What it does
A Whisper encoder produces the state and typed questions fan out over it. Choice, Score, and Noul all answer against the encoded clip, and no decoder ever runs.
Benchmarks & research
Prosodia
Every voice agent transcribes before it decides anything. Prosodia asks the audio directly, so a misheard word can no longer poison the judgment downstream.
Bookmark: Prosodia
Keep this in your collection.
Leave a noteWhat would you try with this?
: Prosodia
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
To start, collect short voice clips of spoken requests and list the choices your system must make. A developer can adapt this project to examine raw sound directly. Your app can then answer multiple questions at the same time without converting spoken words into written text first.
Keep in mind that this project is an early research test. It limits sound clips to thirty seconds, and its emotion detection struggles to spot feelings like sadness. Treat the design as an experimental idea to explore rather than a finished tool for everyday use.