Sinsajo — Real-Time Local Voice Transcription

Why Sinsajo?

Real-time

~150–300ms end-to-end latency. See text as you speak, with Voice Activity Detection (Silero VAD v5) sending only speech segments.

100% Private

Self-hosted Rust server. No API keys, no cloud. Your audio never leaves your network. Optional local WAV/OGG recording.

Multi-platform

Flutter client for Android, iOS, Desktop and Web. Rust server runs anywhere — Docker, bare metal, VPS.

Multi language

Multi-language target selection (en/es/fr/de/pt). English via Parakeet, translation via Canary.

How it works

┌─────────────────┐         WebSocket          ┌──────────────────┐
│  Flutter Client │ ◄─────────────────────────►│   Rust Server    │
│  (Dart/Silero)  │    PCM 16-bit / JSON       │ (transcribe-rs)  │
└─────────────────┘                            └──────────────────┘
       │                                                 │
       │ Silero VAD v5                                   │ ONNX Model
       │ (ONNX Runtime)                                  │
       ▼                                                 ▼
 ┌─────────────────┐                             ┌──────────────────┐
 │  AudioService   │                             │  Canary 180M     │
 │  + vad package  │                             │  Flash /         │
 │  (ML-based)     │                             │  Parakeet TDT    │
 └─────────────────┘                             └──────────────────┘

1. Capture — Flutter records PCM 16-bit @ 16kHz
2. Detect — Silero VAD v5 detects speech (ONNX Runtime)
3. Transmit — Only speech chunks sent over WebSocket
4. Transcribe — Rust server runs Canary 180M Flash or Parakeet TDT 0.6B (Int8)
5. Display — Real-time text + optional server-side WAV/OGG recording

Sinsajo needs two parts to work

Client + Server required

Sinsajo is not standalone. You always need a Server (Rust — runs the AI model) and a Client (Flutter — captures audio and shows text). They talk over WebSocket ws://YOUR_SERVER_IP:8765. Run one server, connect any clients to it.

1. Start a Server — pick one

Docker Compose (recommended)

Best for a Linux server or homelab. Models and recordings are persisted in volumes.

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
# -> ws://localhost:8765 (auto-downloads ~2-3GB model on first run)

Docker CLI

Same image, single command — no compose file needed.

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

Precompiled binary

No Docker needed. Download the binary for your platform and run it.

# GitHub Releases -> server/v* assets
# - sinsajo-server          (linux x86_64)
# - sinsajo-server.exe      (windows x86_64)
./sinsajo-server --model ParakeetTDT --autodownload-model --port 8765
# or on Windows: .\sinsajo-server.exe --model ParakeetTDT --autodownload-model

Grab it from server-latest — no compile, just run.

Build from source

For any architecture that meets the minimum requirements (Rust 1.70+, ~500MB RAM per Int8 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

See Server Guide for CLI flags and GPU acceleration.

2. Connect a Client — pick one

Web App (no install)

Fastest way to try it — nothing to install.

1. Make sure your server is running and reachable (ws://YOUR_SERVER_IP:8765)
2. Open the Web App → Try Web App
3. In Settings set the server IP, tap the mic and speak

Requires HTTPS for microphone access (GitHub Pages provides it).

Android APK

For Android phones/tablets. Share via link or QR.

• Download app-release.apk from client-latest
• Enable Install unknown apps → install → set server IP in Settings → speak

APKs are built on every client/v* tag via GitHub Actions — not stored in git.

Build Flutter app

For any platform that meets client requirements (Flutter 3.x). Compile for Android, iOS, Linux, macOS, Windows or Web.

git clone https://github.com/lutgaru/Sinsajo.git
cd Sinsajo/sinsajo_client
flutter pub get
# optional: edit lib/providers/transcription_provider.dart -> kWsUrl
flutter run          # run on connected device
flutter build apk    # Android
flutter build web --base-href /Sinsajo/app/

See Client Guide.

Try it in 60 seconds

1. docker compose up -d → server on ws://localhost:8765
2. Open Web App → Settings → set localhost or your LAN IP → tap mic.

Choose your way to use Sinsajo

Server → DockerLinux server with Docker Compose or docker run. Recommended for homelab/VPS. Auto-downloads model.

Server → BinaryPrecompiled x86_64 binaries for Windows & Linux. Download from server-latest and run — no Docker, no Rust needed.

Server → From sourceClone and cargo run --release. For any architecture meeting minimum requirements.

Client → Web AppNo install. Open in browser, set server IP and speak. Works on desktop & mobile browsers.

Client → Android APKSide-load on Android. Download from client-latest or share QR with testers.

Client → From sourceFlutter codebase — compile for Android, iOS, Desktop or Web on any supported platform.

At a glance

Metric	Value
Latency	~150–300ms chunk → text
CPU	~20–40% (1 core)
RAM	~500MB (Int8 quantized)
Accuracy	~95% clean Spanish
Bandwidth	~256KB/s (PCM, VAD reduces ~60%)

APK Distribution

APKs are not stored in git. Every tagged release client/v* builds the APK via GitHub Actions and attaches it to the GitHub Release. Link directly to https://github.com/lutgaru/Sinsajo/releases/tag/client-latest/download/app-release.apk from your landing buttons or share a QR.

Next steps

Configure target language

Pick output language in-app. Canary translates, Parakeet is English-only. Server advertises supported languages via model_info.

Save recordings

Enable save_audio in Settings — server saves WAV/OGG per session to records/ for export.

Tune VAD

Adjust thresholds in audio_service.dart for quiet rooms or noisy environments.