Files
agent-estate-wiki/systems/phonon-tts.md
T
Tony0410 4a6e86948d wiki: 2026-11-14 nightly audit + corrections — Phonon TTS on CT200, Mnemosyne capture stall open
- new: systems/phonon-tts.md, incidents/2026-11-14-mnemosyne-capture-stall.md
- CT200 inventory re-verified: 64 non-seeder running / 72 total (delta = the TTS trio, no removals)
- fleet ledger: 11-14 = 167 rows / 166 completed / 0 failed; first red last_status since 11-13
  (Mnemosyne Health Monitor, error, streak 1); cron_incidents 0 -> 1 unclosed (138 rows)
- corrections pass: withdrew the 'unidentified extra container' claim; closed the alert-delivery
  question (delivered 00:03:39, for_failure=1)
- updated: index.md, current-state.md, log.md, systems/scheduled-tasks.md, systems/mnemosyne.md,
  infrastructure/docker-services.md
2026-11-15 00:42:01 +08:00

8.1 KiB
Raw Blame History

title, type, status, created, updated, verified_on, review_after, confidence, tags, sources
title type status created updated verified_on review_after confidence tags sources
Phonon TTS on CT200 (ptts-openai-server, ptts-ws-server, phonon-demo) system active 2026-11-14 2026-11-15 2026-11-15 2026-12-15 high
tts
phonon
ptts
docker
ct200
voice
huggingface
websocket
openai-api
kokoro
edge-tts
live-ct200-docker
live-ct200-docker-inspect
live-ct200-files
live-http-probe
live-state-db
live-agent-log
live-mnemosyne-db

Phonon TTS on CT200

A third TTS stack arrived on CT200 on 2026-11-14 — three containers that together form a self-hosted "ptts" (Phonon TTS) service: an OpenAI-compatible HTTP API (:8881), a static WebAssembly demo UI (:8882), and a WebSocket TTS server (:8883). It runs alongside the two voice services that host already carried (kokoro-tts :8880, edge-tts :5051).

Page created by the 2026-11-14 nightly audit from live probes (see Verification evidence). Everything below is a live read unless it says otherwise.

Containers (live docker ps / docker inspect, 2026-11-15 ~00:3x AWST)

Container Image Ports (host → container) Created (UTC) Last started (UTC) Restart policy
ptts-openai-server xn-ptts-openai-server:v0.7.1 8881 → 10200 (Wyoming also exposed, unmapped) 2026-11-14T02:31:56 = 10:31:56 AWST 2026-11-14T09:54:11 = 17:54:11 AWST unless-stopped
phonon-demo nginx:alpine 8882 → 80 2026-11-14T03:00:11 = 11:00:11 AWST 2026-11-14T03:00:11 unless-stopped
ptts-ws-server ptts-ws-server:0.7.1 (locally built, 116 MB) 8883 → 8080 2026-11-14T03:43:59 = 11:43:59 AWST 2026-11-14T09:54:21 = 17:54:21 AWST unless-stopped

Image names/tags for the first row are as reported by docker ps; the ws server's image is a local build (docker image ls ptts-ws-server → 0.7.1, no registry prefix), so it is not re-pullable — the build inputs are not in the wiki and were not located in this audit.

Endpoints (verified live from CT460 over the tailnet, 2026-11-15)

URL What answers
http://<CT200 tailnet IP>:8881/health health JSON, HTTP 200
http://<CT200 tailnet IP>:8881/v1/models OpenAI-shaped model list → gradium/phonon-40m-en
http://<CT200 tailnet IP>:8881/v1/audio/speech POST works — returned a valid MP3 (test voice nova)
http://<CT200 tailnet IP>:8882/ the demo UI — nginx:alpine, 19,212 B body, byte-identical in size to the file below
http://<CT200 tailnet IP>:8883/ HTTP 404 on a plain GET — expected: it is a WebSocket endpoint (/speech/tts)

The :8881 API is the piece that other services can consume (OpenAI-compatible /v1/audio/speech).

