Sound Library

Your own local sound-effects library, with a search engine that speaks your language — for DaVinci Resolve and any other editor.

No subscriptions. Nothing uploaded to the cloud. Every sound commercially licensed and traceable.

_{Searching caballo — Spanish for horse — across a library whose filenames are entirely in English.}

---

What this is

Sound-effects platforms charge €10–30 a month, and the day you stop paying you lose the right to keep using what you already put in your videos.

This is the alternative. A script downloads thousands of free, permanently commercial-use sounds, merges them into one library with metadata, and gives you a local web app to search, audition and move them into your editor.

A typical install lands ~5,000 sounds and 16 hours of audio in about 24 GB.

Features

• Search that speaks Spanish even though the files are in English. Type caballo and it finds Hoof_Gallop, Pony Whinny and FEETHors_Draft Horse. Handles partial words (caball) and typos (caballlo).
• 12 auto-classified categories: impacts, whooshes, UI, ambiences, foley, voice, cinematic, weapons, vehicles, magic & sci-fi, water, animals.
• Similar sounds via vector similarity — from one gallop it surfaces every other hoof and trot, even with no words in common.
• Player with waveform, seeking, looping and keyboard shortcuts.
• Favourites exported to a folder you drag straight into Resolve's Media Pool.
• Licences always visible and filterable, with an auto-generated CREDITS.md for anything requiring attribution.
• Add your own libraries from the interface or the terminal.

---

Install

You need Python 3.8+ and ffmpeg:

sudo apt install python3 ffmpeg git      # Debian/Ubuntu
brew install python3 ffmpeg git          # macOS

Then:

git clone https://github.com/VictorEscribano/biblioteca-de-sonidos.git
cd biblioteca-de-sonidos
./install.sh

The installer checks dependencies and disk space, downloads the libraries, indexes everything and installs the desktop launcher. Expect 30 min to 2 h depending on your connection — it's ~20 GB.

./install.sh --budget-gb 5      # lightweight install
./install.sh --skip-download    # only index what you already have
./install.sh --no-desktop       # skip the desktop icon

Downloads are resumable: Ctrl-C and relaunch picks up where it left off.

---

Usage

Open the Sound Library icon, or from a terminal:

./sonidos            # opens http://sfx.localhost:7777
./sonidos estado     # summary of what you have
./sonidos index      # reindex after adding sounds
./sonidos icono      # reinstall the desktop launcher

Action	How
Search	Type, or press /
Play	Click a row, or space
Navigate	↑ ↓ through results
Favourite	Star on the row, or F
Find similar	Node icon on the row
Copy path	Copy icon on the row

Getting sounds into DaVinci Resolve

The whole library, always available: in Resolve's Media Storage panel, browse to library/ and add it as a favourite. It'll be in every project without reimporting.

A specific selection: star what you want, hit Export favourites, and drag the resulting folder into the Media Pool. It uses hard links, so it costs no extra disk space.

Adding your own libraries

From the interface, hit "Add library": point it at a folder or .zip already on your disk, name the vendor and licence, and it becomes its own section under Sources. Nested zips are handled — download-centre bundles often wrap everything in an outer archive.

From the terminal:

./sonidos add ~/Downloads/my-library.zip \
    --vendor "Studio name" \
    --license "Commercial royalty-free"

Nothing is copied — it links, so it costs no extra space.

---

Where the sounds come from

Source	Licence	Commercial use	Attribution
Sonniss GDC Bundles	Own royalty-free	Yes, unlimited	No
Kenney.nl	CC0	Yes	No
Freesound (optional)	CC0 / CC-BY	Yes	CC-BY only

Everything install.sh fetches is commercial-use without attribution. Add Freesound and any CC-BY sounds are listed in CREDITS.md with author and link, while the interface shows each sound's licence as it plays.

Important

Sonniss's licence explicitly prohibits using their audio to train AI models.

Freesound (optional)

Adds ~1,700 curated sounds in original quality. Needs OAuth2 credentials from https://freesound.org/apiv2/apply:

python3 tools/dl_freesound.py --setup
python3 tools/dl_freesound.py --run --budget-gb 5
./sonidos index

Respects their API limits (60 requests/minute, 2,000/day).

