🎙️ TranscribeAI

100% Local AI Transcription with Speaker Diarization
No API key. No cloud. No cost. Runs completely offline on your machine.

---

✨ Features

	Feature	Description
⚡	Dual Engine	faster-whisper (CPU) + mlx-whisper (Apple Silicon GPU, 2-5x faster)
🗣️	Speaker Diarization	Auto-identifies Speaker 1, 2, 3... using MFCC + Agglomerative Clustering
🌍	99+ Languages	Indonesian, English, and 99+ languages with auto-detection
📁	Multi-Format	Input: MP3, MP4, WAV, M4A, OGG, FLAC, WEBM → Output: SRT, TXT, DOCX
🧠	5 AI Models	tiny (39M) → large-v3 (1.5B) — choose speed vs accuracy
📊	Smart Progress	5-stage: Upload → Model → Transcription → Speaker ID → Export
💾	Auto Cache	Downloads model once, loads instantly from cache afterward
🌙	Dark Theme UI	Professional web UI with audio player, search & drag-drop
🔒	100% Offline	Zero data leaves your machine. Your audio stays yours.

---

🚀 Quick Start

🍎 macOS — Apple Silicon (M1/M2/M3/M4)

git clone https://github.com/romizone/transcribeAI.git
cd transcribeAI

# ⚙️ Setup
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install mlx-whisper          # 🔥 GPU acceleration

# ▶️ Run
python3 app.py

🌐 Open http://localhost:8080 in your browser

🍎 macOS (Intel) / 🐧 Linux

git clone https://github.com/romizone/transcribeAI.git
cd transcribeAI

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

python3 app.py

🪟 Windows

git clone https://github.com/romizone/transcribeAI.git
cd transcribeAI

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

python app.py

💡 Or use setup scripts: ./setup.sh (macOS/Linux) or setup.bat (Windows)

---

📥 Pre-download Models

Download models ahead of time so transcription starts instantly:

source venv/bin/activate

python3 download_models.py small    # 📦 Download recommended model
python3 download_models.py all      # 📦 Download all models
python3 download_models.py          # 📋 Check download status

🍎 On Apple Silicon, MLX models are auto-downloaded for GPU acceleration

---

🧠 Models

Model	Params	Size	Speed	Best For
tiny	39M	~75 MB	⚡⚡⚡⚡⚡	Quick drafts, short clips
base	74M	~145 MB	⚡⚡⚡⚡	Casual transcription
small ⭐	244M	~465 MB	⚡⚡⚡	Recommended — best balance
medium	769M	~1.5 GB	⚡⚡	Higher accuracy needed
large-v3	1550M	~2.9 GB	⚡	Maximum accuracy

---

🔧 Engines

Engine	Device	Speed	Install
🔥 mlx-whisper	Apple Silicon GPU	2-5x faster	pip install mlx-whisper
🖥️ faster-whisper	CPU (all platforms)	Baseline	Included in requirements.txt

🤖 The app auto-detects Apple Silicon and defaults to mlx-whisper when available

---

💻 CLI Usage

Transcribe directly from terminal — no browser needed:

# 🎵 Simple transcription
python3 transcribe_cli.py audio.mp3

# 🇮🇩 Indonesian, medium model, 3 speakers
python3 transcribe_cli.py video.mp4 --language id --model medium --speakers 3

# 📂 Custom output folder + multiple formats
python3 transcribe_cli.py audio.wav --output ./results --format srt txt docx

---

📂 Project Structure

transcribeAI/
├── 🐍 app.py               # Flask backend (dual engine, diarization, API)
├── 📄 templates/
│   └── index.html           # Web UI (dark theme, progress, audio player)
├── 🖥️ transcribe_cli.py     # CLI version
├── 📥 download_models.py    # Pre-download models offline
├── 📋 requirements.txt      # Python dependencies
├── ⚙️ setup.sh / setup.bat  # Setup scripts
├── ▶️ run.sh / run.bat      # Run scripts
└── 🔧 .env.example          # Configuration template

---

⚙️ How It Works

🎤 Audio Input
    │
    ▼
🧠 Whisper Transcription
    │  faster-whisper (CTranslate2 INT8)
    │  mlx-whisper (Apple MLX GPU)
    │
    ▼
🔇 VAD Filter
    │  Silero VAD removes silence
    │
    ▼
🗣️ Speaker Diarization
    │  MFCC (20 coeff) + Delta + Spectral + Pitch
    │  → StandardScaler → Agglomerative Clustering
    │
    ▼
📄 Export
    ├── 🎬 SRT (subtitles)
    ├── 📝 TXT (readable transcript)
    └── 📑 DOCX (formatted document)

---

🛠️ Troubleshooting

Problem	Solution
🚫 Port 5000 in use (macOS)	AirPlay uses port 5000. TranscribeAI uses port 8080 by default
❌ ModuleNotFoundError	Activate venv first: source venv/bin/activate
⚠️ python3 aliased wrong	Use venv directly: ./venv/bin/python3 app.py
⏳ Stuck at "Memuat model..."	First run downloads ~465MB model (one-time). Pre-download: python3 download_models.py small

---

🏗️ Tech Stack

Layer	Technology
🖥️ Backend	Flask, faster-whisper, mlx-whisper
🎵 Audio	librosa, numpy, pydub
🗣️ Speaker ID	scikit-learn (Agglomerative Clustering)
📄 Export	python-docx
🎨 Frontend	Vanilla HTML/CSS/JS (zero framework dependencies)

---

📜 License

MIT License — free for personal and commercial use.

---

🇮🇩 Made in Jakarta, Indonesia
Built with ❤️ by Romi Nur Ismanto · @romizone

⭐ Star this repo if you find it useful!