Trust & Content Safety
ClearFlow DeepSense Content Trust Infrastructure v2.0
Last updated: March 21, 2026
Our Principles
Provenance first. ClearFlow-originated audio has C2PA credentials, SHA-256 chains, and watermarks. Detection only runs on external audio.
Risk signal, never verdict. We report confidence bands and risk indicators — never "this IS a deepfake."
Honest grading. Per-language accuracy grades are published. We do not claim uniform accuracy across all languages.
Safety is free. Pipeline inline scanning costs zero FlowCoins. Safety should never be paywalled.
Detection Layers
DeepSense uses multiple independent analysis layers with Bayesian fusion. No single layer can determine the result alone.
Spectral analysis calibrated per language. Detects statistical anomalies in frequency distribution.
Tests audio stability across 5 codec simulations (MP3, Opus, G.711, noise, pitch shift). Real audio is stable; synthetic is erratic.
Analyzes Open Quotient, H1-H2 spectral tilt, jitter, shimmer, and Normalized Noise Energy from voiced segments.
Measures coefficient of variation in segment durations. Natural speech has irregular timing; TTS is regular.
Layers are weighted via Bayesian fusion. Layer disagreement (> 0.4 delta) triggers "inconclusive" rather than a false verdict.
Per-Language Accuracy Grades
Detection accuracy varies significantly by language. Tonal languages and low-resource languages have lower accuracy. These grades are published transparently via our API.
| Language | Grade | Notes |
|---|---|---|
| English (US/GB) | B | Best accuracy. Most training data. |
| Spanish (ES/MX) | C | Good but dialectal variation reduces accuracy. |
| Portuguese (BR) | C | Similar to Spanish accuracy. |
| German | C | Compound word prosody adds challenge. |
| French | C | Liaison patterns complicate analysis. |
| Hindi | D | Limited training data. Retroflex consonants. |
| Korean | D | Agglutinative morphology affects timing. |
| Japanese | D | Pitch accent patterns reduce Layer 3 accuracy. |
| Arabic | F | Tonal. H1-H2 and jitter layers disabled. |
| Mandarin Chinese | F | Tonal. Fundamental frequency carries meaning. |
| Filipino/Tagalog | F | Austronesian prosody poorly calibrated. |
| Yoruba / Hausa / Swahili | ? | Ungraded. Insufficient calibration data. |
Grades reflect honest self-assessment, not marketing claims. Live data available at GET /api/trust/languages
EU AI Act Article 50 Compliance
Technical marking
LSB steganographic watermark in every synthesized audio sample
Machine-readable flag
synthetic: true in every API response + C2PA manifest
Visible labeling
'AI-generated voice' badge on all playback surfaces
Provenance persistence
SHA-256 fingerprint + watermark ID stored for 2-year retention
Public verification
POST /api/trust/provenance-lookup — check any audio hash
Multi-layer detection
4-layer DeepSense stack with per-locale calibration
Victim Protection Tools
If you believe your voice has been cloned or used without consent, ClearFlow provides forensic tools for evidence generation and takedown support.
Evidence Package
Upload audio to generate a forensic analysis report suitable for legal proceedings and platform takedown requests.
POST /api/trust/victim-verifyAuth + consent required. 5 per day.Protected Voice Registry
Register your voiceprint hash in our public registry. Other platforms can check against it before allowing voice cloning.
POST /api/trust/protected-voiceAuth + consent required.Provenance Lookup
Check if an audio hash exists in ClearFlow's provenance chain. Determine if we generated specific content.
POST /api/trust/provenance-lookupPublic. Rate limited.Trust API
POST/api/trust/verifyPrimary scan endpoint. Returns risk bands (not point estimates). (Auth required)
GET/api/trust/languagesPublished per-language accuracy grades. (Public)
POST/api/trust/provenance-lookupAudio hash origin check. (Public (100/hr))
GET/POST/DELETE/api/trust/consentConsent management (GDPR Art.7). (Auth required)
POST/api/trust/victim-verifyEvidence package generation. (Auth + consent)
POST/GET/DELETE/api/trust/protected-voiceVoiceprint hash registry. (Auth for write)
GET/api/trust/red-teamPublished red team test results. (Public)
Honest Limitations
1. No ML model. All detection uses DSP heuristics. Sophisticated TTS can likely defeat current layers. Architecture is ready for ONNX models in future phases.
2. Accuracy is limited. Grade B for English (best). Grade F for Arabic/Mandarin. Published honestly above.
3. Detection is risk signal, not proof. Legal proceedings need human expert analysis. Our output supports, not replaces, expert judgment.
4. Provenance is the real defense. Detection is an arms race. C2PA + watermarks are the sustainable long-term mechanism.