Published Local SamplesPublished WAV FilesOffline Ledger Signatures
Published WAV Comparison: Locally Labeled Speech
Listen to the published WAV files and inspect the current Worker decisions on those exact bytes. The labels are local assertions or documented generation recipes; they are not all independently verified upstream. The 29-file sample registry lists SHA-256 hashes, labels, and the evidence tier for each source.
Byte identity: the scores in this section are computed on the exact public WAV bytes. The 168-trial Python, JavaScript and SSL ledgers are bound to the separately stored evaluation-corpus bytes by manifest SHA-256 pins. Four public files, synthetic_studio_brian.wav, synthetic_studio_eric.wav, synthetic_ad_miguel_sovereign.wav and synthetic_abundera_narrator.wav, have different file hashes from their evaluation-corpus counterparts. Measured: identical decoded audio (same PCM hash, mono, 16 kHz, same length); the public copies carry one extra 34-byte metadata chunk, and Python scores were identical on both versions. The other 25 public files match their corpus hashes. The sample registry records this per file as pcm_identical_container_differs.
Selected Speech Label AgreementDEMONSTRATION CORPUS
26/26
Current decisions agree with the locally assigned labels for all 26 binary speech trials (29 total audio files). Curated demonstration subset; not a cross-dataset benchmark EER.
Latency Scope
Unmeasured
No deployed end-to-end request latency has been measured for this published sample set.
Attestation Protocol
Offline Ed25519 Receipts
The offline Python proof ledger includes signed receipts. The deployed API currently returns unsigned responses. Verify a ledger signature and independently pin its key before relying on it.
Interactive Perception Test
Can Human Ears Beat Acoustic Signal Analysis? Take the Acoustic Turing Challenge
Test your perception against locally labeled synthetic and human speech recordings.
The detector measures spectral phase, estimated glottal pulse shape, jitter, and a diagonal bispectral coupling estimate in decoded audio. These are statistical features of the recording, not a complete bicoherence map. They do not establish the speaker's physiology or prove how the audio was produced.
Recording conditions, speech content, codecs, and generators can change these measurements. The decision score is uncalibrated, and a single recording cannot establish a universal biological threshold.
Selected Published WAV Results
These are current Worker decisions on the exact downloadable WAV files at threshold 0.50. Scores are uncalibrated decision values.
Published sample
Source label
Decision score
Worker verdict
Studio Brian
Synthetic
0.9933
Synthetic detected
Human monotone reader
Human
0.0100
Authentic human
Biden robocall sample
Synthetic
0.5498
Synthetic detected
Biden public address sample
Human
0.2315
Authentic human
Playable comparison: listener impression and measured score
Listen & Verify
Play both published WAVs and compare your own impression with the Worker decision scores. No listener study is reported here:
Authentic Human Monotone Reader
HUMAN-LABELED · decision score 0.0100
Human Speaker with Flat, Robotic Pitch Cadence (16 kHz WAV, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5387; LPC residual skew: +0.8784; pitch period variation: 2.3496%; bispectral coupling: 0.6170. These measurements do not establish the source or vocal physiology.
Sample Pair 1: Clean Male Reader vs Studio Persona Eric (Studio sample)
Matched Acoustic Benchmark
These two local files have human and synthetic labels. Listen to them and inspect the measured scores; no listener performance or generator mechanism was tested here.
Authentic Human Speech (Male)
HUMAN-LABELED · decision score 0.0100
Local human label; recipe fetched Narsil/asr_dummy/1.flac (16 kHz, 16-bit PCM mono, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5353; LPC residual skew: +0.7552; pitch period variation: 2.0026%; bispectral coupling: 0.3547. These measurements do not establish the source or vocal physiology.
Sample Pair 2: Clean Female Reader vs Abundera QR Overview Video Ad
Selected Commercial Audio Examples
This pair compares a locally labeled human recording with a site-claimed video-ad extraction. The source MP4 is not available in this repository; the measured score applies only to the published WAV.
Authentic Human Speech (Female)
HUMAN-LABELED · decision score 0.0100
Local human label; recipe references LibriSpeech clean validation speaker 2277, with an existing-file fallback (16 kHz, 16-bit PCM mono, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5387; LPC residual skew: -0.9791; pitch period variation: 1.2903%; bispectral coupling: 0.3067. These measurements do not establish the source or vocal physiology.
Sample Pair 3: Studio Neural Voice Cloning / Kokoro 82M vs Authentic Conversational Human
Studio Local TTS
These published WAVs have local human and synthetic labels. The original studio stem is not pinned, so the named generator attribution remains a site claim.
Human Conversational Reader
HUMAN-LABELED · decision score 0.0100
Local source label: LibriSpeech Reader Speaker #3 (16 kHz, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5619; LPC residual skew: +1.2001; pitch period variation: 1.8950%; bispectral coupling: 0.2721. These measurements do not establish the source or vocal physiology.
Sample Pair 4: F5-TTS Recipe Attribution vs Authentic Human
Flow Matching Frontier
The local generation script contains an F5-TTS inference recipe, but no pinned execution record links the published WAV to a specific model build. The paired files below show measured Worker decisions, without attributing individual features to a synthesis mechanism.
