Voice UI

Voice UI is AR Labs' generic browser voice interface: the reusable part of Salesbot, separated out so that other AI agents (language models run in a loop with tools) can have a voice. It is for anyone building a browser voice agent on Pipecat. We took it from the first working version of Salesbot, our browser voice agent for sales discovery, and removed everything specific to sales: the conversation stages, the progress display and the branding. What remains is a browser page with a connect button, microphone toggle, an audio-reactive visualiser, a live transcript and an event log, connected to a Pipecat backend that listens, thinks and speaks.

flowchart LR
  P["Browser page"] <-->|"WebRTC"| TR["Transport: local or Daily"]
  TR <--> B["Pipecat backend"]
  B --> S["Deepgram speech-to-text"]
  S --> L["Configurable language model"]
  L --> T["Cartesia text-to-speech"]
  T --> B

The backend uses Silero for voice-activity detection and can run with two transports: a direct WebRTC connection for use on one machine, with no hosted service needed, or Daily's hosted WebRTC rooms when the browser and the backend are on different machines. The language model, the agent's instructions and the voice are settings, so the same interface can front a different agent without code changes. A short research note alongside it surveys how voice platforms let an outside agent supply the replies while the voice service handles audio, transcription and turn-taking.

Voice UI is a snapshot rather than a finished package. The extraction into a reusable library has not been completed, and the work is paused. Salesbot has moved well beyond the version this was taken from. VOX Presenter, our voice sessions for discussing documents, follows the same browser and Pipecat pattern.