HEARSAY audio authentication, documented
GitHub

Systems and subsystems

The speech gate and the default-answer policy

Five cheap cues decide whether a file contains speech to judge. A file that fails skips fusion and is placed in a block strictly below every judged file, at the real end of the ranking, where an uncertain file belongs under the sponsor's cost.

In plain words

Our first deep probe scored pure silence at 0.99source: README.md, section orchestration-what-routing-changes synthetic. Under a cost where a false alarm counts 9.33source: README.md, section how-it-decides misses, a handful of silent or musical files at the top of the ranking would cost more than the entire fusion decision. The gate catches such files; the policy puts them where a "we don't know" belongs. The thresholds were set beyond the extreme value of every cue in the full test set, so the gate touches 0 of 1,671source: README.md, section the-eight-forensic-techniques test files and only ever catches silence, tones, static and noise.

What goes in and comes out

In: the prepared clip. Out: a speech_gate result whose score is always the neutral value (fusion never sees it), with features for the five cues, the reasons that fired and is_speech; and, from the policy, the final placement of every file's probability.

Where it lives

src/hearsay/detectors/speech_gate.py: analyze, the detector, and apply_default_answer. The runner applies the policy exactly once in final_score (pipeline.py, lines 656–668).

The five cues and where each is set

Any one cue firing gates the file. The test set's extremes that sit behind the thresholds: quietest file −34.6source: src/hearsay/detectors/speech_gate.py line 15 (GitHub; assumes main is public) dBFS, lowest voiced fraction 0.099source: src/hearsay/detectors/speech_gate.py line 15 (GitHub; assumes main is public), lowest loudness spread 6.0source: src/hearsay/detectors/speech_gate.py line 15 (GitHub; assumes main is public) dB, highest flatness 0.39source: src/hearsay/detectors/speech_gate.py line 15 (GitHub; assumes main is public) (speech_gate.py, lines 14–16). Clips shorter than half a second are zero-padded; pitch comes from YIN between 60 and 400source: src/hearsay/detectors/speech_gate.py line 76 (GitHub; assumes main is public) Hz.

The default-answer policy

Status

Built about 06:10 on Saturday; the block's formula was corrected at 08:30 when the pipeline turned out to score some real files as low as 0.0009source: docs/reports/2026-09-26_gate-and-drift.md, which the first version would have overlapped. The architecture document still shows the first formula.

Numbers

Gated: 0 of 1,671source: README.md, section the-eight-forensic-techniques test files; 31 holdout rows and 9 In-the-Wild rows in the ablation, costing 0.001source: README.md, section orchestration-what-routing-changes under the brief's cost on In-the-Wild. Placement is optimal under the brief's cost when the block's fake rate is below 80%source: README.md, section orchestration-what-routing-changes and under the sponsor code's reading when it is above 9.7%source: README.md, section orchestration-what-routing-changes; see the metric page.

Known limits

Rhythmic polyphonic music can pass the cues; the preflight and the high-score flag remain the backstop. The module's docstring says "a tenth" of frames voiced where the constant is five percent.

Tests that pin it

tests/test_speech_gate.py: silence, a chord, a tone and white noise are gated with the right reason; speech-like input passes; quiet real-level speech is not silence; contract and determinism; the policy pins gated files strictly below every scored file; ordering uses the weak signal and keys and puts failures at the bottom; only the gate flag gates; a data-gated test that real test clips pass. tests/test_pipeline.py pins that the map is applied once and that a decode failure sits below the block.

Sources

The gate and drift report; README, "The abstention path"; consult item 5.