jevbooks

← All projects

audio-jevlike

alperiox/audio-jevlike

Audio-native Jev-shaped decision model from speech, no ASR.

What it is

Prosodia is an audio-native Jev-shaped decision model that takes speech as its state. It encodes audio once with a frozen Whisper encoder and answers typed questions in parallel as calibrated probability distributions, without ASR or generated text.

How it uses Jev

Jev makes decisions on a shared audio encoding. Typed questions (Choice/Score/Noul) branch over this state in parallel, returning calibrated distributions over emotion, sentiment, and acoustic properties. Question text and option sets are supplied at request time; results are used as probabilities in code.

Primitives:choicescorenoul

Technique worth stealing

Prefill-only readout with pointer mechanism: logits are inner products between branch vector and option embeddings, enabling dynamic candidate sets.

Try it

uv sync; uv run pytest; scripts/fetch_meld.sh; uv run python scripts/extract_features.py --encoder whisper; uv run python scripts/run_ablation.py ...

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.96
  • Shows a System One patternyes0.79
  • Handles uncertaintyno0.05
  • Measuredno0.07
  • Runnableyes0.63
  • Worth recommendingunclear0.56
  • Model replicayes0.93
  • Problem scopescore on a 0–2 scale0.72
  • About Jevyes0.97

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.