yt-whisper 0.1.0

Extract text from YouTube videos using Whisper

yt-whisper

Transcribe any YouTube video to text in one command.

yt-whisper downloads the audio from a YouTube URL and runs it through faster-whisper (CTranslate2 backend — comparable speed to whisper.cpp) or Moonshine for English-only content. Output can be plain text, SRT/VTT subtitles, or JSON with timestamps.

---

Features

• One command — paste a URL, get a transcript
• Multiple models — tiny to large-v3-turbo (Whisper) or Moonshine for ultra-fast English transcription
• Multiple formats — .txt, .srt, .vtt, .json, or all at once
• Auto language detection — or force a specific language
• Translation — translate any language to English via Whisper
• GPU support — CUDA auto-detected, or force cpu/cuda/auto
• Rich terminal UI — progress spinners, video metadata panel, transcription stats

---

Requirements

• Python 3.9+
• ffmpeg — required for audio extraction

# macOS
brew install ffmpeg

# Ubuntu / Debian
sudo apt install ffmpeg

# Windows
winget install ffmpeg

---

Installation

pip install yt-whisper

With Moonshine support (English-only, faster on short clips):

pip install 'yt-whisper[moonshine]'

Install from source:

git clone https://github.com/V0latix/YT-Whisper
cd YT-Whisper
pip install -e .

---

Quick Start

yt-whisper "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

The transcript is printed to the terminal and saved as <video-title>.txt in the current directory.

---

Usage

yt-whisper [OPTIONS] URL

Options

Flag	Short	Default	Description
--model	-m	base	Model name (see Models)
--language	-l	auto	Force language code, e.g. fr, en
--task	-t	transcribe	transcribe or translate (→ English)
--output-dir	-o	.	Directory to save output files
--format	-f	txt	txt, srt, vtt, json, or all
--device	-d	auto	cpu, cuda, or auto
--print/--no-print	-p	print	Print transcript to stdout
--version	-v	—	Show version and exit

Examples

# Basic transcription → saves <title>.txt
yt-whisper "https://youtube.com/watch?v=..."

# Higher accuracy with a larger model
yt-whisper "https://youtube.com/watch?v=..." --model large-v3

# Generate SRT subtitles
yt-whisper "https://youtube.com/watch?v=..." --format srt

# Generate every format at once
yt-whisper "https://youtube.com/watch?v=..." --format all

# Force French, then translate to English
yt-whisper "https://youtube.com/watch?v=..." --language fr --task translate

# Save files to a specific folder
yt-whisper "https://youtube.com/watch?v=..." --output-dir ./transcripts

# Run on GPU
yt-whisper "https://youtube.com/watch?v=..." --device cuda

# Silent — save file only, no stdout
yt-whisper "https://youtube.com/watch?v=..." --no-print

# Use Moonshine for fast English-only transcription
yt-whisper "https://youtube.com/watch?v=..." --model moonshine-base

---

Models

Whisper — multilingual

Model	Params	Languages	Relative speed
tiny	39 M	multilingual	~32×
tiny.en	39 M	English only	~32×
base	74 M	multilingual	~16×
base.en	74 M	English only	~16×
small	244 M	multilingual	~6×
small.en	244 M	English only	~6×
medium	769 M	multilingual	~2×
medium.en	769 M	English only	~2×
large-v1	1 550 M	multilingual	1×
large-v2	1 550 M	multilingual	1×
large-v3	1 550 M	multilingual	1×
large-v3-turbo	809 M	multilingual	~2×

Moonshine — English-only ([moonshine] extra required)

Moonshine scales compute with audio length (unlike Whisper's fixed 30 s chunks), making it significantly faster for short clips.

Model	Params	WER
moonshine-tiny	26 M	12.7 %
moonshine-base	61 M	10.1 %

Models are downloaded automatically on first use and cached at ~/.cache/huggingface/.

---

Output Formats

Format	Description
txt	Plain text transcript
srt	SubRip subtitles — for video players (VLC, mpv, …)
vtt	WebVTT subtitles — for web players and YouTube
json	Full data with per-segment timestamps
all	All four formats at once

Output files are named after the video title and saved to the current directory (or --output-dir).

---

License

MIT