01 · Origin
threshold ≥ 0.92
WavLM SSL file embedding
AI vs human source indicator
DeepFakeDetection
Voice integrity screening for the AI audio era.
Multi axis analysis for synthetic speech, replay, channel artifacts, and partial fabrication, built for human review.
Product
Screen synthetic speech with evidence, not a single score
DeepFakeDetection analyzes uploaded or recorded audio across four independent integrity axes: voice origin, replay and rerecording, mixer and channel processing, and partial fabrication segments. It then returns structured evidence for human review.
The live product at deepfakedetection.dev runs the same Phase 9 inference stack used in our research pipeline: decode, segment, extract features (including WavLM SSL embeddings), score each axis, and fuse results with safe forensic wording.
Outputs are experimental screening indicators for research, media verification, and awareness workflows, not legal proof of authenticity or fraud. Manual review is always recommended.
Architecture
Four evidence axes
Each model reports separately. Fusion explains agreement and conflict. It does not invent a court ready authenticity verdict.
02 · Replay
threshold ≥ 0.65
File-level acoustic features
Rerecording, not AI-generated
03 · Mixer / channel
threshold ≥ 0.75
File-level acoustic features
Processing effects, not AI-generated
04 · Partial segments
threshold ≥ 0.95
Segment acoustic + SSL + localization
Candidate spans for manual review
- 01
Decode & normalize
Load audio as 16 kHz mono for consistent feature extraction.
- 02
Segment the file
Slide 4 s windows with 2 s hop to score short spans for partial-fabrication cues.
- 03
Extract features
File-level acoustic + WavLM SSL embeddings; per-segment acoustic, SSL, and localization features.
- 04
Run four models
Origin (SSL), replay (acoustic), mixer/channel (acoustic), and partial segment (combined). Each axis stays separate.
- 05
Fuse & explain
Phase 8F multi-axis fusion, candidate segments, safe forensic summary, and manual-review recommended wording.
Use cases
Built for real audio workflows
Pair automated evidence with human judgment. Explore public case studies for how multi axis screening behaves on real world clips.
View case studies01
Media verification
Screen interview clips and viral audio before publication. Review multi axis evidence alongside editorial judgment.
02
Security awareness
Demonstrate voice cloning risk in workshops with live analysis, not canned demos or single fake or real scores.
03
Call center triage
Flag suspicious recordings for human review. Designed to support operators, not automate legal decisions.
04
Research and evaluation
Reproduce multi axis runs via the live lab or API, inspect features, and compare axis outputs in structured reports.
Research to product
From forensic audio R&D to a cloud voice integrity platform
Developed from academic R&D in forensic audio analysis and now being built into a cloud based voice integrity platform.
Working product
Signed in users can upload, record, or generate lab audio and run the live multi axis analysis pipeline today.
Domain: deepfakedetection.dev
Contact: founders@deepfakedetection.dev
Run a real analysis in minutes
Create an account, open the analysis lab, and screen a recording through the same multi axis engine powering our public demo.