The library

Rendered from docs/reports/2026-09-26_gate-and-drift.md; the markdown file is the record.

Non-speech gate, default-answer policy, and speaker-embedding drift

Sat Sep 26, 2026, 06:00–07:00. Follows architecture.md section 10, decision 4 (default answer for undetermined files) and the v4 brief's section 8.7 (rubric technique 6, speaker-embedding consistency). Both are contract detectors in src/hearsay/detectors/, registered by hearsay.detectors.engineered, exported in the M4 format by scripts/score_detector.py. Neither is a fusion column: the gate's score is constant and its decision lives in features["is_speech"]; drift's score is mild and, on this data, points the wrong way for a stacker (below).

1. The problem the gate solves

The deep detectors are trained on speech and score non-speech near 1.0: the public-data M1 probe scored pure silence 0.99. With a false alarm at 9.33× a miss, a file with no speech to judge belongs at the real end of the ranking, deliberately and reproducibly, with a reason the explanation report can print.

2. The gate (hearsay.detectors.speech_gate)

Speech, for the gate, is material that is

1. not silent: RMS above −70 dBFS (digital silence is −180; the quietest NSA test file is −34.6);
2. voiced: at least 5% of frames within 30 dB of the loudest have a YIN pitch in 65–390 Hz (NSA test minimum 0.099);
3. moving in pitch: std of log F0 over voiced frames ≥ 0.02 (a tone or chord locks YIN to one value: 0.00; NSA test minimum 0.099);
4. moving in loudness: std of frame loudness ≥ 3 dB (a steady tone, hum or stationary noise: under 1 dB; NSA test minimum 6.0);
5. not spectrally flat: long-term spectral flatness < 0.5 (white noise 0.998; NSA test maximum 0.39).

Any failed cue gates the file; the failed cues are the reasons (reason_silence, reason_unvoiced, reason_tone, reason_static, reason_noise) and go into the evidence sentence. Score is always 0.5. Cost about 0.13 s per clip (YIN).

Set from the test set, not guessed. outputs/inventory/nsa_test_speech_gate.csv holds the five cues for all 1,671 test files; every threshold sits beyond that file's extreme, so no NSA test file is gated. A sixth cue, "a few spectral lines" (share of long-term energy in the top five bins ≥ 70%, meant for tones and chords), was tried and dropped: 77 real test files exceed it, because speech puts most of its long-term energy in a few low bins; tones and chords are caught by cues 3 and 4 anyway.

Cases (tests/test_speech_gate.py): the submission preflight's silence and three-note chord, a 220 Hz tone, white noise, a speech-like synthetic (moving pitch, syllable-rate loudness, a pause) that must pass, the same at −64 dBFS RMS that must not read as silence, and 60 real test clips of which ≥ 98% must pass (100% do).

Known limit. Rhythmic polyphonic music moves in pitch and loudness and can pass this gate. The submission preflight (silence, chord) and the flag on any fused score above 0.8 remain the backstop; a learned speech/music discriminator would be the next step if the draft review shows music in the test set.

3. The policy (speech_gate.apply_default_answer)

Corrected Sat 08:30 after the fusion consult (docs/consults/2026-09-26_fusion-strategy_RESPONSE.md, item 5), which proved the pinned block is the one direction-robust hedge only if it sits strictly below every determinate score; the first version (0.02 + 1e-4 × fused) did not, because the pipeline scores some real files as low as 0.0009.

• Determinate files (speech): the fused score is mapped monotonically into [0.001, 1.000] (0.001 + 0.999 × fused); the ranking among them is unchanged and none can fall into the block.
• Gated files: a pinned block in [0.000, 0.001), strictly below every determinate score under either scoring convention. Within the block, files are ordered by whatever weak signal remains (order_by: the raw M1b score is the natural choice; voiced fraction and duration can be folded into it; the fused score by default) with a 1% share of the block's range given to a deterministic hash jitter of the file key, so no two files share a value (a superset of thresholds can only help the sweep, never hurt). Decode failures and files with no signal (NaN, or failed=True) go to the bottom tenth of the block, [0, 0.0001), by jitter alone.
• Only the gate flag gates a file. High-disagreement or low-confidence files are a fusion matter and stay in the determinate range (the consult compresses them toward its low end there); they must not be collapsed into the block.

