--- name: soundboardio-create-board description: "Create a soundboard on SoundBoardio: a public GitHub repository soundboardio- with the sounds, icons and soundboardio.json manifest, checked and ready to play. Use when someone wants to make, set up, fix or publish their own soundboard of quotes or sounds." --- # Create a SoundBoardio board A [SoundBoardio](https://soundboardio.com/) board is a plain public GitHub repository: sounds, icons and a `soundboardio.json` manifest. SoundBoardio reads it straight from GitHub, so there is nothing to build or deploy, and every push shows up on the board. ## Ask first - What the board is about, its name and the language of its sounds. - Where the sounds are: finished files, or a longer video or recording plus the moments to cut out. - The text of every sound. Take it from the user or from a transcript, never make one up. - The GitHub account or organisation that will own the repository, and the author name and link shown on the board. - An image for the icon. Every board needs one; a still from the source video works. ## Repository Name it `soundboardio-`, with `` in lowercase letters, digits and dashes, and make it public. Without the `soundboardio-` prefix SoundBoardio can't open it. Start from an empty repository: ``` soundboardio-/ ├── soundboardio.json ├── README.md ├── icons/ │ ├── icon-192.png │ ├── icon-512.png │ ├── icon-maskable-512.png │ └── og.jpg └── sounds/ ├── first-quote.mp3 └── second-quote.mp3 ``` The board opens at `https://soundboardio.com/app///`, without the prefix. The README needs only the board's name and that link. ## Sounds One MP3 per tile. File names in lowercase ASCII with dashes: no spaces, and never `#`, `?` or `%`, which break the sound's URL. Uploads of the same video and compilations often overlap: compare their lengths and transcripts first, and cut from the most complete, best-sounding one. A tile plays the moment it is tapped, so cut the silence at the start, and keep the loudness even across the board: ```sh # cut a clip out of a longer recording ffmpeg -ss 00:01:23.40 -to 00:01:26.10 -i source.mp4 -vn clip.wav # fade the ends, trim the silence, even out the loudness, hold the peaks down, mono 96 kb/s MP3 ffmpeg -i clip.wav -af "afade=t=in:d=0.01,afade=t=out:st=2.67:d=0.03,silenceremove=start_periods=1:start_threshold=-50dB,areverse,silenceremove=start_periods=1:start_threshold=-50dB,areverse,loudnorm=I=-16:TP=-1.5:LRA=11,aresample=44100,alimiter=limit=0.79:level=false:latency=true" -ac 1 -ar 44100 -b:a 96k sounds/first-quote.mp3 ``` The fade-out starts 0.03 s before the clip ends (`st`, here 2.70 − 0.03), so a cut through background noise doesn't click. Keep the limiter: without it, resampling and the MP3 encoder push the peaks above 0 dB and some devices crackle. `loudnorm` can't measure a clip shorter than about 0.4 s: leave it out for those. `silenceremove` trims only real silence. When music or room noise runs under the speech it trims nothing, so place the cut points yourself, in a gap between two words. `silencedetect` lists the gaps, with times counted from `-ss`: ```sh ffmpeg -ss 60 -to 80 -i source.mp4 -af silencedetect=noise=-40dB:d=0.05 -f null - ``` When it finds no gap, the background is louder than `noise`: raise it to -30 or -25 dB. Cut in the middle of the gap nearest the word timestamps (see Text). Drawn-out last words are easy to clip and a start can catch the tail of the previous word: check both ends of every clip. ## Text Every tile needs the words it plays. Take them from the user, or transcribe the source with a speech-to-text model that gives word timestamps (Whisper large-v3 works well); the timestamps also show where to cut. Then transcribe every cut clip again: extra words at either end mean the cut is off. Clips under about 2 s often come out garbled, so trust the long transcript for those. A model told to expect one language translates speech in another one: transcribe foreign lines in their own language. Never fill in a word you can't make out. Write `…` in its place in `lyrics`, make no tile of it, and give the user a list of the unclear spots: the source, the time and what it might be. When the speech mixes languages or dialect, ask once how to spell such words (say "good" or "gut") rather than deciding word by word. ## Icons Every board needs icons: the validator fails a board without them. Start from one image of at least 512 × 512 px. When the user has none, a still from the source video works, or the video's thumbnail: ```sh ffmpeg -ss 00:00:05 -i source.mp4 -frames:v 1 art.png yt-dlp --skip-download --write-thumbnail --convert-thumbnails png -o art "