☕ Support This Project

Support my work: coff.ee/theseraphim

Audio Transcription Tool

Complete system for audio transcription with speaker diarization, using AI models that work entirely offline. Optimized for Windows with intelligent installation and automatic compatibility handling.

Features

• Accurate transcription using OpenAI Whisper models
• Intelligent speaker diarization with automatic method selection:
  • Python 3.12+: Resemblyzer (no compilation issues, no tokens needed)
  • Python 3.11-: pyannote.audio (traditional method)
• Precise timestamps for every segment and word
• Automatic audio cleaning to improve quality
• Support for multiple formats (M4A, WAV, MP3, FLAC, etc.)
• Multiple output types (JSON, TXT, Markdown)
• Offline operation - no external APIs required
• Optimized GPU acceleration with CUDA 12.4 support
• Local model caching for offline usage
• Detailed logging with precise timestamps
• Zero-config installation for most scenarios

Installation (Windows)

🚀 One-Click Installation (Recommended)

Run PowerShell as Administrator and execute:

# Full installation with all features
.\install.ps1

# With model downloads for offline usage
.\install.ps1 -DownloadModels -DownloadDiarizationModels -HuggingFaceToken "hf_your_token"

# CPU-only installation (no CUDA)
.\install.ps1 -NoGPU

# Skip speaker diarization if not needed
.\install.ps1 -SkipDiarization

The installation script automatically:

• ✅ Detects your Python version and chooses the best compatibility approach
• ✅ Installs optimized PyTorch with CUDA 12.4 for maximum GPU performance
• ✅ Handles all dependency conflicts and compilation issues
• ✅ Sets up speaker diarization without HuggingFace token hassles (Python 3.12+)
• ✅ Downloads models locally for offline usage (optional)
• ✅ Creates a ready-to-use virtual environment

Installation Options

# Installation script options
.\install.ps1 [OPTIONS]

Options:
  -SkipDependencies         # Skip system tools (Python, FFmpeg, etc.)
  -NoGPU                   # Force CPU-only PyTorch
  -DownloadModels          # Download Whisper models
  -DownloadDiarizationModels # Download speaker diarization models locally
  -SkipDiarization         # Skip speaker diarization setup
  -ForceNonAdmin           # Run without Administrator (limited functionality)
  -HuggingFaceToken        # Specify HF token for pyannote models
  -Help                    # Show detailed help

System Requirements

Minimum:

• Windows 10/11
• PowerShell 5.1+
• Administrator privileges (for installation)
• 4GB RAM
• 2GB disk space

Recommended:

• Windows 11
• 16GB RAM
• NVIDIA GPU with 6GB+ VRAM (automatically detected)
• 10GB free disk space

Manual Installation (Advanced Users)

If you prefer manual control:

# 1. Create virtual environment
python -m venv venv
venv\Scripts\Activate.ps1

# 2. Install dependencies
pip install -r requirements.txt

# 3. For GPU acceleration (optional)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124

# 4. For speaker diarization (Python 3.12+)
pip install resemblyzer scikit-learn

# 5. For speaker diarization (Python 3.11-)
pip install pyannote.audio

Models Used

Whisper (Transcription)

Whisper models are downloaded automatically on first use or with -DownloadModels:

Model	Size	VRAM	Speed	Quality
tiny	39 MB	~1GB	~32x	Basic
base	74 MB	~1GB	~16x	Good
small	244 MB	~2GB	~6x	Very good
medium	769 MB	~5GB	~2x	Excellent
large-v3	1550 MB	~10GB	~1x	Best

Speaker Diarization

Resemblyzer (Python 3.12+, default):

• Advantages: No compilation, no tokens needed, works immediately
• Size: ~50MB
• Requirements: None beyond installation script

PyAnnote (Python 3.11-, optional):

• Model: pyannote/speaker-diarization-3.1
• Size: ~300MB
• Requirements: HuggingFace token and model acceptance

Usage

Basic Examples

# Activate virtual environment first
venv\Scripts\Activate.ps1

# Simple transcription
python main.py input.m4a -o output --format txt

# With speaker diarization
python main.py input.m4a -o output --format md --diarize

# High quality with GPU acceleration
python main.py input.m4a -o output --format json --model-size large-v3 --clean-audio --diarize

# Force CPU usage
python main.py input.m4a -o output --format txt --device cpu

Full Parameters

python main.py INPUT_FILE -o OUTPUT_PATH [OPTIONS]

Required arguments:
  INPUT_FILE                Input audio file
  -o, --output OUTPUT_PATH  Output path (without extension)

Options:
  --format {json,txt,md}    Output format (default: txt)
  --model-size {tiny,base,small,medium,large,large-v2,large-v3}
                            Whisper model size (default: base)
  --diarize                 Enable speaker diarization
  --language LANG           Audio language (auto-detect if not specified)
  --device {auto,cpu,cuda}  Processing device (default: auto)
  --download-models         Download models before processing
  --log-level {DEBUG,INFO,WARNING,ERROR}
                            Logging level (default: INFO)
  --clean-audio             Apply audio cleaning

Practical Examples

# Italian meeting with multiple speakers
python main.py meeting.m4a -o meeting_transcript \
  --format md --diarize --language it --clean-audio

# English interview, JSON output for analysis
python main.py interview.wav -o interview_data \
  --format json --model-size medium --diarize

# Long podcast with best quality
python main.py podcast.mp3 -o podcast_transcript \
  --format txt --model-size large-v3 --log-level DEBUG

Output Structure

TXT Format

Audio Transcription
Generated: 2025-06-17 14:30:00
Language: Italian
Speakers detected: 2

==================================================

[00:00:05] SPEAKER_00: Buongiorno e benvenuti alla nostra riunione.
[00:00:12] SPEAKER_01: Grazie, sono felice di essere qui.
[00:00:18] SPEAKER_00: Iniziamo con il primo punto all'ordine del giorno.