---

How the search works

This was the hard part. The library is in English and uses abbreviated UCS naming (FEETHors_Draft Horse Walk, VEHWagn_Wood Cart), so searching caballo returned nothing and even horse found only 6 of the 15 equine sounds present.

The trick is to expand the documents at index time, not the queries:

1. tools/thesaurus.py defines 109 concepts holding ~1,200 Spanish and English terms. Indexing Hoof 2_Rocks_Gallop-4-Step appends caballo, horse, galope, relincho, casco, equino…, so whatever you type is already literally in its search text.
2. UCS prefixes get split and translated: FEETHors → feet + hors → footsteps + horse.
3. At query time (web/search.js) three passes of decreasing precision run: exact token → prefix → trigram. Ranking is BM25, boosted when the match lands in the filename.
4. Similar sounds uses cosine similarity in TF-IDF space.

Across 5,165 sounds: 8,281 terms, index built in ~150 ms, queries in ~0.2 ms. All in the browser, no search server.

---

Layout

├── install.sh           installer
├── sonidos              launcher (abrir, index, estado, add, icono)
├── tools/
│   ├── categories.py    taxonomy and keyword classifier
│   ├── thesaurus.py     ES-EN thesaurus and UCS abbreviations
│   ├── build_index.py   dedupe, metadata, classification, index.json
│   ├── serve.py         local server (Range, export, import)
│   ├── crawl_gamesounds.py / dl_gamesounds.py / dl_kenney.py / dl_freesound.py
│   ├── add_pack.py      add your own libraries
│   └── install_desktop.sh
└── web/                 interface: HTML + CSS + JS, no dependencies

Generated content (library/, _staging/, index.json, CREDITS.md) is not in the repo: this is the tool, you download your own audio.

Design notes

• Hard links instead of copies. library/ and exports/ share inodes with _staging/, so the per-category structure doesn't duplicate ~20 GB.
• Dedupe by size + hash of the first 256 KB. Hashing 20 GB in full would be glacial, and for exact duplicates this is enough.
• The server implements Range requests, which http.server lacks. Without them the browser can't seek — it downloads the whole WAV before playing.
• Waveforms are decoded in the browser only for files under 10 MB; above that you get a plain bar.
• The server listens on both loopback addresses. sfx.localhost resolves to ::1 before 127.0.0.1, and serving IPv4 only makes the browser eat a connection refusal before retrying.
• [hidden] { display: none !important } is deliberate. Any rule setting display outbeats the browser's default for the hidden attribute, which left the modal and player bar permanently visible.

---

Contributing

Most valuable contributions, in order:

Thesaurus terms. This is what improves the tool most and is the easiest to contribute. If you search for something and don't find it, open an issue with the query and the sound you expected, or add the term to tools/thesaurus.py:

["horse", "caballo", "equino", "pony", "hoof", "gallop", "relincho", ...],

Each group holds interchangeable terms across both languages. Two traps worth knowing, both found by testing against real data and annotated in the code:

• Ambiguous synonyms drag in noise. snort sat in the horse concept and pulled in pigs; plate sat in the glass concept (via Spanish plato) and pulled in metal plates.
• Vendor names contaminate. A pack from "Digital Rain Lab" injected "water" and "rain" into every one of its files. That's why the pack name is indexed as literal text and never activates concepts.

New sources. If you know free commercial-use libraries that can be fetched by script, open an issue. Be aware many sit behind Cloudflare and return 403 to curl.

Categories and keywords in tools/categories.py, if you spot misclassifications.

Before opening a PR

python3 -m py_compile tools/*.py     # Python syntax
node --check web/app.js              # JS syntax
python3 tools/build_index.py         # must exit 0

The project has no dependencies beyond requests, and the interface uses no framework. Let's keep it that way.

---

Licence

Code under MIT. Audio files are not covered by it — each sound keeps its original provider's licence. See the CREDITS.md generated on every index build.