GitHub - itsmevictor/clean-transcribe: A simple CLI to transcribe Youtube videos or local audio/video files and produce LLM-cleaned transcripts for analysis, reading, or subtitles.

3 min read Original article ↗

Clean Transcriber

A command-line tool to turn any YouTube video, local audio or video file into a clean, readable text transcript. It uses the transcription model of your choice (local or API-based) for transcription and your preferred LLM to automatically clean and reformat the output.

Features

  1. Multiple input formats: Supports various audio and video formats for flexible usage (e.g., YouTube URL, .mp3, .wav, .m4a, .opus, .mp4, .mkv, .mov).
  2. Multiple output format: Generate clean transcripts in TXT, SRT, or VTT formats.
  3. Flexible transcription models: Choose from various local (Whisper, Voxtral) and API-based (OpenAI, Gemini, Mistral) models for different use cases.
  4. LLM-powered cleaning that removes filler words, fixes grammar, and organizes content into readable paragraphs.
  5. Wide LLM support - use Gemini, ChatGPT, Claude or any other (local) LLM for cleaning.

Quick Start

# Transcribe a YouTube video
clean-transcribe "https://www.youtube.com/watch?v=VIDEO_ID"

# Transcribe a local video file
clean-transcribe "/path/to/your/video.mp4"

# Transcribe a specific segment of a video
clean-transcribe "https://www.youtube.com/watch?v=VIDEO_ID" --start "1:30" --end "2:30"

# Create clean subtitles from a video
clean-transcribe "https://www.youtube.com/watch?v=VIDEO_ID" -f srt -o subtitles.srt

Installation

Option 1: use pip:

pip install clean-transcribe

Option 2: Install as package

git clone https://github.com/itsmevictor/clean-transcribe && cd clean-transcribe 
pip install -e .
clean-transcribe "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

Configuration

Key Options

  • --format, -f: Output format (txt, srt, vtt)
  • --model, -m: Transcription model (whisper-tiny, whisper-base, whisper-small, whisper-medium, whisper-large, whisper-turbo, whisper-1-api, gpt-4o-transcribe-api, gpt-4o-mini-transcribe-api, gemini-2.5-pro-api, gemini-2.5-flash-api, gemini-2.5-flash-lite-api, gemini-2.0-flash-api, voxtral-mini-api, voxtral-small-api, voxtral-mini-local, voxtral-small-local)
  • --start: Start time for transcription (e.g., "1:30")
  • --end: End time for transcription (e.g., "2:30")
  • --transcription-prompt: Custom prompt for Whisper to guide transcription
  • --llm-model: LLM for cleaning (gemini-2.0-flash-exp, gpt-4o-mini, etc.)
  • --cleaning-style: presentation, conversation, or lecture
  • --save-raw: Keep both raw and cleaned versions
  • --no-clean: Skip AI cleaning

LLM-Powered Cleaning Setup

Quick Setup (Recommended)

# Install and configure Gemini (fast + cost-effective)
llm install llm-gemini
llm keys set gemini
# Enter your Gemini API key when prompted

# Or use any other LLM provider

# OpenAI
llm keys set openai

# Anthropic Claude  
llm install llm-claude-3
llm keys set claude

Uses Simon Willison's excellent llm package for provider flexibility.

Cleaning Process

What it does:

  • Removes filler words (um, uh, so, like, you know, etc.)
  • Fixes grammar and punctuation errors
  • Organizes content into logical paragraphs
  • Maintains original meaning and context

Cleaning styles:

  • presentation: Professional tone, organized paragraphs
  • conversation: Natural flow, minimal cleanup
  • lecture: Educational format, clear sections for notes

Note: SRT/VTT preserve timing while cleaning text content.

Feedback

I'd love to hear your feedback! If you encounter any issues, have suggestions for new features, or just want to share your experience, please don't hesitate to open an issue.