Song Layerer

This project is a Node.js-based tool that allows for experimental audio generation, by time-stretches collections of tracks to the same duration (taking the longest as basis), and layers them together.

Inspired by Every Recording of Gymnopédie 1.

Features

• Download top N search results from YouTube
• Use self-selected audio files as input
• Time-stretch audio files to match the longest duration
• Layer multiple audio files into a single output

Dependencies

• Node.js
• ffmpeg
• yt-dlp
• Rubberband (MacOS ARM version included in ./lib/rubberband/)

yt-dlp Notes

YouTube frequently breaks older yt-dlp builds. If downloads fail with messages like "Signature extraction failed", HTTP 403, or "Only images are available", update yt-dlp (macOS: brew upgrade yt-dlp).

If YouTube is rate-limiting or blocking your IP, you may need cookies. This tool supports:

• YTDLP_COOKIES_FROM_BROWSER=chrome (or firefox, etc.)
• YTDLP_COOKIES=/path/to/cookies.txt

Installation

1. Clone this repository
2. Install dependencies: pnpm i
3. Ensure ffmpeg and yt-dlp are installed and accessible in your PATH

Usage

Run the main script with the following command:

node main.js [options] "search query"

Options:

• --download or -d: Download songs from YT first (default: false)
• --number or -n: Number of songs to download (default: 5)
• --exclude or -e: Comma-separated list of keywords to exclude from search (e.g guitar,drums)

Example:

node main.js -d -n 3 -e "guitar,🎸" "Clair de Lune"

Alternative usage:

Manually populate the output/[search_query]/songs/ directory with audio files, and run node main.js "search query". This skips the download step.

UI (React)

This repo includes a tiny local UI (Vite + React + Tailwind + shadcn/ui).

• Dev: pnpm dev then open http://localhost:5173
• Prod build: pnpm build then pnpm start (serves UI + API on http://localhost:5174)

Output

The processed files will be saved in the output/ directory, organized by search query:

output/

  └── [search_query]/

    ├── `songs/`

    ├── `stretched_songs/`

    └── `layered_[sanitized_search_query].wav`

File Structure

• main.js: Main entry point, orchestrates the process
• songdl.js: Handles downloading cover versions from YouTube
• timestretch.js: Time-stretches audio files to match the longest duration
• layerer.js: Layers the time-stretched audio files into a single output

Troubleshooting

This tool does not promise a consonant result. To improve output quality, you can try the following:

• check the downloaded songs for artifacts at the beginning and end of the file (e.g applause, vocals, etc.). remove these songs.
• check the length of the downloaded songs. remove outliers (e.g much shorter or longer)

Once you have a good set of songs, you can run the tool without the --download option to skip the download step.