to today's transformer-driven pipelines. What changed, and what stayed the same.
If transcription gets your team's names, acronyms, and product codes right, the rest of the polish is downhill.
One email a fortnight. Product news, voice-AI research notes, and the occasional customer story. No spam.