JSON Format

{
  "metadata": {
    "created_at": "2025-06-17T14:30:00",
    "language": "Italian",
    "has_speakers": true,
    "total_segments": 45,
    "diarization_method": "Resemblyzer"
  },
  "transcription": {
    "text": "Full transcription text...",
    "language": "Italian",
    "segments": [
      {
        "id": 0,
        "start": 5.2,
        "end": 8.7,
        "text": "Buongiorno e benvenuti",
        "speaker": "SPEAKER_00",
        "words": [
          {
            "word": "Buongiorno",
            "start": 5.2,
            "end": 5.8,
            "probability": 0.95
          }
        ]
      }
    ]
  }
}

Markdown Format

# Audio Transcription

**Generated:** 2025-06-17 14:30:00  
**Language:** Italian  
**Speakers:** 2  
**Diarization:** Resemblyzer

---

## Transcription with Timestamps

### SPEAKER_00

**00:00:05**: Buongiorno e benvenuti alla nostra riunione.

**00:00:18**: Iniziamo con il primo punto all'ordine del giorno.

### SPEAKER_01

**00:00:12**: Grazie, sono felice di essere qui.

Smart Installation Features

Automatic Python Version Handling

The installation script detects your Python version and automatically chooses the best approach:

Python 3.13+: Uses Resemblyzer for speaker diarization (recommended)

• ✅ No compilation issues
• ✅ No HuggingFace tokens needed
• ✅ Works immediately after installation

Python 3.11-3.12: Offers both Resemblyzer and pyannote.audio options

• ✅ Fallback mechanisms for maximum compatibility
• ✅ Handles compilation issues automatically

GPU Optimization

• Automatic CUDA detection and driver compatibility check
• Optimized PyTorch installation with CUDA 12.4 support
• Eliminates Triton warnings and performance issues
• CPU fallback for systems without compatible GPUs

Local Model Caching

# Download models for offline usage
.\install.ps1 -DownloadDiarizationModels -HuggingFaceToken "hf_your_token"

• Downloads pyannote models to local cache
• Enables completely offline operation
• Bypasses HuggingFace authentication during transcription
• Automatic cache setup and management

Troubleshooting

Installation Issues

1. "Administrator privileges required"

ERROR This script requires Administrator privileges!

Solution: Run PowerShell as Administrator or use:

.\install.ps1 -ForceNonAdmin -SkipDependencies

2. Python version compatibility

WARNING Python 3.13 detected - using Resemblyzer for speaker diarization

Solution: This is normal! The script automatically uses the best method for your Python version.

3. CUDA installation issues

WARNING Failed to install CUDA toolkit

Solution: Use CPU-only mode:

.\install.ps1 -NoGPU

Runtime Issues

1. Speaker diarization not working

INFO No speaker diarization available

Check: Verify installation completed successfully:

python -c "import resemblyzer; print('Resemblyzer available')"

2. GPU not detected

INFO Using device: cpu

Solution: Check CUDA installation:

python -c "import torch; print(torch.cuda.is_available())"

3. Model download issues

WARNING Model download failed

Solution: Models download automatically on first use, or use:

.\install.ps1 -DownloadModels

Performance Optimization

For long files (>1 hour):

• Use --model-size small or base
• Avoid --clean-audio if not needed
• Consider using --device cpu for very long files

For best quality:

• Use --model-size large-v3
• Enable --clean-audio
• Ensure GPU acceleration: --device cuda

For maximum speed:

• Use --model-size tiny
• Disable diarization or use Resemblyzer method
• Skip audio cleaning

Logging

The system uses detailed logging with timestamps:

[20250617-143000] INFO - Starting audio transcription process
[20250617-143001] INFO - Using device: cuda
[20250617-143001] INFO - Resemblyzer available - speaker diarization enabled
[20250617-143002] INFO - Processing audio file: meeting.m4a
[20250617-143005] INFO - Loaded audio: 1547.2s, 16000Hz
[20250617-143006] INFO - Loading Whisper model: large-v3
[20250617-143008] INFO - Starting transcription...
[20250617-143045] INFO - Detected language: Italian
[20250617-143046] INFO - Performing speaker diarization...
[20250617-143078] INFO - Diarization completed: 2 speakers detected
[20250617-143079] INFO - Saving output in md format...
[20250617-143080] INFO - ✅ Transcription completed with speaker diarization

Supported Audio Formats

• Input: M4A, WAV, MP3, FLAC, OGG, WMA, AAC
• Internal processing: WAV 16kHz mono
• Maximum duration: Limited only by available memory

Advanced Configuration

Environment Variables

# For offline pyannote usage (if downloaded locally)
$env:HF_HUB_OFFLINE = "1"

# Force CPU usage
$env:CUDA_VISIBLE_DEVICES = ""

# Custom cache directory
$env:TRANSFORMERS_CACHE = "D:\models\cache"

Custom Diarization Settings

The system automatically chooses the best diarization method, but you can verify which is being used:

python -c "
import sys
sys.path.append('src')
from diarizer import Diarizer
print('Diarizer using optimal method for your Python version')
"

License

This project uses:

• Whisper: MIT License (OpenAI)
• Resemblyzer: Apache 2.0 License
• PyAnnote: MIT License
• Other dependencies: Various open source licenses

See the LICENSE files of the individual dependencies for full details.

---

🎯 Ready to transcribe? Run .\install.ps1 as Administrator and you'll be set up in minutes!