ReduxStation TTS gateway

Self-hosted TTS service backing SStts on the game side. Wraps Piper (VITS + ONNX on CPU) behind an aiohttp gateway and exposes the HTTP contract the game-side subsystem expects.

Companion game-side code lives in ReduxStation/ReduxStation:

• code/controllers/subsystem/tts.dm (the subsystem)
• code/datums/http.dm (rust-g HTTP wrapper)
• code/controllers/configuration/entries/tts_config.dm (config entries)

What it does today

• Hosts 2 Piper voice models: en_GB-vctk-medium (109 UK English speakers, the "comical" tg-style sound) and en_US-libritts_r-medium (904 American LibriTTS-R speakers for raw variety). About 1013 voices total.
• Implements GET /tts-voices, GET /pitch-available, GET /tts, GET /tts-blips, GET /health-check matching the API contract SStts uses.
• Returns ogg/vorbis audio with an audio-length: HH:MM:SS.fff header.
• Auth via the Authorization header, constant-string compared, defaults to coolio (override in production).

What is on the roadmap

• ffmpeg filter chain (radio click splicing, silicon convolution, mask voice filters). The gateway currently ignores filter and special_filters query parameters.
• Pitch shift via rubberband. /pitch-available returns 500 so SStts keeps pitch_enabled = FALSE and the URL templating substitutes 0 in.
• Real blips synthesis. /tts-blips currently returns the same audio as /tts.
• Result caching (sha256 LRU of recent utterances).

Build

docker build -t reduxstation-tts:latest .

The build downloads both voice models from Hugging Face at image build time (about 125 MB). First build takes a couple of minutes; subsequent rebuilds reuse the voice-data layer.

Run

docker run --rm \
    -p 127.0.0.1:5500:5500 \
    -e TTS_AUTHORIZATION_TOKEN=$(openssl rand -hex 32) \
    --name tts \
    reduxstation-tts:latest

Set TTS_AUTHORIZATION_TOKEN to a strong random string in production. The default value coolio is fine for local smoke testing. The game container must send the same value in TTS_HTTP_TOKEN in hippiestation_config.txt.

Smoke test

TOKEN=coolio
curl -fsS -H "Authorization: $TOKEN" http://localhost:5500/tts-voices | head -5
curl -fsS -H "Authorization: $TOKEN" http://localhost:5500/health-check
curl -fsS -H "Authorization: $TOKEN" \
    -H "Content-Type: application/json" \
    --data '{"text":"hello there general kenobi"}' \
    "http://localhost:5500/tts?voice=VCTK%20Man%2002&identifier=test&filter=&pitch=0&special_filters=" \
    -o /tmp/test.ogg
ffprobe /tmp/test.ogg

The first request after start triggers ONNX warmup. Expect 1 to 4 seconds for the cold path. Subsequent requests on the same voice land in tens to low hundreds of milliseconds depending on host CPU.

Wire to the game container

In the deployment compose file (ReduxStation/config/docker-compose.yml), add a service alongside the existing game container:

  tts:
    image: ghcr.io/reduxstation/tts:latest
    restart: unless-stopped
    expose:
      - "5500"
    environment:
      - TTS_AUTHORIZATION_TOKEN=${TTS_TOKEN}
    cpus: "8"
    mem_limit: "4g"
    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://localhost:5500/health-check"]
      interval: 5s
      timeout: 2s
      retries: 3

Then in ReduxStation/config/hippiestation_config.txt:

TTS_HTTP_URL http://tts:5500
TTS_HTTP_TOKEN <same value as TTS_TOKEN env>
TTS_MAX_CONCURRENT_REQUESTS 4
TTS_HTTP_TIMEOUT_SECONDS 30

When TTS_HTTP_URL is unset, SStts returns SS_INIT_NO_NEED at boot and the subsystem disables itself silently. Safe to leave the config keys absent while iterating on the gateway image.

Capacity expectations

On the live VPS (AMD EPYC 9645 Zen 5, 16 vCPUs):

• Cold-start per worker: 1 to 4 seconds (ONNX model load).
• Per-request synthesis: 25 to 80 ms for typical SS13 utterances (5 to 50 chars).
• Memory: about 250 MB resident per worker, about 500 MB total for the two voice models plus Python overhead. Headroom is large; the mem_limit: 4g above is loose.
• Throughput: comfortably 30 to 60 requests per second sustained per single aiohttp worker.