Sample Pair 5: Kokoro ONNX Recipe Sample vs Authentic Female
Recipe documented
The locally documented recipe for this WAV calls Kokoro ONNX, despite its StyleTTS2-derived filename. The paired local WAV examples below show this detector's decision scores on one synthetic and one human recording.
Authentic Human Speech (Female)
HUMAN-LABELED · decision score 0.0100
Local human label; recipe references LibriSpeech clean validation speaker 2277, with an existing-file fallback (16 kHz, 16-bit PCM mono, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5387; LPC residual skew: -0.9791; pitch period variation: 1.2903%; bispectral coupling: 0.3067. These measurements do not establish the source or vocal physiology.
Sample Pair 6: Azure Jenny Neural Sample vs Monotone Human
Evasion Defense Benchmark
The paired local WAV examples below show the current Worker result for an Azure Jenny sample and a monotone human reader. Feature values are measurements of these recordings, not proof of a particular synthesis mechanism.
Human Monotone Cadence
HUMAN-LABELED · decision score 0.0100
Flat, robotic-sounding genuine human speech (16 kHz, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5387; LPC residual skew: +0.8784; pitch period variation: 2.3496%; bispectral coupling: 0.6170. These measurements do not establish the source or vocal physiology.
Sample Pair 7: Azure Multilingual v2 (Microsoft Copilot Andrew) vs Authentic Male
Copilot Voice Engine
The local generation script uses en-US-AndrewMultilingualNeural for this sample. The Worker decision applies to this WAV; the recorded features do not identify a specific synthesis artifact.
Authentic Human Speech (Male)
HUMAN-LABELED · decision score 0.0100
Local human label; recipe fetched Narsil/asr_dummy/1.flac (16 kHz, 16-bit PCM mono, 3.50s)
Worker decision score: 0.0100 (threshold 0.50); STFT phase dispersion: 0.5353; LPC residual skew: +0.7552; pitch period variation: 2.0026%; bispectral coupling: 0.3547. These measurements do not establish the source or vocal physiology.
Listen to the published WAVs and compare their current Worker decisions. Generator names are local labels with different evidence tiers; the score does not establish the production method:
Abundera Video Ad Fleet: In-the-Wild Audio Extractions
3 published ad excerpts classified synthetic
These three local WAV excerpts are described on this site as extractions from Abundera video ads. Their source MP4 files are not part of this repository, so the extraction history cannot be verified here. The current Worker classifies each published WAV as synthetic:
The local 168-trial sweep uses convenience subsets of external collections. It is not an evaluation on their complete official partitions, and the source archives are not pinned to immutable revisions.
See the trial ledger for exact selected filenames and processed-audio hashes. Full-dataset error rates cannot be inferred from these subsets.
Named voice sample comparisons
3 locally labeled sample pairs
These examples compare a locally labeled synthetic recording with a locally labeled human recording. The original media is not pinned here; the reported decisions apply to the published WAVs:
Incident context: The FCC documented an AI-cloned Biden robocall sent before the 2024 New Hampshire primary. The published WAV here has a local incident label, but no pinned upstream source establishes that it is the recording investigated by the FCC. The comparison recording is likewise not pinned to an upstream release.
LOCAL SYNTHETIC LABEL: Joe Biden NH Robocall
SYNTHETIC-LABELED · decision score 0.5498
Local label: 2024 NH Primary Robocall Audio (16 kHz, 6.00s)
Worker decision score: 0.5498 (threshold 0.50); STFT phase dispersion: 0.5451; LPC residual skew: +0.5942; pitch period variation: 2.2299%; bispectral coupling: 0.1751. These measurements do not establish the source or vocal physiology.
The two WAVs carry In-the-Wild dataset labels in the local manifest. Their original upstream files are not pinned here, so the scores should be read as decisions on these published bytes.
LOCAL SYNTHETIC LABEL: Barack Obama Clone
SYNTHETIC-LABELED · decision score 0.9933
Local dataset label: In-the-Wild ITW-6 (16 kHz, 4.62s)
Worker decision score: 0.9933 (threshold 0.50); STFT phase dispersion: 0.5549; LPC residual skew: +0.3595; pitch period variation: 2.2128%; bispectral coupling: 0.5078. These measurements do not establish the source or vocal physiology.
The two WAVs carry In-the-Wild dataset labels in the local manifest. Their original upstream files are not pinned here, and acoustic scores do not identify a voice-production mechanism.
LOCAL SYNTHETIC LABEL: Donald Trump Clone
SYNTHETIC-LABELED · decision score 0.9608
Local dataset label: In-the-Wild ITW-41 (16 kHz, 4.64s)
Worker decision score: 0.9608 (threshold 0.50); STFT phase dispersion: 0.5463; LPC residual skew: +0.4352; pitch period variation: 1.6192%; bispectral coupling: 0.2271. These measurements do not establish the source or vocal physiology.
