Mission Brief
Project Overview
VAANI analyzes short speech clips and returns a Human, AI, or Inconclusive verdict with the evidence laid out: per-model scores, acoustic features, comparable reference clips, and a reliability report. Two independent signals run on every clip: a fusion head trained on Wav2Vec2 embeddings plus pitch, spectral, and zero-crossing features, and the published Spectra-AASIST3 anti-spoofing model as a second opinion. A disclosed rule table combines them, and when the signals disagree, scores sit in an ambiguous band, or fusion entropy runs high, the API answers Inconclusive rather than forcing a guess. Reference retrieval finds the closest clips in a held-out index by cosine similarity, and a deterministic explanation engine produces the report text, so identical input always yields identical output.