Skip to content

kylefoxaustin/media-transcriber

BranchesTags

Repository files navigation

Media Transcriber 🎬

Local AI-powered transcription for YouTube videos and local media files. Runs entirely on your hardware with GPU acceleration.

v1.1.0 NEW: Speaker Identification — Create a voice profile and extract only YOUR voice from videos!

Features

  • 🚀 GPU Accelerated - CUDA support via faster-whisper (CTranslate2)
  • 🎯 High Accuracy - Whisper large-v3 model with word-level timestamps
  • 📺 YouTube Support - Videos, playlists, and channels
  • 📁 Local Files - MP4, MKV, MOV, MP3, WAV, and many more formats
  • 🎤 Speaker Identification - Extract only YOUR voice from multi-speaker videos (NEW in v1.1.0)
  • 🗣️ Speaker Diarization - Identify who said what (optional)
  • 📊 Multiple Formats - SRT, VTT, TXT, JSON output
  • 🔄 Batch Processing - Process folders, playlists, or URL lists
  • 🌐 Web Interface - Drag-and-drop uploads with real-time progress
  • 💾 Resume Support - Skip already-processed files
  • 🏠 Fully Local - No cloud APIs, everything on your machine

What's New in v1.1.0 🎤

Speaker Identification lets you:

  • Create a "voice profile" from a sample of you speaking
  • Automatically find and extract YOUR segments from any video
  • Perfect for pulling your voice out of family videos, meetings, podcasts
# One-time: Create your voice profile
python voice_profile.py create kyle ~/Videos/me_talking.mp4

# Extract your voice from any video
python extract_speaker.py kyle family_reunion.mp4

# Output: Just your text, SRT subtitles, or audio-only file

Quick Start

Option A: Docker Web UI

git clone https://github.com/kylefoxaustin/media-transcriber.git
cd media-transcriber
docker compose up -d
# Open http://localhost:8000

Option B: Command Line

git clone https://github.com/kylefoxaustin/media-transcriber.git
cd media-transcriber
bash setup.sh
source venv/bin/activate

# Transcribe
python transcribe.py https://www.youtube.com/watch?v=VIDEO_ID
python transcribe.py /path/to/video.mp4

Speaker Identification (v1.1.0)

Step 1: Create Your Voice Profile

You need a sample where only you are speaking (30-60 seconds is ideal):

# From a video file
python voice_profile.py create kyle ~/Videos/me_presenting.mp4

# From a specific time range (you at 0:30-1:30)
python voice_profile.py create kyle video.mp4 --start 30 --end 90

# From an audio file
python voice_profile.py create kyle my_voice_memo.m4a

Step 2: Extract Your Voice from Any Video

# Extract your segments (outputs text, SRT, audio, JSON)
python extract_speaker.py kyle family_video.mp4

# Just get the text
python extract_speaker.py kyle video.mp4 --output-text

# Just get subtitles of your parts
python extract_speaker.py kyle video.mp4 --output-srt

# Just get audio of you speaking
python extract_speaker.py kyle video.mp4 --output-audio

# Lower threshold = more matches (may include false positives)
python extract_speaker.py kyle video.mp4 --threshold 0.5

Voice Profile Management

# List all profiles
python voice_profile.py list

# Test if a file matches your voice
python voice_profile.py test kyle unknown_video.mp4

# Delete a profile
python voice_profile.py delete kyle

# View profile details
python voice_profile.py info kyle

How It Works

  1. Voice Embedding: Uses SpeechBrain's ECAPA-TDNN model to create a 192-dimensional "voiceprint" from your sample
  2. Transcription: Whisper transcribes the target video into segments
  3. Matching: Each segment's audio is compared to your voiceprint using cosine similarity
  4. Extraction: Segments above the threshold (default 0.6) are extracted

Tips for Best Results

  • Good samples: Clear audio, just your voice, minimal background noise
  • 30-60 seconds: Enough to capture your voice characteristics
  • Multiple samples: You can add more samples to improve accuracy: 
    python voice_profile.py create kyle sample1.wav sample2.wav sample3.wav
  • Threshold tuning: Start at 0.6, lower to 0.5 if missing segments, raise to 0.7 if getting false positives

