How it works
The fence sits between the microphone and ASR: everything outside the accepted zone is discarded before a single word is transcribed.
What a voice fence is
A voice fence is an audio front-end that decides whose speech is allowed into the translation pipeline at all. It is not a translation feature — it runs before recognition, discarding audio that does not belong to the people in your conversation.
It is separate from conversation mode. Conversation mode manages turns and language directions; the fence simply decides what counts as speech worth processing. An app can have one without the other, but the two are most useful together.
Why gating matters
Speech recognition fails differently from typing: in a market or restaurant the microphone hears the tables next to you, background music, and passers-by. Without gating, fragments of other people's conversations get transcribed and translated as if you had said them — confusing at best, embarrassing at worst.
Because everything downstream depends on the audio it receives, a bad front-end caps the quality of the whole chain. Better gating usually improves perceived translation accuracy more than a better translation model does.
How the gate decides
Fences combine several signals. Beamforming across the phone's microphones estimates the direction a sound arrived from, so speech from behind or to the side can be attenuated. Loudness and reverberation cues estimate distance, filtering out far-away voices. Voice enrollment — a short sample recorded once per speaker — lets the app accept only the enrolled voices and reject everyone else, which is the strictest and most reliable form of fencing.
Apps built around noisy-environment use, such as PolyPort.chat with its voice-enrollment approach, treat the fence as a core feature. General-purpose apps typically rely on the operating system's generic noise suppression instead, which is tuned for phone calls rather than for rejecting a second human voice.
Limits and trade-offs
A tight fence can cut off a quiet speaker, someone leaning back, or a third person who legitimately joins the conversation. A loose fence lets bystanders in. Most apps expose no control over this, so the practical workaround is physical: hold the phone closer to the speaker and angle it away from the noise source.
Frequently asked questions
Is a voice fence the same as noise cancellation?
No. Noise cancellation removes non-speech sound. A voice fence rejects speech too — specifically speech from people who are not part of your conversation.
Do I need to enroll my voice?
Only for apps that use voice-profile fencing. Direction- and distance-based fences work with no setup, but are easier to fool in a crowd.
Does the fence run on the phone or in the cloud?
On the device, almost always. It has to run before audio is uploaded, both for latency and because sending rejected audio to a server would defeat the point.
