← Work indexexperiment

Case study · 2026

Voice Clone Agent

A real-time assistant that combines native interaction, speech models, GPU services, and an explicit own-voice consent boundary.

Role
System designer and application engineer
Outcome
A native macOS client and self-hosted speech pipeline with tested handoff behavior, evaluation tooling, and consent enforced on every synthesis request.
Capabilities
Applied AI · Native applications · Real-time media
Stack
Swift · Python · WebSockets · GPU inference

Generated output

Hear generated evaluation samples

Evaluation audio only · no reference recordings

  1. Generated English greeting

    Hello, this is a generated evaluation sample from the real-time voice assistant.

  2. Generated technical response

    The client streams one session at a time while the renderer and audio handoff remain serialized.

Case record

The problem

A convincing voice assistant has to feel continuous across listening, reasoning, speaking, interruption, and presentation. Improving one model does not solve the handoff problems between those stages.

The constraints

  • Speech and rendering services have different latency and lifecycle behavior.
  • The native client must recover cleanly from interruption and reconnection.
  • GPU capacity is shared and cannot support an unrestricted public endpoint.
  • Voice cloning is limited to the owner’s consented profile.

Pivotal decisions

The system separates orchestration, transcription, language generation, speech synthesis, and optional presentation services behind explicit protocols. The macOS client owns the user-facing lifecycle while renderer handoff is serialized to prevent overlapping work from corrupting presentation state.

Consent is not a disclaimer. A self-consent record is required and revalidated for every synthesis request, and the public case study exposes generated evaluation output rather than the private reference recordings used to construct the profile.

What it demonstrates

Emerging model capability becomes a product only when latency, ownership, interruption, privacy, and failure are treated as first-class design inputs.