Basic Transcription

CLI Usage

# YouTube video
python transcribe.py https://www.youtube.com/watch?v=VIDEO_ID

# Local video file
python transcribe.py /path/to/video.mp4

# Local audio file  
python transcribe.py /path/to/podcast.mp3

# Multiple files
python transcribe.py video1.mp4 video2.mov audio.mp3

# From a text file (URLs or paths, one per line)
python transcribe.py --file inputs.txt

# YouTube playlist
python transcribe.py --playlist "https://www.youtube.com/playlist?list=PLxxxxx"

# Folder of media files
python transcribe.py --folder /path/to/videos/

# Folder recursive (include subfolders)
python transcribe.py --folder /path/to/videos/ --recursive

Model Selection

Model VRAM Speed Quality
tiny ~1GB ⚡⚡⚡⚡⚡ ★★☆☆☆
base ~1GB ⚡⚡⚡⚡ ★★★☆☆
small ~2GB ⚡⚡⚡ ★★★☆☆
medium ~5GB ⚡⚡ ★★★★☆
large-v3 ~10GB ⚡ ★★★★★ 
# Use a smaller/faster model
python transcribe.py --model medium video.mp4

# Use CPU (slower but no GPU needed)
python transcribe.py --device cpu --model small video.mp4

Docker Web UI

Starting the Service

docker compose up -d
# Open http://localhost:8000

Features

  • YouTube Tab: Paste URLs, watch real-time progress
  • Upload Tab: Drag-and-drop local files
  • Download results in SRT, VTT, TXT, JSON formats

Management

./manage.sh start    # Start container
./manage.sh stop     # Stop container
./manage.sh logs     # View live logs
./manage.sh gpu      # Verify GPU access

Output Structure

output/
├── transcripts/
│   └── VIDEO_ID/
│       ├── metadata.json
│       ├── Video_Title.srt
│       ├── Video_Title.vtt
│       ├── Video_Title.txt
│       └── Video_Title.json
│
└── speaker_extract/           # Speaker ID outputs
    ├── video_kyle.txt         # Just your text
    ├── video_kyle.srt         # Your subtitles
    ├── video_kyle.wav         # Audio of just you
    └── video_kyle.json        # Full data with confidence scores

Supported Formats

Video: MP4, MKV, AVI, MOV, WMV, FLV, WebM, M4V, MPEG, MPG, 3GP

Audio: MP3, WAV, M4A, AAC, OGG, FLAC, WMA, Opus

Installation

Quick Setup

git clone https://github.com/kylefoxaustin/media-transcriber.git
cd media-transcriber
bash setup.sh
source venv/bin/activate

Manual Setup

# System dependencies
sudo apt update
sudo apt install ffmpeg python3-pip python3-venv

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install packages
pip install -r requirements.txt

cuDNN Setup (if needed)

If you get Unable to load libcudnn_ops.so:

sudo apt install libcudnn9-cuda-12
sudo ldconfig

Requirements

  • Python 3.10+
  • NVIDIA GPU with CUDA support (or CPU fallback)
  • ~10GB disk space for Whisper large-v3 model
  • ~1GB for speaker recognition model
  • ffmpeg
  • Docker (optional, for web UI)

Performance

RTX 5090 + large-v3:

  • Transcription: ~15x realtime (1 hour video ≈ 4 minutes)
  • Speaker ID: ~30 segments/minute

RTX 3080 + large-v3:

  • Transcription: ~5x realtime (1 hour video ≈ 12 minutes)

Troubleshooting

"CUDA out of memory"

python transcribe.py --model medium video.mp4

"SpeechBrain not installed"

pip install speechbrain torchaudio

Low speaker match accuracy

  • Use a cleaner voice sample (just you, no background noise)
  • Try a longer sample (60+ seconds)
  • Lower the threshold: --threshold 0.5

No matches found

  • Threshold might be too high: try --threshold 0.4
  • Voice sample quality might be poor
  • Background noise in target video

Author

Kyle Fox - GitHub

License

MIT License - Use freely for personal and commercial projects.

Credits

Version History

  • v1.1.0 - Speaker identification: voice profiles and speaker extraction
  • v1.0.0 - Initial release: YouTube + local file transcription

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases 2

v1.1.0 - Speaker Identification Latest
Dec 18, 2025
+ 1 release

Packages

No packages published

Languages