Can AI Hear You Sing?
Benchmarking real-time speech-to-text on music: six engines, four languages, one laptop. Guess who won? The LOCAL model.
danicuki.substack.com/p/can-…
The @rotkiapp is a nice app for tracking your crypto portfolio privately. Problem is that the free plan has some awful limitations and none of the subscription plans fit my use case: "Supporter plan" could be cheaper and actually has no extra features compared to free and the "Basic plan" is too expensive for a small user like me.
2/ How it works: mic → speech-to-text → a Director model (Gemma 4 on the laptop, or Gemini Flash-Lite) decides the next brushstroke → an image model paints over the previous canvas.
3/ When it knows the song, it paints ahead and holds: 11 of 12 lines revealed within 1 ms of being sung. When it doesn't, it listens and paints what it hears. The Napoli finale was the second path, on purpose.
4/ Open source (MIT): github.com/danicuki/sinestes… · sinestesia.art
Next: Warsaw, @googlecloud Builder Connect, in November. Grazie, Napoli.
25 years after living in Treviso, my first talk in Italian: @googledevs DevFest Napoli, singing "Volare" and the AI painted it live on the wall. Then a song it had never heard, written that week: Napoli, Napoli, Napoli.
Built with @googlegemma + @GeminiApp . Video (IT):
youtube.com/watch?v=DH0KZMx6…
This does not mean people do not care about privacy.
It means privacy usually loses when it requires sacrifice — and wins when it is packaged as convenience, protection, or financial upside.
9/10



