System rebuild in progress — dataset pipeline and model training being restructured · May 2026

Real-time voice fraud detection

Detect AI voice
fraud on live calls.

Voice deepfakes are becoming more realistic and accessible every day. PulmoForge detects them in real-time.

Current status · May 2026
Rebuild Phase
Dataset pipeline
Rebuilding
Model training
In progress
Public beta model
Upcoming
Demo (prior build)
Available
What's happening and why.

rebuild phase

Dataset + training pipeline restructure underway

STEP 01
Dataset reset

Rebuilding the dataset from only commercially viable datasets. Earlier contaminated Datasets are being removed from the training pipeline.

In progress
STEP 02
Updated Dataset

New better datasets are being added and outdated/flawed datasets are being removed for maximum effects from the training pipeline.

In progress
STEP 03
Model retraining

Full retraining of model under a cleaner pipeline and different techniques. Beta version will be released after this.

Upcoming

Earlier iterations included synthetic speech sources that were not fully verified for licensing compatibility or distribution consistency — introducing risk of dataset contamination and evaluation drift. The system is being reset to a clean training baseline so that the next public model reflects results we can fully stand behind.

How we compare to other detectors
System Meeting audio Compression trained Status
PulmoForge ✓ Zoom + Opus ✓ Yes ⟳ Rebuilding
Legacy detectors ✗ Mostly lab audio ✗ Rarely Modern gen AI bypasses
Enterprise fraud tools ✓ Phone systems Unknown Closed platforms

Most voice fraud systems were not built around Zoom compression pipelines. PulmoForge was trained directly on compressed meeting audio.

What the system does.
Codec resistance
Codec-resistant detection Trained on Zoom's Opus 32kbps compression pipeline. Detects synthesis artifacts that survive real-world codec degradation — where most detectors fail.
False positive rate
Low false positive rate Designed to minimize false accusations on genuine voices. Results are probabilistic — intended to assist, not replace, human review.
Architecture
WavLM backbone Fine-tuned on large-scale speech data. Attention pooling isolates subtle synthesis artifacts across codec-compressed audio. Architecture unchanged through the rebuild.
Inference mode
Live call monitoring Per-second risk scoring on rolling windows. Built for real-time calls and meetings. Will return post-rebuild with retrained model weights.

Zoom destroys the signals most detectors depend on.

Opus compression removes many high-frequency artifacts that traditional detectors rely on. PulmoForge was trained directly on codec-degraded speech to learn synthesis patterns that survive real-world meeting conditions.

Built for actual calls — not clean studio audio.

What PulmoForge detects
  • ElevenLabs voice clones through Zoom
  • RVC voice conversion
  • XTTS neural TTS
  • Real-world impersonation attempts
Status — May 2026

PulmoForge is in a controlled rebuild. The public beta model is being retrained on a clean, fully verified dataset. Free beta access will be available once the new model passes internal evaluation — join the waitlist to be notified first.

Run Current Demo → Join Waitlist