Skip to content
Maskflare
Talk to usStart free trial
Menu

Voice AI

Mask Personal Data in Realtime Voice AI, With Honest Limits

A voice agent on the OpenAI Realtime API takes a call. The caller spells out an email address and reads a card number. Some of that travels as audio, some as text events in the same session.

Maskflare's realtime route masks the text events before they reach the provider. It does not mask audio. That boundary decides whether this fits your use case, so it is stated up front.

Start free trial Discuss this use case
mask path
  1. 01

    Request from your application

  2. 02

    Sensitive values masked

  3. 03

    Forwarded to the AI provider

Why the usual approach falls short

Voice agents send audio and text over one live session. Masking the text events cuts what the provider gets in transcripts and tool calls; audio is not masked.

The Maskflare approach

Connect your realtime client to the Maskflare gateway instead of the provider. Text events in the session, such as text input and instructions, are masked with your rules before they are forwarded. Audio frames pass through unchanged, so anything the caller says out loud reaches the provider as audio. People also say identifiers differently than they type them: "jane dot doe at acme dot com". Spoken-form detection, which catches that pattern in transcribed text, is available as an opt-in setting because it raises false positives in ordinary speech.

Benefits

Text events masked live

Text input and instructions in the realtime session are masked before the provider sees them.

Spoken forms, opt-in

Detection for spoken-style identifiers such as "jane dot doe at acme dot com" can be switched on where it pays off.

Clear boundary

Audio is not masked. You can decide with that known, instead of finding out in a review.

Same rules as text AI

Voice routes share the rule set used for chat and embeddings in the same environment.

Built on

Maskflare Detection Engine →Maskflare AI Gateway →Maskflare Agent & MCP Security →

Understand the underlying risks

Frequently asked questions

Is audio masked?

No. Only text events in the realtime session are masked. If a caller reads out a card number, the provider receives it in the audio.

Then what is the point?

A lot of the sensitive data in a voice agent moves as text, for example instructions that carry the caller's account details. Masking those text events cuts what the provider receives, even though audio still carries what is spoken.

What is spoken-form detection?

Rules that recognize identifiers written the way people say them, such as "jane dot doe at acme dot com". It is opt-in because phrases like that can also appear in ordinary speech.

Which providers are supported?

The OpenAI Realtime API through the gateway's realtime route. Check with us for other live-voice APIs before you plan around them.

What about recorded calls?

Transcripts sent to a model as text are masked like any other prompt. See the customer support AI page for that workflow.

See it on your own data

Try Maskflare free.
Bring your hardest prompts.

28-day free trial or 10,000 requests, whichever comes first. Point one app at the gateway or enroll a few laptops,
and watch what gets masked. Or talk to us first.

Start free trial Talk to us