ffmpeg-minimal/README.md

# Minimal audio-only ffmpeg build (rAIdio.bot)

## Why this exists

The binary we previously shipped was BtbN's kitchen-sink `win64-gpl` build:
~80 external libraries, including a dozen **patent-encumbered video encoders**
(libx264/H.264, libx265+libkvazaar/HEVC, libvvenc/H.266, libxvid/MPEG-4,
libxavs2/libdavs2/libuavs3d/AVS, libopencore-amr/AMR). rAIdio.bot is an audio
app — our code only ever invokes audio operations:

- probe duration: `-i <file> -f null -`
- decode to PCM: `-c:a pcm_s16le` / raw `f32le`
- encode: `libmp3lame`, native `flac`, native `pcm`, `libvorbis`/`libopus`
- always with `-vn` (drop video)

Consumers of `tools/ffmpeg.exe` (both subprocess, both audio-only):
1. the Rust app (`src-tauri/src/backend/ffmpeg.rs`, `commands/*`)
2. pydub / whisper on the ComfyUI side — `launcher.rs:552` prepends `tools/`
   to the Python subprocess PATH "so pydub/ffmpeg is findable".
   (`torchaudio.save` is patched to soundfile, so it needs no ffmpeg.)

So we ship a video-codec kitchen sink — with commercial patent exposure — to run
`wav→mp3/flac` audio conversion. This build replaces it with exactly what we use.

## What this build keeps / drops

- **No `--enable-gpl`, no `--enable-nonfree`** → the binary is **LGPL-2.1-or-later**
  (ffmpeg core) instead of GPL-3.0. No GPL source-offer obligation.
- **External libs: only** `libmp3lame` (LGPL-2.1), `libvorbis`+`libogg` (BSD),
  `libopus` (BSD), `zlib` (Zlib). That is the entire third-party attribution
  surface — down from ~80.
- **Encoders/muxers limited to audio** (pcm/flac/mp3/ogg/opus/wav). No video
  encoders of any kind are present in the artifact → removes the codec
  patent-pool distribution exposure.
- **Full native audio decode + container demux kept** (so any user import still
  loads: m4a/AAC, opus, ogg, flac, wav, mp3, wma, and audio tracks out of
  mp4/mkv/mov/avi/asf). Patent-encumbered **video decoders** (h264/hevc/vvc/
  vc1/mpeg2video/mpeg4/msmpeg4*) are explicitly disabled — we never decode video.

## Build

Reproducible Docker cross-compile (Ubuntu + mingw-w64), mirroring the toolchain
shape of the upstream BtbN build:

```
powershell tools/ffmpeg-minimal/build.ps1
```

Outputs `tools/ffmpeg-minimal/out/ffmpeg.exe` + `ffprobe.exe` + `MANIFEST.txt`
(the exact configure line + component list + lib versions). Nothing is staged to
the depot and the app is not launched — this is for smoke-testing first.

## Smoke matrix (run before swapping the shipped binary)

Against `out/ffmpeg.exe`:

| # | What | Command shape | Expect |
|---|------|---------------|--------|
| 1 | version/license | `ffmpeg -version` | `configuration:` shows **no** `--enable-gpl`; LGPL |
| 2 | probe duration | `ffmpeg -i in.mp3 -f null -` | exits 0, prints duration |
| 3 | decode→pcm (our path) | `ffmpeg -i in.mp3 -vn -c:a pcm_s16le -ar 44100 out.wav` | valid wav |
| 4 | mp3 export | `ffmpeg -i in.wav -codec:a libmp3lame out.mp3` | valid mp3 |
| 5 | flac export | `ffmpeg -i in.wav -c:a flac out.flac` | valid flac |
| 6 | ogg/opus export | `ffmpeg -i in.wav -c:a libvorbis out.ogg` / `libopus out.opus` | valid |
| 7 | BPM probe path | `ffmpeg -ss 0 -i in.mp3 -t 30 -vn -ac 1 -ar 22050 -f f32le pipe:1` | raw stream |
| 8 | audio out of video | `ffmpeg -i in.mp4 -vn -c:a pcm_s16le out.wav` | audio extracted |
| 9 | pydub load (Py side) | `pydub.AudioSegment.from_file("in.m4a")` with this exe on PATH | loads |
| 10 | whisper decode | run a transcribe on `in.m4a` | succeeds |

All green → safe to stage into the depot (replacing the BtbN binary) and update
the SBOM ffmpeg entry to the new minimal component set.