Iso-Vox: Solving the Cocktail Party Problem in Voice AI

Iso-Vox is the Jobix target speaker extraction layer. It isolates the person on the call from background speech and noise before any transcription happens.

The cocktail party problem

Standard speech-to-text assumes one clean speaker. Real calls do not look like that: a prospect answers from an open-plan office, a car with road noise, a call centre floor, or a kitchen with a television on. When several voices overlap, conventional transcription blends them, and the AI agent responds to words the caller never said.

How target speaker extraction works

Iso-Vox builds a short voice profile from the first moments of the call, then filters the incoming audio stream against that profile in real time. Competing speakers and steady background noise are suppressed before the audio reaches the transcription model, so the agent only ever reasons over the words of the person it is actually talking to.

Why it matters for latency

Filtering has to be cheap or the conversation stops feeling human. Iso-Vox runs inline in the audio path rather than as a separate post-processing pass, which keeps end-to-end voice latency under the 500ms average the platform targets. Cleaner input also reduces retries and clarification questions, which shortens calls.

Where it is applied

Iso-Vox is enabled by default on every Jobix voice agent, inbound and outbound, across all supported languages. It is most visible on consumer outbound calling, field-service inbound, and multilingual support queues, where background speech is the norm rather than the exception.