Experimental Synthetic Voice Detection

Noise or Voice is an experimental acoustic research prototype built around interpretable DSP features rather than a large learned classifier. Its current JavaScript engine reaches 73.21% accuracy on a selected 168-trial stress corpus at threshold 0.50, but only 37.50% on the 40-trial In-the-Wild subset. These are internal convenience samples, not full standardized challenge results. A verdict supports further review; it does not authenticate a speaker.

Research question: how well do interpretable acoustic features distinguish synthetic speech under channel and domain changes? Current artifacts include a complete-recording detector, public sample demonstrations, internal score ledgers and documented generalization failures. Independent validation and score calibration remain open. Read Cisco Caceres’s research agenda.

# 1. In-line synchronous REST verification
curl -X POST https://noiseorvoice.com/api/v1/verify \
  -H "Authorization: Bearer $NOISEORVOICE_API_KEY" \
  -F "audio=@synthetic_edgetts_aria.wav" \
  -F "sensitivity=0.85" \
  -F "extract_telemetry=true"

{
  "verdict": "SYNTHETIC_DETECTED",
  "confidence_score": 0.999,
  "decision_score": 0.999,
  "score_semantics": "uncalibrated_detection_score",
  "score_calibrated": false,
  "confidence_calibrated": false,
  "scoring_engine": "noiseorvoice-js-dsp/2.4-SPECTRA",
  "telemetry": {
    "vocoder_phase_incoherence": 0.517,
    "glottal_jitter_variance": 2.4347,
    "rolloff_95_hz": 2219
  },
  "attestation": null,
  "signing_status": "unavailable"
}

Measured Results and Known Failures

The Python implementation differs: 73.21% accuracy / 33.98% EER overall, and 37.50% accuracy / 60.00% EER on the same In-the-Wild subset. Four threshold decisions differ. There is no complete ASVspoof, WaveFake, MLAAD, or FakeAVCeleb evaluation for this detector.

All 29 published WAVs are checked for exact hashes, playback, and browser/API result parity. The offline Python proof ledger has signed receipts; deployed API responses are currently unsigned. Of these selected examples, 26 speech clips match their recorded labels; tone and silence are inconclusive. The mixed track is not counted as binary accuracy. Some source labels lack independent upstream verification.

The broader corpus is a nonrandom convenience sample with overlapping sources and related recordings. Full corpus audio is not bundled for public reproduction. No validated streaming latency, representative telephony performance, controlled listener study, or adversarial bypass rate is claimed.

6 October 2026 research diagnostic: a frozen DSP snapshot was tested on 12 human-speech targets across 36 clean and artificial-babble trials. At threshold 0.50, the six held-out test speakers produced human false alarms of 4/6 clean, 4/6 at 10 dB, and 5/6 at 5 dB. This small cached read-speech diagnostic contains no synthetic class; it cannot estimate synthetic misses, EER, authentication performance, or real-time behavior. Read the frozen snapshot commitments, methods, failures and limitations.

2026-10-06 ODSS two-class research diagnostic: a previously frozen JavaScript DSP reference core; these historical results have not been rerun with the current detector source (ca6a48f4f408) evaluated 32 test recordings from 4 speakers at unchanged threshold 0.50. It falsely flagged 10/16 human recordings and missed 4/16 synthetic recordings, with 0 processing errors and 0 abstentions. Balanced correct fraction was 56.25%; the descriptive speaker-cluster interval [40.625%, 78.125%] is fragile with only 4 clusters. These exploratory failures do not establish deployment readiness or establish authentication, streaming, radio or general-population performance. The original ASVspoof selection is unchanged; its later completed diagnostic is reported separately below. Read the canonical protocol, licensed sources, deviations and limitations; download the frozen aggregate.

