The library

Rendered from docs/handoffs/2026-09-26_m3-spectra-aasist-handoff.md; the markdown file is the record.

Handoff: M3, Spectra-AASIST as the second deep detector (Sat Sep 26, 2026, ~03:35)

Purpose of this chat: run the off-the-shelf Spectra-AASIST detector over the NSA fold file and test set, and export its scores in the fusion format so M4 (fusion) gets a second deep score stream that is decorrelated from M1. Time box: 45 min of build plus the scoring run; docs/plan.md says "drop on friction". No training, no fine-tuning: this is a scoring rung.

Deliver:

• outputs/detector_scores/spectra_aasist.csv with columns path, fold, split, score, logit, split ∈ {inner_oof, holdout, test}, one row per fold-file row (20,000) plus one per test file (1,671), same shape as outputs/detector_scores/handcrafted_v1.csv.
• models/m3_spectra_<stamp>/meta.json via hearsay.metrics.report on the outer holdout: overall, per generator, per bona fide source, sponsor_code_asis and sponsor_code_flipped, plus the direction check.
• A short note in docs/reports/2026-09-26_m3-spectra.md (numbers, timing, what it's blind to), and an M3 line in docs/STATUS.md.
• Do not write TSVs or touch submissions/. The main chat turns detector scores into submissions.

Context

HEARSAY is the HackGT 13 NSA HEARSAY entry (software only). Read CLAUDE.md, docs/plan.md, docs/STATUS.md first. Scoring is 60% normalized minDCF with C_FA = 4, π_synth = 0.3 (false alarms on real speech cost 9.33× a miss); output 1.0 = synthetic, never flip.

Four chats run in parallel. Stay in your lane.

Device decision (make it first, with Nathan): Spectra-AASIST wraps a full XLS-R 300M, so it costs about what M1's extraction cost: ~0.2 s per 4 s window on MPS, roughly 3–5× that on CPU. 21,671 clips at up to 16 windows each is ~1–2 h on MPS, 4–8 h on CPU. The main chat is using the MPS GPU (an In-the-Wild extraction was running at 03:30: pgrep -fl extract_embeddings). Options:

1. Coordinate MPS with the main chat: run when their extraction finishes, or run at lower priority (nice) in segment mode with max_windows=4.
2. CPU with max_windows=2 on the 6 spare cores (nice -n 10), which fits the 45-min box only for a subset; use --limit for a pilot, then let it run in the background. Ask before taking the GPU. Never run on vast.ai (that's the M5 chat's budget).

What already exists:

• src/hearsay/spectra.py: load_spectra(device), spectra_logits(model, x, max_windows=16) (repeat-pads short clips into 64,600-sample windows with 50% hop, pre-emphasis, mean of window logits), synth_logit(logits) = logit_spoof − logit_bonafide (increases with synthetic likelihood). Weights in weights/Spectra-AASIST/ (model.py + model.safetensors, 1.26 GB; load_spectra patches the hub id to the local XLS-R copy).
• scripts/spectra_direction.py: asserts direction (AUC > 0.5 and spoof median above bona fide median) on a labeled manifest and fits an optional class-balanced Platt map. Already run on 494 ASVspoof 2019 eval clips: outputs/spectra/asv19eval500_scores.csv (raw logits per file; spoof synth_logit ≈ +11, bona fide ≈ −7 on the first rows). Result on that ASVspoof sample: AUC 1.0, EER 0.0, spoof median synth_logit +9.7 vs bona fide −7.6, 0.169 s per clip on MPS; provisional Platt map (a = 4.42, b = 10.51) in outputs/calibration/spectra_platt_asv19eval500.json. That is public data with a known silence shortcut; expect worse on NSA.
• scripts/score_detector.py shows the fusion-format export for non-learned detectors (split=inner_oof for inner rows, holdout, test; NaN score for failed rows; logit = log(s/(1−s)) on a clipped score). Mirror its export block. scripts/train_handcrafted.py:75-88 has by_group() for the per-generator / per-source readouts; copy it.
• Fold file: splits/nsa_folds.csv (path, label, generator, speaker, source, group, fold; holdout = playht + wavegrad2 + 26 bona fide groups; inner folds "0"–"4"). Never regenerate it. Test manifest: outputs/manifests/nsa_test.csv (filename, path, relative paths; row order is the submission order).

Working branch / worktree

main in ~/Projects/hearsay (shared worktree). Uncommitted at handoff: the M5 chat's untracked spec/plan/consult files under docs/ (not yours; never stage them). Stage only your own files with git add <paths>; never -A. Don't push without asking Nathan.

