DeepAudioGuard
Know whether the voice in a recording is real or AI-generated. It listens to an audio or video file, examines it the way a forensic audio expert would, scores it with a neural network trained to recognise AI-cloned voices, and returns one evidence-backed verdict — Verified Authentic or Synthetic Deepfake — with a downloadable report explaining why
What it is
AI-powered deepfake audio forensic analysis
Imagine a lie-detector, but for voices instead of people. DeepAudioGuard takes a recording — a phone call, a video clip, a voice note — breaks it into small pieces, and examines each piece the way a forensic audio expert would: checking for unnatural pitch, breathing, silence, and the digital fingerprints that AI voice generators leave behind. It also runs the recording through a neural network trained specifically to recognise AI-cloned voices.
In Standard Mode the file is chopped into four chunks and each chunk is scored in parallel by both the signal-processing checks and the neural network running on the local machine, CPU or GPU. In Enhanced Mode, where a Forensic Verdict Engine API key is configured, the raw chunk-level metrics are additionally sent to a large-language-model synthesis step that writes the clinical-style narrative and reconciles any disagreement between the DSP checks and the neural score.
It is built for people who need to double-check a suspicious voice: journalists verifying a leaked recording, security teams screening a phone call, or everyday users who received a strange voice message and want a second opinion.
It fuses signal-processing science with a trained neural network into one clear, evidence-backed verdict.
- Signal layers fused into one verdict — DSP forensics, an acoustic neural model and language-model synthesis
3
Signal layers fused into one verdict — DSP forensics, an acoustic neural model and language-model synthesis
- Chunks scored in parallel by both the DSP checks and the neural network
4
Chunks scored in parallel by both the DSP checks and the neural network
- Underlying DSP measurements exposed — phase discontinuity, spectral centroid drift, jitter and the rest
7
Underlying DSP measurements exposed — phase discontinuity, spectral centroid drift, jitter and the rest
- Standard Mode runs with no external API calls once the model checkpoint is installed
Offline
Standard Mode runs with no external API calls once the model checkpoint is installed
How it works
From an uploaded file to a clinical verdict
01
Upload
Drop in an .mp4 or .wav file
02
Extract & chunk
The audio is extracted and split into four pieces
03
Parallel analysis
DSP forensics and the acoustic neural model, side by side
04
Synthesis
The verdict engine merges all signals
05
Verdict
Authentic or synthetic, with a report
Product exclusiveness
What sets DeepAudioGuard apart
Triple-signal fusion
Combines DSP forensics, a fine-tuned acoustic neural model and a language-model synthesis layer into a single verdict; most competitors ship only one or two of these layers.
Flagship capability · three layers, one verdict
Fully self-hostable and offline-capable
The core detection pipeline — DSP plus the acoustic model — runs with no external API calls, unlike most commercial tools, which are cloud-only SaaS.
No external calls · not cloud-only
Real-time progressive streaming
Chunk-level results appear live via Server-Sent Events instead of waiting for a single end-of-file score.
Server-Sent Events · results as they land
Forensic-grade PDF reporting
The report is generated entirely client-side — no data leaves the browser for report creation.
Client-side · nothing leaves the browser
Transparent, inspectable metrics
Exposes the seven underlying DSP measurements — phase discontinuity, spectral centroid drift, jitter and the rest — rather than a single opaque score.
Seven measurements · not one opaque score
Clinical dashboard UX
A purpose-built dark forensic theme with animated metric bars and timelines, closer to a diagnostic tool than a generic upload form.
Metric bars · timelines · diagnostic theme
Free and open
No per-scan credit system and no seat-based enterprise pricing, which most direct competitors rely on.
No per-scan credits · no seat pricing
How it compares
Measured against the two kinds of tool people buy instead
Dedicated audio deepfake focus
- Enterprise voice-security platforms
- Yes — Audio specialist
- Consumer self-serve detectors
- Partly — Often video & image too
- DeepAudioGuard
- Yes — DSP + acoustic model + synthesis
Real-time call screening
- Enterprise voice-security platforms
- Yes — Call-centre grade
- Consumer self-serve detectors
- No — Upload only
- DeepAudioGuard
- No — File and chunk based
Runs self-hosted and offline
- Enterprise voice-security platforms
- No — Cloud-only SaaS
- Consumer self-serve detectors
- No — Cloud-only SaaS
- DeepAudioGuard
- Yes — DSP + acoustic model, no external calls
Underlying metrics exposed
- Enterprise voice-security platforms
- No — Single score
- Consumer self-serve detectors
- No — Single score
- DeepAudioGuard
- Yes — Seven DSP measurements
Chunk-level results while it runs
- Enterprise voice-security platforms
- No — End-of-file score
- Consumer self-serve detectors
- No — End-of-file score
- DeepAudioGuard
- Yes — Server-Sent Events
Report built without sending data out
- Enterprise voice-security platforms
- No — Generated server-side
- Consumer self-serve detectors
- No — Generated server-side
- DeepAudioGuard
- Yes — Client-side PDF
Access model
- Enterprise voice-security platforms
- No — Enterprise / custom pricing
- Consumer self-serve detectors
- Partly — Free tier, then paid
- DeepAudioGuard
- Yes — Free, self-hosted
The architecture
Four layers — one path from a file to a verdict
Only layer three runs side by side, and it runs four chunks at once
Read top to bottom to follow one file through the system.