Call shape unchanged for the runner: apply_default_answer(fused, is_speech); the keyword options order_by, keys, failed refine the block. Zero test files are gated today, so this changes no submitted ranking; the determinate map shifts every score by at most 0.001.

4. Speaker-embedding drift (hearsay.detectors.speaker_drift)

ECAPA-TDNN speaker embeddings (SpeechBrain spkrec-ecapa-voxceleb, Apache-2.0, 89 MB, fetched to weights/spkrec-ecapa-voxceleb; loaded from there, never from the Hub at run time) on 1.0 s windows at a 0.5 s hop over the silence-trimmed clip (cap 8 s), unit-normalized; features are the minimum, mean and std of the pairwise cosines and the minimum adjacent-window cosine. Rule: minimum pairwise cosine below 0.05 → score 0.6 ("voice drifts across the clip"), else 0.5. Cost about 0.25 s per clip on CPU; the encoder loads in under 2 s.

Calibration (outputs/inventory/speaker_drift_calibration.csv; inner rows only): 120 LibriSpeech real, 60 LJ real, 160 DiffSSD fakes (20 per generator), 80 synthetic two-speaker splices (the first half of one LibriSpeech speaker's clip glued to the first half of another's), 80 same-speaker splices as control.

At 0.05 the flags fall on LibriSpeech real 11.7%, LJ real 6.7%, fakes 1.9%, same-speaker splices 12.5%, two-speaker splices 93.8%. One-second ECAPA windows are noisy for one speaker too (median minimum cosine 0.20 for real, 0.28 for fakes), which is why the rule sits near zero.

A finding worth a line in the README: the fakes are the most consistent voices. Every drift statistic points the same way: fakes have higher minimum, adjacent-minimum and mean cosines than real speakers (class AUCs in the table below). A cloned voice does not vary from second to second the way a person does. That makes drift evidence of editing or of a real speaker, not of synthesis; a fused column would learn "drift = real", the opposite of the rubric's framing, so the export is for the routing log and per-file evidence only.

These are single-statistic AUCs on a 340-clip inner sample, not a validated detector; they say the direction, and they mark "speaker-embedding consistency" as a candidate v4-style feature family (spk_*, four columns, 0.25 s per clip) for whoever reopens the feature lane. Not pursued here: the lane is closed and the column would need the ECAPA encoder inside every extraction worker.

Export summaries (scripts/score_detector.py, all 21,671 rows, no failures; gate 405 s on 6 workers, drift 2,043 s on 4 threads).

The 84 gated training clips are very noisy recordings and very noisy vocoder output (spectral flatness above 0.5); none of the test set is. The drift rate on the test set (16%) is above every training corpus; per generator in training the rule fires on 0–4% of fakes and 7% of real clips, so on a 70%-real test set with phone and field recordings a rate in the teens is what "drift marks real speakers and harder conditions" predicts. That is one more reason it is a routing and evidence detector, not a fusion column: fused, it would push 16% of the test set, mostly real, toward synthetic.

5. Exports and where they plug in

• outputs/detector_scores/speech_gate.csv (score 0.5 throughout) and speech_gate_results.csv with is_speech and the reasons per path; outputs/detector_scores/speaker_drift.csv / _results.csv with the cosine statistics per path. All 21,671 rows.
• Orchestrator: read is_speech from the gate's features (or the results file) and apply apply_default_answer to the fused score; print the gate's evidence for gated files.
• Docker: needs speechbrain (now in pyproject.toml) and weights/spkrec-ecapa-voxceleb copied into the image; HF_HUB_OFFLINE=1 is fine because the detector loads from the local directory.