Docs · Speech to speech

Source docs/STT-TTS-ARCHITECTURE-20261007.md in the ZoeyOS source. Last verified 2026-10-07. State current.

Speech to speech: the ear, the bridge and the voice

As they were running on 2026-10-07, read from the services themselves. The team's own notes were checked and not used as the source.

The path

room microphone (reSpeaker beam array), raw capture
        │
        ▼
ear rail :8190          voice activity (TEN VAD), the WeSpeaker speaker stamp
        │
        ▼
recogniser :8096        Audio8 ark-asr 0.6B, int8 ONNX, on the CPU, its own environment
        │
        ▼
ear bridge :8194        the turn, the quiet wait, the speak; the reply brain is Zoey's gateway → ops :8000
        │
        ▼
voice :7789             Chatterbox Turbo, 24 kHz PCM, watermark on
        │
        ▼
PipeWire                the speakers

announcer, separate: Audio8's TTS runtime :8024. Not the conversation voice.

Health at measurement, idle room: the ear on the beam microphone, TEN VAD at 0.5 (Silero installed beside it at 0.35, not the live switch), WeSpeaker as the voice encoder. The bridge waits two seconds of quiet before she speaks and flushes fast at 0.3 s.

The units

voice	Chatterbox Turbo on :7789. Replaced CosyVoice the morning of 2026-10-07; the contract (health, stream, abort) did not move, so the bridge kept up without a restart.
recogniser	the ear worker on :8096, its own Python environment
ear rail	capture, VAD, stamp, HTTP to the recogniser, on :8190
ear bridge	hears a turn, asks Zoey, speaks through :7789; abort on :8194
announcer	Audio8's TTS runtime on :8024, for house lines, never her voice
ops	the reply brain: vLLM 0.30.0, Qwen3.8-27B-NVFP4, :8000
crew	the other lane: vLLM 0.27.1, the same weight family, :8005

Versions read from the environments

Chatterbox	chatterbox-tts 0.1.7, MIT, Resemble AI; Python 3.14.7, torch 2.14.1, transformers 5.2.0
Perth watermark	resemble-perth 1.0.1, on
TEN VAD	1.0.6.8, Apache 2.0 with additional conditions; one pitch file carries LPCNet's BSD code
Silero VAD	6.2.1, installed, not the live switch
WeSpeaker	0.0.1, Apache-2.0
ear worker	Python 3.12.13, numpy 1.26.4
PipeWire	1.6.2

Two rules the ear keeps

The ear skips the house's own voice before recognising. A line from a bench, or the announcer on air, is dropped before the recogniser sees it. A recogniser restart is an ear restart: the ear hears nothing while it is down.
Length is coached, not capped. The spoken-length cap was reverted 2026-10-06; the running bridge carries no cap.

What this page does not decide

Loudness, the mouth tiles and the sheet lead live with the mouth instruments. This page is the ear, the bridge and the voice as they were running on 2026-10-07.