Client
- React + Vite dashboardAnimated metric bars
- Live chunk timelineServer-Sent Events
- PDF report exportGenerated client-side
API
- REST APITakes .mp4 and .wav
- /api/analyze-chunkedStreams results as they land
- Audio extractionAudio pulled from video
Parallel analysis engine
- DSP forensicslibrosa, scipy — seven measurements
- Acoustic neural modelLoaded once, cached
- ThreadPoolExecutorFour chunks in parallel
Decision
- Forensic decision synthesis engineReconciles DSP and neural score
- Clinical verdictAuthentic or synthetic
Deployment strategy
Three ways to run it — laptop, container or GPU host
Local / developer deployment
Run the API backend and the web frontend as two local processes for development and demos. The trained checkpoint, roughly 407 MB, is stored outside version control and fetched from a release artifact or object storage bucket at setup time.
Containerized deployment
Package the API, the model checkpoint and the audio tooling into a single container image, and serve the web build as static files behind a reverse proxy or from the same container. Use a GPU-enabled base image on a GPU host, and fall back to a CPU image for lower-traffic or cost-sensitive environments.
Cloud / production deployment
Host the API on a GPU-backed compute service so acoustic inference stays fast under load, and put a queue in front of long-running chunk analysis so SSE connections stay responsive during traffic spikes. Serve the frontend from a CDN, keep the API behind HTTPS with the Forensic Verdict Engine key in a secrets manager, and add observability on request latency, chunk-processing time and model load status.
Under the hood
What the platform is made of
The two modes it runs in, what it accepts, and the checkpoint the neural score depends on.
- Standard Mode
- Local analysis. A file is uploaded, chopped into four chunks, and each chunk is scored in parallel by both the signal-processing checks and the neural network running on the local machine, CPU or GPU. This mode works fully offline once the model checkpoint is installed.
- Enhanced Mode
- Cloud synthesis. With a Forensic Verdict Engine API key configured, the raw chunk-level metrics are additionally sent to a large-language-model synthesis step, which writes the clinical-style narrative and reconciles any disagreement between the DSP checks and the neural score into one human-readable verdict.
- Inputs and extraction
- The documented pipeline accepts .mp4 and .wav files. Where the input is video, the audio is extracted before analysis begins.
- The model checkpoint
- The trained checkpoint, roughly 407 MB, is kept outside version control and placed in the checkpoints directory — either trained with the included notebook or obtained pre-trained.
Frequently Asked Questions
Yes for Standard Mode — the DSP forensics and the local acoustic model do not require external calls. Enhanced Mode's narrative synthesis step does require an internet connection and an API key.
The documented pipeline accepts .mp4 and .wav files; audio is extracted from video before analysis.
The tool produces a detailed forensic report with metrics and timestamps intended to support an investigation, but admissibility depends on jurisdiction and how the evidence chain is handled — this should be treated as an analytical aid, not a substitute for legal or expert forensic testimony.
A GPU is recommended for fast inference; the model will run on CPU as a fallback, at reduced speed.
Those tools are cloud-hosted platforms aimed at enterprises and often priced per scan or by contract. DeepAudioGuard is openly available and self-hostable, trading enterprise call-centre integrations for transparency and zero licensing cost.
It is not bundled with the distribution, due to size. You either train your own using the included notebook, or obtain a pre-trained checkpoint and place it in the checkpoints directory.
No detector is perfect. Industry testing shows accuracy drops on compressed, noisy or re-recorded audio, so a synthetic verdict should be treated as strong evidence rather than absolute proof, especially for high-stakes decisions.
The rest of the line
More in SyntheticMediaGuard
Put a suspicious recording through the pipeline
Book a walkthrough and we will show the four chunks scored in parallel, the metrics behind the verdict, and the report it leaves behind.