WhatHuh

Automatically Create Subtitles For Video Files

So if you have older movies or tv shows that are impossible to get subtitles for or you want to easily sub movies you create this is about as easy as it can be.

WhatHuh comes from the frustration of there not being an easy way to get or create subs so that I can still enjoy watching things despite my hearing loss.

Features

• Cross-platform - AvaloniaUI GUI works on Windows, macOS, and Linux
• CLI Support - Spectre.Console-powered command line interface for scripting
• Voice Activity Detection (VAD) - Silero VAD filters out silence to reduce hallucinations
• Multiple Whisper Backends - Automatically uses CUDA, Vulkan, CoreML, OpenVINO, or CPU
• LLM Refinement - Optional Ollama integration to clean up transcripts
• Advanced Audio Preprocessing - FFmpeg filters for noise reduction and normalization
• Native AOT - Self-contained executables with no runtime dependency

Requirements

• FFmpeg - Must be installed and available in PATH
• Ollama (optional) - For LLM transcript refinement

CLI Usage

Commands

Transcribe Options

Setup Options

Examples

# Transcribe a single file
whathuh -t video.mp4

# Use large-v3 model for best accuracy
whathuh -t video.mp4 -m large-v3

# English only, disable VAD
whathuh -t video.mp4 -l en -n

# Transcribe entire folder with medium model
whathuh -t "C:\Videos" -m medium

# Transcribe MKV files with LLM refinement
whathuh -t *.mkv -r

# Transcribe entire folder, use large-v3 model, English language, and LLM refinement with llama3.2
whathuh -t F:\Temp\ -m large-v3 -l en -r --llm-model llama3.2:latest

# Setup all dependencies
whathuh -s -a

# Download large-v3 model
whathuh -s -w large-v3

# Show system info
whathuh -i

Whisper Models

This project uses OpenAI's Whisper model to create the subtitles. There are several model options:

Use Large v3 If You Can - Models are downloaded automatically on first use.

Runtime Selection

Whisper.net automatically selects the best available runtime:

1. CUDA - NVidia GPUs with CUDA drivers Requires CUDA 13.1 or later and updated NVIDIA drivers!
2. Vulkan - Vulkan support
3. CoreML - Apple Silicon Macs
4. OpenVINO - Intel hardware acceleration
5. CPU - Universal fallback

Voice Activity Detection

The Silero VAD integration filters out silence and non-speech audio before transcription. This:

• Reduces Whisper hallucinations during silence
• Speeds up processing by skipping non-speech segments
• Improves overall transcription quality

LLM Refinement (Optional)

For possibly better results, enable LLM refinement with Ollama. I really need to stress that this isn't always perfect and results vary widely based on model and on the text it is working with.

With that said:

1. Install Ollama
2. Pull a model: ollama pull phi3:mini
3. Enable refinement in GUI or use --refine in CLI

The LLM corrects:

• Spelling and grammar errors
• Disfluencies (um, uh, you know)
• Acronym capitalization
• Homophones

Building from Source

# Clone the repository
git clone https://github.com/Echostorm44/WhatHuh.git
cd WhatHuh

# Build all projects
dotnet build

# Run GUI
dotnet run --project src/WhatHuh.Gui

# Run CLI
dotnet run --project src/WhatHuh.Cli -- --help

Credits

• Whisper.net by Sandro Hanea
• Silero VAD by Silero Team
• Spectre.Console
• AvaloniaUI
• Image icons by rsetiawan - Flaticon

License

MIT License - See LICENSE for details