Backend Overview

Understanding the Subtide backend server.

Architecture

The Subtide backend is a Flask-based REST API that handles:

Video Processing - Downloading and extracting audio
Transcription - Converting speech to text using Whisper
Translation - Translating text using LLM APIs
Streaming - Real-time translation for Tier 4

graph TB
    subgraph Backend[Subtide Backend]
        API[Flask API] --> Whisper[Whisper Service]
        Whisper --> Translation[Translation Service]

        Whisper --> MLX[MLX / Faster Whisper]
        Translation --> LLM[OpenAI / OpenRouter]

        API --> YouTube[YouTube Service<br/>yt-dlp]
    end

    style API fill:#4a90e2,stroke:#2c5aa0,color:#fff
    style Whisper fill:#7b68ee,stroke:#5a4fcf,color:#fff
    style Translation fill:#50c878,stroke:#3a9d5f,color:#fff
    style MLX fill:#ffa726,stroke:#f57c00,color:#fff
    style LLM fill:#26c6da,stroke:#00acc1,color:#fff
    style YouTube fill:#ff6b6b,stroke:#d63031,color:#fff

Deployment Options

Option	Best For	Setup Difficulty
Binary	Personal use, quick start	Easy
Python Source	Development, customization	Medium
Docker	Production, teams	Medium
RunPod	GPU acceleration, cloud	Medium

Quick Start

Binary (Fastest)

# Download from releases
chmod +x subtide-backend-macos
./subtide-backend-macos

Docker (Production)

cd backend
docker-compose up subtide-tier2

Source (Development)

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
./run.sh

Whisper Backends

Subtide supports multiple Whisper implementations:

MLX (Apple Silicon)

Optimized for M1/M2/M3/M4 Macs:

WHISPER_BACKEND=mlx ./subtide-backend

Uses unified memory efficiently
No separate GPU memory needed
Best for macOS users

Faster-Whisper (NVIDIA)

CUDA-accelerated for NVIDIA GPUs:

WHISPER_BACKEND=faster ./subtide-backend

Requires CUDA toolkit
Significant speedup on supported GPUs
Best for Linux/Windows with NVIDIA

OpenAI Whisper (Cloud)

Original Whisper implementation:

WHISPER_BACKEND=openai ./subtide-backend

CPU-based (slower)
Works everywhere
Fallback option

Model Selection

Model	Size	Speed	Quality
`tiny`	~39 MB	Fastest	Basic
`base`	~74 MB	Fast	Good
`small`	~244 MB	Medium	Better
`medium`	~769 MB	Slow	Great
`large-v3`	~1.5 GB	Slowest	Best
`large-v3-turbo`	~800 MB	Fast	Excellent

Recommended: large-v3-turbo for best speed/quality ratio.

Hardware Requirements

Apple Silicon

Mac	Memory	Recommended Model
8 GB	Limited	`tiny`, `base`
16 GB	Good	`small`, `base`
32 GB	Excellent	`large-v3`
64 GB+	Optimal	Any model

NVIDIA GPUs

GPU	VRAM	Recommended Model
RTX 3060	12 GB	`medium`
RTX 3090/4080	16-24 GB	`large-v3`
RTX 4090	24 GB	Any model

Environment Variables

Core configuration:

Variable	Description	Default
`PORT`	Server port	`5001`
`WHISPER_MODEL`	Model size	`base`
`WHISPER_BACKEND`	Backend type	Auto
`CORS_ORIGINS`	Allowed origins	`*`

See Configuration for full list.

Health Check

Verify the backend is running:

curl http://localhost:5001/health

Expected response:

{"status": "healthy"}

Next Steps

Docker Deployment - Container deployment
RunPod Deployment - Cloud GPU deployment
Local LLM Setup - Use local models
API Reference - API documentation