Media verification
Screen interview clips or viral audio before publication; pair with editorial review.
FYP 2025Phase 9 · Four evidence axes
DeepFakeDetection
Synthetic speech screening with four separate integrity checks, built for demos, research, and thesis defense.
Signal view
WavLM SSL · acoustic features
LIVE VISUALIZER · UI PREVIEW
Active models
4 axis indicators (Phase 9B)
Origin axis
WavLM SSL · origin_file_model
Replay axis
Acoustic · replay_file_model
Mixer / channel
Acoustic · mixer_file_model
About the system
A complete pipeline from raw audio to explainable verdict
Multi-axis voice integrity checks (origin, replay, channel, and partial segments) with experimental evidence indicators only.
The web dashboard calls the Phase 9 FastAPI service in new backend/release/No mock scores. Each upload runs decode, segmentation, four axis models, and multi-axis fusion with safe forensic wording.
Outputs are experimental evidence indicators for research and demos, not legal proof of authenticity or fraud. Manual review is recommended.
Architecture
Model & inference at a glance
Values match the Phase 9 release backend: model registry, segmentation defaults, and candidate thresholds.
Phase 9B · experimental
Four experimental evidence models
Origin (WavLM SSL), replay (acoustic), mixer/channel (acoustic), and partial segments (combined features), fused into separate indicators, not one binary fake/real verdict.
release/models/ · joblib + metadata
Origin
WavLM SSL file embedding
AI vs human source indicator
Replay
File-level acoustic features
Rerecording, not AI-generated
Mixer / channel
File-level acoustic features
Processing effects, not AI-generated
Partial segments
Segment acoustic + SSL + localization
Candidate spans for manual review
SEGMENTATION
4 s windows · 2 s hop @ 16 kHz
segmentation.pyOUTPUT
Separate evidence axes + fusion status
/analyzeAxis thresholds (candidate)
Origin threshold0.92 (candidate)
Replay threshold0.65 (candidate)
Mixer threshold0.75 (candidate)
Partial threshold0.95 (candidate)
Validation · Phase 9C / 9D
Phase 9B models are experimental forensic indicators. Phase 9C smoke tests pass on release sample cases; broader Phase 9D batch validation is documented separately.
- 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.
Platform
What you can demonstrate today
01
Four independent checks
Phase 9Voice source, replay signs, channel/mix effects, and edited segments are scored separately, not collapsed into one fake score.
02
SSL origin model
WavLMWavLM embeddings power the AI-vs-human origin axis with calibrated evidence bands in the UI.
03
Segment highlights
SegmentsPartial module surfaces timestamp candidates on the waveform for optional manual listening.
04
Live web demo
IntegratedNext.js dashboard uploads audio to the Phase 9 FastAPI service in `new backend/release/`.
05
Safe forensic wording
SafetyExperimental prototype language, manual review flags, and no court-ready verdict claims.
06
History (optional)
FirestoreSigned-in users can save summary results to Firebase when rules allow.
Use cases
Where teams apply screening
Always pair model output with human judgment. Suitable for FYP demos, workshops, and media workflows.
Real-world case studiesSecurity awareness
Demonstrate voice-cloning risk in workshops using live multi-axis evidence, not canned demos.
Research & FYP
Reproduce Phase 9 release runs via CLI or API and compare axis outputs in JSON reports.
Call-center triage
Flag suspicious recordings for human review, not automated blocking or legal verdicts.
Upload real audio. Get a real report.
Sign in, open the forensic lab, and run your file through the hybrid model: explanations, thresholds, and multiclass hints included.