Completed ASVspoof 5 diagnostic. 6 October 2026 Pacific ASVspoof 5 research diagnostic: the historical frozen DSP snapshot (ca6a48f4f408) scored all 192 registered recordings at unchanged threshold 0.50. On the 96-record evaluation subset it falsely flagged 22/48 human recordings and missed 8/48 synthetic recordings; 66/96 were correct (68.75%), with 0 processing errors and 0 abstentions. This developer-selected, speaker-disjoint two-TTS-family diagnostic has no independent blinded custodian. Source/channel/training exposure and separate synthetic-content reuse scope remain unresolved. Scores are uncalibrated; these historical results do not validate the current deployed detector, authentication, streaming, radio use or general-population performance. Read methods and retained errors; download the aggregate.

Internal 18,000-trial ASVspoof 5 comparison (not blind). On 18,000 developer-selected, labeled trials from the ASVspoof 5 evaluation partition (2,000 human, 16,000 synthetic, attacks A17 to A32), every fixed rule was scored on every row: Worker 59.86% balanced and 81.18% raw, rejecting 1351/2000 human trials and missing 2036/16000 synthetic trials; learned feature model 55.52% balanced and 54.37% raw, rejecting 860/2000 human trials and missing 7353/16000 synthetic trials; author-released SSL model 59.07% balanced and 90.89% raw, rejecting 1637/2000 human trials and missing 2/16000 synthetic trials. The Worker abstained on 36 synthetic trials, which count as errors, and two retained human trials had prior developer exposure. Balanced accuracy is the primary figure because raw accuracy is inflated by the 8:1 synthetic share; all three balanced figures are near 60% or lower. Historical source scope: this comparison ran frozen DSP source fdfc11cd0cb2 (/reproduce/dsp-fdfc11cd0cb2456b364b4ae770549b0a85dcbd27c3bbbd91dbf0c732a8c16f8e.txt); this published ledger was not generated by the current source (afd2fa0e7600). Adaptation and input differ: the learned model was fit on a study training subset, the Worker and learned model read the full WAV, and the SSL model read only the first 4.04 seconds. This is an internal, non-blind comparison with known labels. It is not independent, not a full or official ASVspoof 5 challenge result, and it establishes no superiority. The full audio is not bundled. Source: ASVspoof 5 database, DOI 10.5281/zenodo.14498691, Open Data Commons Attribution License. Read the qualified results; download the summary.

Interactive Spectrogram & Synthetic Artifact Detector

Inspect acoustic feature estimates and a spectrogram directly in your browser. Demo analysis stays on your device; voluntary research contributions use a separate upload flow.

{
  "status": "ready",
  "verdict": "STANDBY",
  "processing_mode": "browser_demo_local_analysis",
  "signed": false
}

Take the 10-Clip Acoustic Turing Challenge: Compare Your Guesses with the Detector

Listen to 10 randomly selected source-labeled recordings and compare your guesses with the current detector. This interactive exercise is not a controlled perception study or evidence that the detector outperforms human listeners.

Selected Public-Figure Audio Case Studies

Explore selected clips labeled synthetic and human in the source collection. The current API engine scores these specific files; their source chain of custody and labels have not been independently verified here. Play each clip or load it into the client-side spectrogram. These examples do not establish performance on new speakers, channels, or generators.

Detection Threshold & Error Rate Tradeoff Explorer

Explore how the decision threshold changes errors on the published 168-trial JavaScript score ledger. These observed rates describe this selected corpus, not expected production fraud or caller rejection rates.

How the Detector Works

The engine estimates phase behavior, spectral rolloff, LPC residual skewness, periodicity, and reflection-coefficient changes. Their distributions overlap across human and synthetic speech. Hand-set feature rules need validation on representative recordings.

US Patent 11,924,717 concerns communications analytics in a walkie-talkie network. It does not cover or validate this detector's later synthetic-speech methods. Read the patent.

Pilot Interfaces & Evaluation Receipt Design

Private pilot designs for reviewing recorded audio. Operational use requires independent evaluation, configured security controls, and a documented response to incorrect verdicts.

Frequently Asked Technical Questions