Edge Cases: Adaptive VAD Gating & Non-Vocal Signal Handling
Zero CPU Waste
A production detection system must not produce false classifications on silence or non-speech background signals. Noise or Voice uses adaptive energy-flux VAD gating to return an inconclusive result for non-speech inputs:
Ambient Room Silence
Low-level background room tone (peak amplitude < 0.01).
Worker verdict: INCONCLUSIVE_SILENCE. Decision score is not evaluated for this non-speech sample.
Wiener Spectral Flatness: Geometric to arithmetic mean ratio of PSD across K bins. Values approaching 1 describe a flatter spectrum; lower values describe spectral concentration. This statistic does not establish speech provenance or distinguish human from synthetic speech on its own.
Offline Python Proof Ledger: Signed Evaluation Receipts
Offline Ed25519
The offline Python core proof ledger publishes per-audio SHA-256 hashes, acoustic telemetry, and signed receipts. The deployed API currently has no signing key configured and returns unsigned responses. The larger evaluation sweep publishes per-trial hashes and decisions; it does not include a signed receipt for every row.
Epistemological Distinction: Cryptographic Attestation vs Scientific Ground Truth
An Ed25519 signature can verify that a receipt was signed by the holder of a particular key and has not changed since signing. A SHA-256 digest can bind it to specific audio bytes. Neither establishes the evaluator binary, execution time, label correctness, or detector accuracy on its own.
The canonical JSON ledgers include source-code digests and audio hashes. The 29-file local suite contains 26 binary-labeled speech recordings, two non-speech cases, and one mixed recording. The 168-trial sweep links to its corpus manifest digest; these are internal regression results, not an independent blind evaluation.
Subset Purpose:Selected diagnostic regression corpus with source and codec variation; no adaptive-attack trials.
Selection Protocol:Deterministic convenience selection from source releases; the selected rows are not a random or representative draw.
Manifest & Hashes:Complete trial list and per-trial SHA-256 hashes published in evaluation-sweeps.json.
Comparability Scope:Internal diagnostic results; no same-set official challenge evaluation has been performed.
The 168-trial internal sweep uses selected rows from several source corpora:
WaveFake Multi-Vocoder Suite (N=64): First eight rows per class in the source Parquet: seven neural vocoders and one LJSpeech human class. This selection is deterministic but not random.
In-The-Wild Deepfake Suite (N=40): First two authentic and synthetic recordings for each shared speaker in sorted speaker order, capped at 20 per label.
Core Golden Suite (N=26 in sweep, N=29 in gate): Continuous regression benchmark spanning locally labeled generator samples, including an F5-TTS recipe and a Kokoro ONNX recipe under a StyleTTS2-derived filename and paired authentic human recordings.
Codec-Degraded Suite (N=18): Controlled transcode evaluations across PSTN G.711 μ-law (8 kHz telephony) and Opus (16 kbps & 32 kbps VoIP).
LibriSpeech Clean Baseline (N=20): First recording for each of the first 20 speakers encountered in the LibriSpeech clean validation split.
The JSON trial ledger publishes processed-audio SHA-256 digests, locally assigned labels, extracted features, and prediction verdicts. Source releases are not pinned to immutable revisions. See the site sample registry for per-WAV provenance tiers.
An exploratory analysis of the 168-trial internal corpus ranked a fractional LPC-derived feature highly. These values describe this selected corpus and do not establish a universal physical origin or out-of-distribution accuracy.
Ledger: docs/PHYSICAL_ACOUSTIC_SIGNATURES.json
Continuous Regression Gate
Local gate
The local gate checks the 26 labeled core speech clips and compares the 22-file diverse suite with its baseline. This is a regression check, not an estimate of population accuracy.
Ledger: tests/regression_baseline.json
Run Local Verification Checks
With Node.js 18 or newer, download the standalone public verifier and run it against this site. It fetches all 29 registered WAVs, the release record and its hash-pinned DSP source, and the Worker result ledger, then checks digests and recomputes verdicts, scores, and features. Source labels remain local assertions.
From a source checkout with the required audio files and Python/Node dependencies, the following commands rerun additional local checks. The 168-trial source corpus is absent from a fresh public checkout, and its upstream releases are not pinned:
# 1. Run the current Python test suite
python3 -m pytest
# 2. Run automated continuous regression test gate across baseline ledger
python3 scripts/regression_gate.py --all
# 3. Run empirical physical acoustic discovery harness (unsupervised statistical signatures)
python3 scripts/discover-acoustic-signatures.py
# 4. Run Python benchmark engine across 29 files (26 binary speech labels, 2 non-speech cases, and 1 mixed file)
python3 -m engine.benchmark
# 5. Run comprehensive evaluation sweep across 168 trials (In-The-Wild, WaveFake, LibriSpeech, Codecs)
python3 -m engine.benchmark --eval-corpus
# 6. Verify single audio sample via CLI with Ed25519 attestation
python3 -m engine.cli verify site/audio/samples/human_real_biden_speech.wav
# 7. Run the JavaScript DSP check on the published WAVs
node scripts/test-js-dsp.mjs
# 8. Run full Playwright E2E UI & verification test suite
node scripts/test-e2e.mjs