NASA Word Song

Generate lyric-synced vertical videos for Instagram Reels using NASA Landsat imagery.

What it does

1. Downloads satellite imagery from NASA's Your Name In Landsat — each letter maps to a unique aerial photo.
2. Uses OpenAI Whisper to transcribe any song with word-level timestamps.
3. Stitches the Landsat images together with smooth crossfade transitions using MoviePy, synced exactly to the lyrics.

Setup

Requires Python 3.10+, FFmpeg, and uv/pip.

# Install dependencies
uv pip install -e .
# or: pip install -e .

A convenience wrapper script ./nasa-word-song is provided. It automatically uses the virtual environment and sets up the correct paths. Use it instead of python -m nasa_word_song.

Quick start

1. Fetch Landsat images for your word

./nasa-word-song fetch "dalas" -o ./landsat_images

2. Generate the video

./nasa-word-song generate song.mp3 -i ./landsat_images -o final_reel.mp4

This will:

• Transcribe song.mp3 with Whisper
• Map each sung word to its first letter's Landsat image
• Render a 1080x1920 (9:16) MP4 with crossfade transitions

Using a songs folder with timestamps

If you keep all songs in one folder, pass the folder path as audio and pick a song with --song:

./nasa-word-song generate ./songs --song "track1.mp3" -i ./landsat_images -o reel.mp4

To generate a video for only a specific segment (perfect for 20-30s Reels), use --start and --end:

./nasa-word-song generate ./songs --song "track1.mp3" --start 0:15 --end 0:45 -i ./landsat_images -o reel.mp4

Timestamps accept seconds (15) or MM:SS (1:30).

Commands

Command	Purpose
fetch <word>	Download Landsat images for every letter in <word>
transcribe <audio>	Save word-level timestamps to JSON
generate <audio>	Render the final Reels video

Options

generate

• audio — Path to audio file or a folder containing songs
• --song — Song filename (required when audio is a folder)
• -i, --images — Folder with Landsat images (required)
• -o, --output — Output MP4 path (default: final_reel.mp4)
• -t, --transcription — Use a pre-computed JSON instead of running Whisper
• -c, --crossfade — Transition duration in seconds (default: 0.2)
• -f, --fallback — Image to show when a letter is missing
• --word-mode — Map full words instead of first letters
• -m, --model — Whisper model: tiny, base, small, medium, large
• --start — Start timestamp in seconds or MM:SS (e.g. 15 or 0:15)
• --end — End timestamp in seconds or MM:SS (e.g. 45 or 0:45)

Example workflow

# 1. Download images for a short phrase
./nasa-word-song fetch "dreams" -o ./landsat_images

# 2. Generate a 30-second clip from a songs folder
./nasa-word-song generate ./my_songs --song "dreams.mp3" --start 0:30 --end 1:00 -i ./landsat_images -o reel.mp4 --crossfade 0.15

macOS Apple Silicon tip

For much faster transcription on M1/M2/M3 Macs, install the optional dependency:

uv pip install -e ".[fast]"

Then edit transcribe.py to use mlx_whisper instead of openai-whisper.