Files on CT200 (owner/paths verified by live ls -la)

Path Contents
/opt/ptts-openai-server/ run.sh (861 B, 2026-11-14 09:01 local) + .env (47 B, mode 600, 2026-11-14 09:00 local — holds the HuggingFace token for the gated model; value not recorded here)
/opt/phonon-demo/index.html the demo page, 2026-11-14 12:30 local; title "Phonon TTS - WebAssembly Demo"; talks to :8883 only (no 8881 reference)
/opt/phonon-demo/phonon-tts/ the WASM/JS assets (device.js, fetch.js, hub.js, gpu/, index.js, index.d.ts, LICENSE-APACHE) — i.e. the upstream phonon-demo front-end
/opt/ptts-voices/ 7 entries (2026-11-14 17:54 local): six .safetensors voices — anthony.safetensors, anthony-clone.safetensors, anthony-m.safetensors, anthony-november.safetensors, elizabeth.safetensors, grace.safetensors — plus a voices.jsonl manifest

The ws server logged num_voices=6 at 2026-11-14 04:42:52 UTC (12:42:52 AWST) and num_voices=7 at 09:54:24 UTC (17:54:24 AWST) — so the voice set was still being changed that evening. Provenance of the voice files is not established: the naming suggests cloned/derived voices, but nothing in the retained logs confirms whose they are or how they were produced.

Deployment timeline (2026-11-14)

AWST Event Source
08:35 – 08:45 First Phonon install session: ptts-openai-server brought up on :8881; the HuggingFace token returned 401 on the gated model and was replaced session transcript (state.db)
08:52 – 09:03 Second session: model switched to gradium/phonon-40m-en; /opt/ptts-openai-server/run.sh (861 B) and its 600 .env written at 09:00–09:01 session transcript + live ls
10:31:56 ptts-openai-server re-created (2nd recreation; the 09:01 container was not the live one) live docker inspect
11:00:11 phonon-demo (nginx:alpine) created → demo UI on :8882 live docker inspect
11:43:59 ptts-ws-server:0.7.1 created → WebSocket TTS on :8883 live docker inspect
12:30 /opt/phonon-demo/index.html last written live ls mtime
12:42:52 ws server logs num_voices=6 live docker logs
17:54:11 / 17:54:21 both ptts servers restarted together; /opt/ptts-voices written; ws logs num_voices=7 live docker inspect + ls + logs

⚠ Provenance gap. The two retained interactive sessions in state.db cover only 08:35–09:03 AWST. Every change from 10:31 onward — the demo UI, the ws-server build, the voice files, the 17:54 restart — is not attributable to any retained transcript: the only sessions active in state.db between 09:03 and 23:59 on 11-14 are cron runs and a heartbeat "Pulse" session (495 messages, none containing docker/TTS content). Working hypothesis, not proven: the later work was done by another estate agent or a session not retained in this profile's store. Do not treat the 08:35/08:52 sessions as the whole of this deployment.

Why it is on 8881 and not 8880

8880 was already taken by kokoro-tts on the same host (ghcr.io/remsky/kokoro-fastapi-cpu:v0.2.1, declared in the ReadLater compose file → infrastructure/docker-services), and edge-tts sits on :5051. Phonon is therefore the third voice backend on CT200 — three TTS stacks on one LXC.

Operations

# reach CT200 (the PVE node is 100.96.100.82; the wiki's CT200 row carries the tailnet IP)
ssh root@100.96.100.82 'pct exec 200 -- docker ps --format "{{.Names}}\t{{.Status}}\t{{.Ports}}" | grep -Ei "ptts|phonon"'
# logs (the ws server is the talkative one)
ssh root@100.96.100.82 'pct exec 200 -- docker logs --tail 40 ptts-ws-server'
# smoke test the OpenAI-compatible path from another host on the tailnet
curl -s -X POST "http://<CT200 tailnet IP>:8881/v1/audio/speech" -H 'Content-Type: application/json' \
  -d '{"model":"gradium/phonon-40m-en","input":"test","voice":"nova"}' -o /tmp/probe.mp3 && file /tmp/probe.mp3

Open items / watch

  1. The HuggingFace token is a standing single point of failure. Both ptts-openai-server and ptts-ws-server serve a gated model, and the token has already expired once (2026-11-14 08:35–08:45 → replaced). The token lives in ~/.hermes/.env on CT460 and in /opt/ptts-openai-server/.env (mode 600) on CT200; whether the ws container reads the same file was not confirmed in this audit. A silent expiry would take the whole stack down again.
  2. ptts-ws-server:0.7.1 is a local-only image — no build source recorded here.
  3. Attribution gap for everything after 09:03 AWST (see the timeline note).
  4. Nothing in the estate consumes :8881 yet, as far as this audit could see: it was verified by probe, not by a caller. If it is wired into Hermes TTS later, the provider config belongs in systems/hermes-agent / the TTS configuration, not here.
  5. The upstream project's own naming is inconsistent (ptts-* containers, phonon-* model/page, a "gradium" publisher prefix) — the same stack appears under three names.