feat(web): real-time voice chat with Live Mode

[crrow]

Description

Add Gemini Live-style voice conversation to the desktop app:

• Web Speech API for real-time STT (browser-side, zero latency)
• LiveKit Agents UI components for voice UI (waveform, controls)
• Server-side TTS (OpenAI) with streaming audio_delta events
• Barge-in support (interrupt Rara mid-speech)

Frontend-first: Steps 1-3 work without backend changes (voice input → text → chat).
Steps 4-7 add TTS backend for voice responses.

Design doc: docs/plans/2026-04-13-realtime-voice-chat.md

Component

web (frontend, UI)

Alternatives considered

• LiveKit full stack (Server + Agent): too heavy, adds unnecessary infrastructure
• Browser speechSynthesis for TTS: robotic voice, not production quality
• Server-side Whisper STT: 1-2s latency vs real-time Web Speech API

[m13v]

the Web Speech API for STT is the right call for zero-latency input, but be ready for browser inconsistencies. Chrome gives you interim results (partial transcripts that update in real time), Firefox does not. Safari on macOS uses Apple's speech framework underneath and has different accuracy characteristics. for barge-in support, the key is detecting speech onset fast enough to interrupt TTS playback. we found that VAD (voice activity detection) in the browser needs a threshold calibration step, otherwise ambient noise triggers false interrupts constantly. the LiveKit UI components are solid for the waveform visualization, but their server-side TTS path adds 200-400ms of latency that makes conversation feel sluggish. if you're already using OpenAI TTS, their streaming endpoint with audio_delta events can cut that to under 100ms for first audio byte.

[m13v]

our voice pipeline with VAD calibration, streaming transcription, and the audio processing flow: https://github.com/m13v/fazm/blob/main/Desktop/Sources/TranscriptionService.swift

fazm is an open source macOS AI agent that uses voice as the primary input method. the streaming transcription and barge-in patterns might be useful as reference.