Streaming TTS - speak as tokens arrive, not after the full reply

[Dix01]

Problem: TTS waits for the full LLM reply before speaking, adding noticeable latency. JARVIS should start talking on the first sentence while the rest still streams - the way the films feel.

Where:

• Backend TTS: jarvis/plugins/voice_tools_optional.py (Piper / Riva / Edge tiers)
• WS streaming: jarvis/server/ws.py
• Frontend playback: web/src/lib/voice.ts

Approach: Chunk the assistant stream on sentence boundaries (. ! ? / newline). Synthesize + enqueue audio per sentence so playback starts after sentence #1. Keep an ordered queue.

Acceptance:

• First audio plays before generation finishes
• Sentences play in order, no overlap
• "cancel" still stops the whole queue

Difficulty: medium.