Environment / setup

cd ~/Projects/hearsay && uv sync
uv run pytest -q          # 215 passed at 3c170d5
uv run ruff check .       # clean
uv run python -c "from hearsay.spectra import load_spectra; m = load_spectra('cpu'); print(type(m).__name__)"

What to do next

1. Device decision with Nathan (above). Check pgrep -fl extract_embeddings and the main chat before touching MPS.
2. Input path parity. Spectra was trained on ~4 s crops. Decide and record whether to score (a) spectra_logits on the raw load_audio clip (current code: repeat-pads short clips, which is the tiling shortcut M1 v2 removed), or (b) the silence-trimmed clip via hearsay.embed.prepare_segment(x) (trim, cap 8 s) and then windows without repeat-padding for clips ≥ 64,600 samples, zero-padding shorter ones. Recommendation: (b), with the same treatment for train and test rows, so silence and tiling can't become shortcuts. Whatever you pick, the same function scores every row (inner, holdout, test).
3. Direction check on NSA first: run scripts/spectra_direction.py on ~300 inner-fold rows (150 per class, sampled from splits/nsa_folds.csv where fold != holdout) before any export. If AUC < 0.5 it exits non-zero; do not "fix" by flipping without telling Nathan.
4. Write scripts/score_spectra.py: loads the fold file + test manifest, scores every row with the chosen path, exports outputs/detector_scores/spectra_aasist.csv (score = sigmoid(synth_logit) or a Platt map fit only on inner-fold rows, never holdout; logit = synth_logit), writes models/m3_spectra_<stamp>/meta.json with hearsay.metrics.report on the holdout, by_group breakdowns, timing, device, max_windows, and the input-path choice. Support --limit N for a pilot and --device.
5. Pilot with --limit 200, check timing, extrapolate, then run the full set in the background (nohup … & or the terminal), logging to outputs/spectra/score_spectra.log.
6. Tests (tests/test_spectra.py, hermetic, no weights): direction contract (synth_logit increases with spoof logit), export schema and counts (20,000 + 1,671, unique paths, split values), NaN on decode error, no repeat-padding if you chose (b). Mark anything that needs weights @pytest.mark.needs_weights @pytest.mark.slow and skip when weights/Spectra-AASIST is missing (pattern in tests/test_audio_and_submission.py:177-194).
7. Report + STATUS line, disclosure line in CLAUDE.md (Spectra-AASIST "how used": M3 score stream; license still unclear, say so). Tell the main chat the file name so fusion picks it up.

IMPORTANT: tests & at-risk artifacts

• Test: uv run pytest -q → 215 passed at 3c170d5; uv run ruff check . → clean. Add your tests; the docs-consistency test bans the bare word "CSV" in CLAUDE.md/docs/plan.md (write .csv) and any π other than 0.3.
• At risk (gitignored, NOT archived): outputs/detector_scores/*.csv, outputs/spectra/, models/. weights/Spectra-AASIST (1.26 GB) and weights/wav2vec2-xls-r-300m are local only.
• In flight (main chat, MPS): extract_embeddings.py --manifest outputs/manifests/itw_stress.csv at 03:30. Check pgrep -fl "extract_embeddings|train_probe".
• The raw audio lives on the external drive (/Volumes/Crucial P3 NVME Gen 3 2TB/hearsay/, symlinked from data/); it must be mounted. ASVspoof/In-the-Wild/MLAAD were copied there at 03:12 and may be symlinked by the time you start.

Analytical notes

• M1 v2 (frozen XLS-R probe, layer 7) is the bar for deep detectors: outer-holdout minDCF 0.1458, EER 2.5% (models/m1_wav2vec2-xls-r-300m_L7_20260926-0302/meta.json); wavegrad2 0.21, LibriSpeech real 0.16. M3 does not need to beat it; its value is a second, differently trained deep score for the stacker. Report where it disagrees with M1 (per generator, per source).
• Handcrafted detector v3: holdout minDCF 0.25, blind to grad_tts, pro_diff and ElevenLabs. Compression forensics: near chance on decoded inputs.
• Known traps from the CPU chat's report (docs/reports/2026-09-26_cpu-detectors.md): the test set is band-limited differently from the training corpora, so anything keyed on the 7–8 kHz band false-alarms on real test audio. Check frac_above_half on the test set in meta.json: with a 70% real test set, a value far above 0.3 means false alarms.
• Sponsor scoring code treats higher = bona fide; hearsay.metrics.report gives both readings. Submit 1.0 = synthetic regardless.
• Spectra's output convention: index 0 = spoof, index 1 = bonafide (model card); synth_logit already handles it. The direction script is the guard.

Pointers

CLAUDE.md → docs/plan.md (ladder row M3, cut order) → docs/STATUS.md → src/hearsay/spectra.py → scripts/spectra_direction.py → scripts/score_detector.py (export format) → scripts/train_handcrafted.py:75-88 (by_group) → src/hearsay/metrics.py → docs/reports/2026-09-26_cpu-detectors.md (band-limit trap).