🎙️ PhoWhisper Tool - Advanced AI Audio Transcription Assistant (2026)

PhoWhisper Tool is a specialized AI assistant for Vietnamese and multilingual audio, designed to transcribe audio recordings, YouTube videos, lectures, and meetings. The tool automatically converts speech to text (Transcription) and utilizes the most advanced LLMs (Gemini 3.1, Gemma 4) to summarize and refine the content.

✨ Key Features

• 🚀 GPU Acceleration (CUDA 12.4): Leverages the power of RTX 30/40/50 series for up to 20x faster processing.
• 🧠 Gemini 3.1 Advanced Integration: Supports high-end features like Thinking Budget (deep reasoning) and Google Search Grounding (real-time fact-checking).
• 🏠 Local LLM Privacy: Integrated with Ollama and Gemma 4 (31B) for absolute privacy with offline processing.
• 📥 YTDown Core Engine: Specialized YouTube/Google Drive download engine that bypasses bot restrictions with multi-threaded speed.
• 🔊 Smart Audio Processing: Automatic volume normalization, AI-powered Noise Reduction, and Speed Optimization.
• 🎨 Modern GUI: User-friendly graphical interface with real-time progress tracking and integrated Q&A for lecture content.

🛠️ System Requirements

• Python: 3.10 - 3.12 (3.12 Recommended)
• FFmpeg: Required (For audio/video processing)
• NVIDIA GPU: RTX 20 series or newer (Required for CUDA acceleration)
• CUDA Toolkit: 12.4+ (If using GPU)

📥 Quick Installation Guide

1. Install FFmpeg & CUDA

• Windows: Download FFmpeg from gyan.dev and add to PATH. Install CUDA Toolkit 12.4.
• macOS: brew install ffmpeg
• Linux: sudo apt install ffmpeg

2. Install Python Dependencies (GPU Support)

Open Terminal/PowerShell in the project directory and run:

# 1. Uninstall old versions (if any)
pip uninstall -y torch torchvision torchaudio

# 2. Install PyTorch with CUDA 12.4 support (Latest 2026)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

# 3. Install other dependencies
pip install -r requirements.txt

⚙️ API Configuration (.env)

Create a .env file in the root directory and configure it as follows:

# Google Gemini API (Get yours at aistudio.google.com)
GEMINI_API_KEY=your_api_key_here

# Advanced Gemini Model Options (Latest 3.1 Series)
# Suffixes: -maxthinking (deep reasoning), -search (Google search grounding)
GEMINI_MODEL=gemini-3.1-pro-preview-maxthinking

# Optional Proxy (if using geminicli2api)
# GEMINI_BASE_URL=http://localhost:8888/v1

# Local LLM (Ollama)
LOCAL_MODEL_NAME=gemma4:31b

🚀 Usage Guide

🖥️ Graphical User Interface (GUI)

python gui.py

Supports drag-and-drop, real-time status updates, and direct chat with your transcript content.

⌨️ Command Line Interface (CLI)

Process Local Files:

# Automatically process all files in the /audio folder
python phowhisper.py

# Manual mode to select GPU/CPU and specific files
python phowhisper.py --mode manual

Process URLs Directly:

# Download and process from YouTube
python phowhisper.py https://www.youtube.com/watch?v=xxxx

# Download and process from Google Drive
python phowhisper.py https://drive.google.com/file/d/xxxx/view

Advanced Parameters:

• --noise-reduction: Activate AI-powered noise filtering.
• --asr-model auto|phowhisper|whisper: Choose the ASR model.
• --skip-speed: Skip speed optimization (useful for 100% word-for-word accuracy).

📁 Project Structure

• audio/: Input folder for raw files (automatically cleaned up after processing).
• output/: Output folder for raw transcripts (.txt) and refined summaries (_processed.txt).
• core/: YTDown engine for high-performance audio/video downloading.

🤝 Contact & Contribution

• Author: Phan Chi Kien ([email protected])
• GitHub: https://github.com/phanc/phowhisper-tool

---

Developed with passion for the Vietnamese AI community and beyond.