Server (Rust)

Runs alone, used by the client

The server does transcription and exposes ws://0.0.0.0:8765 (configurable). It is useless without a Client — connect via Web App, APK or a Flutter build. See Client Guide and Quick Start.

Options — pick one

Docker Compose (recommended)

Best for a Linux server, homelab or VPS. Image lutgaru/sinsajo-server from Docker Hub, models/records persisted.

docker-compose.yml

services:
  sinsajo-server:
    image: lutgaru/sinsajo-server:latest
    ports: ["8765:8765"]
    volumes:
      - sinsajo_models:/app/models
      - sinsajo_records:/app/records
    restart: unless-stopped
volumes:
  sinsajo_models:
  sinsajo_records:

docker compose up -d

Or pin a tag: lutgaru/sinsajo-server:latest or server/v* tag. Pass args after image: ... lutgaru/sinsajo-server:latest --model Canary180M --autodownload-model.

Docker CLI

Same image, single command without compose file:

docker pull lutgaru/sinsajo-server:latest
docker run -p 8765:8765 \
  -v sinsajo_models:/app/models \
  -v sinsajo_records:/app/records \
  lutgaru/sinsajo-server:latest --model ParakeetTDT --autodownload-model

Precompiled binary

No Docker, no Rust needed — download and run.

# From GitHub Releases -> server/v* assets:
# - sinsajo-server        (linux x86_64)
# - sinsajo-server.exe    (windows x86_64)
./sinsajo-server --model ParakeetTDT --autodownload-model --port 8765
# Windows PowerShell: .\sinsajo-server.exe --model ParakeetTDT --autodownload-model
# Custom dirs: --model-dir ./models --records-dir ./records

Get it at server-latest. No compilation for x86_64 Windows/Linux.

CLI args:

Arg	Default	Description
--model	prompt	ParakeetTDT or Canary180M
--autodownload-model	false	Auto-download if missing
--port	8765	WebSocket port
--model-dir	./models	Model storage
--records-dir	./records	WAV/OGG output

Build from source

For any architecture meeting minimum requirements (Rust 1.70+, ~500MB RAM for Int8, ~2–3GB disk for model).

git clone https://github.com/lutgaru/Sinsajo.git
cd Sinsajo/server
cargo run --release -- --model ParakeetTDT --autodownload-model
# or
cargo build --release && ./target/release/sinsajo-server --help

Add new models in src/config.rs (MODELS array) with HuggingFace repo + folder. If --model is omitted it reads sinsajo-config.json, then prompts.

Available: ParakeetTDT (0.6B, more accurate, English) and Canary180M (180M, faster, multilingual).

Models

Auto-download handles ~2–3GB on first run. Subsequent starts ~2–3s. Override dirs with --model-dir / --records-dir.

Current models: ParakeetTDT (TDT 0.6B, English) and Canary180M (Canary 180M Flash, multilingual). Edit server/src/config.rs to add more.

GPU acceleration

use transcribe_rs::{set_ort_accelerator, OrtAccelerator};
set_ort_accelerator(OrtAccelerator::Cuda); // or Auto

Translation

Canary supports target_language from client start message. Server forwards to transcribe_with:

model.transcribe_with(&samples, &CanaryParams {
  language: Some("es".to_string()),
  target_language: Some("en".to_string()),
  ..Default::default()
})?;

Parakeet is English-only and ignores the target language. See Protocol.

Persistence

• Models: models/ (or --model-dir)
• Recordings: records/ (WAV/OGG, triggered by client clean)
• Config: sinsajo-config.json (persists last chosen model)

Deployment

The server is not deployed to GitHub Pages (static). Run it on your LAN, VPS, fly.io / Railway, etc. with Docker or a binary. Expose 8765 and point clients to ws://YOUR_IP:8765 or wss:// behind a reverse proxy (Caddy/Nginx). For remote access consider Tailscale/WireGuard.

x86_64 Windows/Linux

Precompiled binary — fastest without Docker.

Other arch

Cargo build — any arch/OS meeting Rust minimums.

Linux server

Docker Compose / CLI — recommended for homelab.