HEARSAY-IV: Speech Transcription You Can Argue With | Meta Hackathon 2026

HEARSAY-IV — Speech Transcription You Can Argue With. What if a speech recognition system could explain exactly why it chose a word, show the alternatives it rejected, and undo a correction without losing its original state? HEARSAY-IV is an experimental, non-neural speech transcription system built for the Meta Global AI Developer Hackathon 2026. Inspired by the original HEARSAY-II blackboard architecture, it combines classical speech recognition with a Saturation Fabric that makes word decisions inspectable. **What the demo shows:** - Real speech transcription with a word-by-word explanation panel. - A classical MFCC/HMM acoustic model, Kneser-Ney trigram language model, and 259,000-word lexicon. - Near-linear scaling: measured growth exponent improved from 3.84 to 1.07. - Processing up to 67,000 competing hypotheses per sentence, faster than real time. - Reversible corrections recorded as facts, with byte-identical undo. - A pre-registered experiment testing whether retaining dormant hypotheses improves recognition compared with pruning. **The experimental results are reported without cherry-picking.** Dormancy did not outperform pruning. On the evaluated test set, the classical decoder had a 64.3% word error rate, while the Fabric had an 85.5% word error rate. The experiment established a scaling improvement and an inspectable decision mechanism, not an accuracy breakthrough. One correction repaired 7 of 16 single-word errors in a separate test. Built using Muse Code with parallel development agents, verifiable tasks, independent review, and human-controlled acceptance decisions. Source code: https://github.com/Michael-A-Kuykendall/hearsay-four (Repository will become public after release checks.) LibriSpeech audio dataset: OpenSLR 12, CC BY 4.0. #HearsayIV #SpeechRecognition #MetaHackathon