Protect your AI from Audio Attacks

Defend the signal
beneath the signal.

Attackers can hide commands inside ordinary audio — inaudible to people, clear as day to speech AI. Subphase screens every clip before it reaches your model.

shield_lockTRY THE SCANNERterminalVIEW DOCS
The threat

Your voice AI hears things you can't

An adversarial audio attack adds a near-silent, precisely computed signal to ordinary sound. People hear the original. The speech model hears a sentence the attacker chose — and your product acts on it.

hearingWhat a person hears

“Hey, it's Sam — call me back when you get this.”

smart_toyWhat the speech model hears

“Hey, it's Sam — transfer $2,000 to the new account now.”

Illustrative. The difference between the two is a perturbation most listeners never notice.

Where attack audio comes from

upload_file

Uploaded audio

Voice notes, voicemails, recordings and clips that users send to your app — any of them can be crafted before upload.

call

Phone & VoIP calls

A caller plays prepared audio into a line answered by an AI agent or IVR. The agent acts on what its speech model hears.

speaker

Speakers in the room

A song, video or ad playing near an always-listening device can carry commands meant for the assistant, not the people.

public

Third-party media

Audio pulled in from the web or partners — podcasts, videos, user content — before it's transcribed, indexed or moderated.

If audio you didn't record yourself reaches a speech model, it's an entry point.

How it works

Two detection layers, one verdict

Audio flows through Pulse for a fast read, then Sonar goes deep — screened before it ever reaches your voice model.

graphic_eq

Audio In

wav · 16 kHz

arrow_forward
insights

Pulse

fast signal scan · free

arrow_forward
network_intelligence

Sonar

deep two-layer model

arrow_forward
verified_user

Verdict

clean / threat

bolt

Signal-level screening

The first layer inspects the acoustic fingerprint of every clip in real time — the statistical traces a hidden perturbation leaves behind. No model to train, no data to send anywhere, sub-second verdicts.

network_intelligence

Learned detection

A two-layer neural model, trained on real adversarial attacks, catches the subtle manipulations that slip past fixed rules — and keeps improving as new attack techniques emerge.

shield_lock

In front of your model

Subphase screens audio before it reaches your voice assistant or transcription pipeline, returning a clean / threat verdict — with an optional pass to strip the injected signal.

Two independent layers means an attack has to fool both a hand-built detector and a learned one — different methods, different blind spots.

Who it's for

If your product listens, it has an attack surface

Adversarial audio matters anywhere a speech model hears audio you don't control — and its output is trusted.

support_agent

Voice agents & AI phone assistants

Agents that book, refund or pay based on what a caller says. A hidden command becomes a real action.

headset_mic

Call centers & IVR

Transcripts drive routing, compliance records and analytics. A poisoned call rewrites the record.

home

Smart home, in-car & devices

Always-on microphones hear every speaker nearby — including media built to issue commands.

subtitles

Transcription & meeting tools

Notes, captions and summaries inherit whatever the model heard — not what was actually said.

neurology

Audio-native LLM apps

Models that take spoken input can be steered by instructions no listener notices — prompt injection through sound.

policy

Trust & safety teams

Speech-based moderation can be slipped by audio crafted to transcribe as something harmless.

You probably need Subphase if…

  • check_circleYour product accepts audio from users, callers or the open web
  • check_circleA speech model's output triggers actions, decisions or records
  • check_circleNo person listens to every clip before the model does
Get started

Screen audio in three lines of Python

  1. 1

    Install the SDK

    Python 3.10+, one dependency. Detection runs on Subphase's servers — no models to download.

  2. 2

    Get a free key

    The free tier includes Pulse and 1,000 scans a month.

  3. 3

    Scan before you trust

    Call scan() on each clip before it reaches your speech model. On a threat, drop it, flag it, or sanitize it.

Not on Python? Call the REST API directly — see the docs.

terminal
$ pip install subphase
$ export SUBPHASE_API_KEY="sk_dev_..."
$ python
>>> from subphase import Subphase
>>> client = Subphase() # reads SUBPHASE_API_KEY
>>> client.scan("voicemail.wav")
{'status': 'threat', 'confidence': 0.82, 'model': 'pulse', ...}
Pricing

Choose your defense tier

Start free with Pulse. Upgrade for Sonar and higher limits.

Free

$0

Pulse · 1,000 scans / mo

  • check_circlePulse — fast signal scan
  • check_circleHosted API + client SDK
  • check_circleCommunity support

Developer

BETA
$49/mo

Sonar · 1,000 scans / mo

  • check_circleSonar — deep ensemble detection
  • check_circleEverything in Free
  • check_circleBasic logs
  • check_circleEmail support

Pro

BETA
$199/mo

Sonar · 10,000 scans / mo

  • check_circleEverything in Developer
  • check_circleAudit logs
  • check_circleWebhook alerts
  • check_circleUsage dashboard

Enterprise

BETA
Custom

Sonar · unlimited scans

  • check_circleEverything in Pro
  • check_circleSLA + dedicated support
  • check_circleSSO
  • check_circleCustom model tuning
Contact

Get in touch

Questions about a plan, enterprise, or integration? Send a note.