🎙 transcribe.py

A clean, terminal-based video and audio transcription tool powered by OpenAI Whisper — fully open-source, runs locally, no API keys required.

---

Features

• 🎬 Transcribe video or audio files directly from the terminal
• ⚡ Choose from 5 Whisper model sizes — from blazing-fast to highly accurate
• 📄 Export to plain text, SRT, WebVTT, or JSON (with timestamps)
• 🌍 Automatic language detection — no configuration needed
• 🔒 Runs entirely offline — your files never leave your machine
• 💅 Clean, interactive UI powered by Rich

---

Requirements

• Python 3.8+
• ffmpeg (system-level)

---

Installation

1. Clone the repo

git clone https://github.com/WiseArts/transcribe.git
cd transcribe

2. Install ffmpeg

# macOS
brew install ffmpeg

# Ubuntu / Debian
sudo apt install ffmpeg

# Windows (via Chocolatey)
choco install ffmpeg

3. Install Python dependencies

pip install openai-whisper rich

Note: The first time you run the tool, Whisper will automatically download the selected model weights and cache them locally. This is a one-time download per model.

---

Usage

python transcribe.py

The tool walks you through three steps:

1. File — enter the path to your video or audio file (drag & drop into the terminal works on most systems)
2. Model — pick a size based on how fast vs. accurate you need it
3. Output format — choose how you want the transcript saved

The output file is saved alongside your source file (e.g. interview.mp4 → interview.srt).

---

Model Options

#	Model	Speed	Quality	VRAM	Best for
1	tiny	██████████	███░░░░░░░	~1 GB	Quick drafts, fast machines
2	base	████████░░	█████░░░░░	~1 GB	Everyday use (default)
3	small	██████░░░░	███████░░░	~2 GB	Better accuracy, still fast
4	medium	████░░░░░░	█████████░	~5 GB	High quality, multilingual
5	large	██░░░░░░░░	██████████	~10 GB	Best possible accuracy

---

Supported File Formats

Video: .mp4 .mov .avi .mkv .webm .flv

Audio: .mp3 .wav .m4a .aac .ogg .flac

---

Output Formats

Format	Extension	Description
Plain text	.txt	Clean transcript, one line per segment
SRT	.srt	Subtitles with timestamps (video players, Premiere, etc.)
WebVTT	.vtt	Web subtitles for HTML5 <video> tags
JSON	.json	Full Whisper output with segment-level confidence data

---

Performance Notes

• CPU vs GPU: The script uses CPU by default (fp16=False) so it works on any machine. If you have an NVIDIA GPU with CUDA, remove the fp16=False flag in the transcribe() call for a significant speedup.
• Speed: As a rough guide on CPU, base transcribes roughly 4–8× real-time speed. A 10-minute video takes around 2–3 minutes.
• Accuracy: Whisper performs best on clear speech with minimal background noise. The medium and large models handle accents and technical vocabulary noticeably better.

---

Dependencies

Package	Purpose
openai-whisper	Speech-to-text transcription
rich	Terminal UI
ffmpeg	Audio extraction from video files

---

License

MIT — do whatever you like with it.