This is a compilation of the libraries associated with handling audio and video
in FFmpeg—libavformat, libavcodec, libavfilter, libavutil and libswresample—for
WebAssembly and asm.js, and thus the web, as well as the ffmpeg
and ffprobe
CLIs themselves. It is compiled via emscripten, is highly customizable, and has
a ruthless commitment to correct licensing.
In short, this is a pure JavaScript and WebAssembly system for low-level audio and video encoding, decoding, muxing, demuxing, and filtering.
FFmpeg is released under the LGPL. Therefore, if you distribute this library,
you must provide sources. The sources are included in the sources/
directory
of the compiled version of libav.js.
In order to reduce license-header Hell, the small amount of wrapper functions provided by libav.js are all released under the so-called “0-clause BSD” license, which does not require that the license text itself appear in derivative works. Built libraries have their correct license headers.
This file is the main README for using and building libav.js, and should be sufficient for many users. More detail on specific concepts is provided in other files:
-
API.md describes the libav.js-specific parts of the API.
-
CONFIG.md describes the configuration system and how to create your own configuration of libav.js.
-
IO.md describes the various I/O modes provided by libav.js.
-
TESTS.md describes the testing framework.
libav.js builds are available on
GitHub and in NPM. Include
dist/libav-version
-variant
.js to use libav.js. The variants are discussed
below.
The simplest way to use libav.js is to include it from a CDN. libav.js uses Web Workers by default, and Web Workers cannot be loaded from a different origin, so if you load it from a CDN, you must disable its own loading of workers. As such, it's only recommended to use libav.js from a CDN if you're already in a worker, and thus don't need sub-workers. Nonetheless, the following is a simple example of using libav.js from a CDN in the browser thread:
<!doctype html>
<html>
<body>
<script type="text/javascript" src="https://cdn.jsdelivr.net/npm/@libav.js/variant-default@5.1.6/dist/libav-5.1.6.1.1-default.js"></script>
<script type="text/javascript">(async function() {
const libav = await LibAV.LibAV({noworker: true});
await libav.writeFile("tmp.opus", new Uint8Array(
await (await fetch("exa.opus")).arrayBuffer()
));
const [fmt_ctx, [stream]] = await libav.ff_init_demuxer_file("tmp.opus");
const [, c, pkt, frame] = await libav.ff_init_decoder(stream.codec_id, stream.codecpar);
const [, packets] = await libav.ff_read_multi(fmt_ctx, pkt);
const frames = await libav.ff_decode_multi(c, pkt, frame, packets[stream.index], true);
alert(`Got ${frames.length} audio frames!`);
})();
</script>
</body>
</html>
Here's a better example, using libav.js locally:
<!doctype html>
<html>
<body>
<script type="text/javascript" src="libav-5.1.6.1.1-default.js"></script>
<script type="text/javascript">(async function() {
const libav = await LibAV.LibAV();
await libav.writeFile("tmp.opus", new Uint8Array(
await (await fetch("exa.opus")).arrayBuffer()
));
const [fmt_ctx, [stream]] = await libav.ff_init_demuxer_file("tmp.opus");
const [, c, pkt, frame] = await libav.ff_init_decoder(stream.codec_id, stream.codecpar);
const [, packets] = await libav.ff_read_multi(fmt_ctx, pkt);
const frames = await libav.ff_decode_multi(c, pkt, frame, packets[stream.index], true);
alert(`Got ${frames.length} audio frames!`);
})();
</script>
</body>
</html>
It's also possible to use libav.js from Node.js, though this isn't a good idea, since you can presumably use a native version of FFmpeg's libraries. The Node interface is only provided for internal testing.
Use .dbg.js
instead of .js
for a non-minified, more debuggable version. Use
.mjs
for the ES6 module version. Use .dbg.mjs
for both. You don't need any
combination; e.g., if you only intend to use imports, you do not need any .js
files.
libav.js exposes a global variable, LibAV
, for all API access. If importing as a
module, LibAV
is the default export.
For certain unusual loading situations, you can set the LibAV
global variable
before importing. In particular, if the base directory (directory in which
libav's files are located) can't be detected for some reason, then you must set
LibAV.base
to the correct base. LibAV.base
does not need to be a full URL,
but should be if loading from another origin.
Bundlers have further concerns. To use libav.js with a bundler, see the section on bundlers below.
LibAV.LibAV
is a factory function which returns a promise which resolves to a
ready instance of libav. The factory function and libav instance methods are
documented in API.md.
You need the main entry file and at least one target, for a minimum of three files, but you should probably include several others.
The main entry file is named as follows: libav-<version>-<variant>.js
. You
only need the variant you intend to use. The debug version is named
libav-<version>-<variant>.dbg.js
, and you can use that in place of the
original, but it is not required. If using ES6 modules, use mjs
in place of
js
.
That entry file will load a target based on the environment it's loaded in and
the options used to load it, as described above. The supported targets are
asm.js, plain WebAssembly, and threaded WebAssembly. It is harmless to include
all of them, as users will not download all of them, only the ones they use.
But, you may also include only those you intend to use. In every case, there is
a .dbg.js
equivalent which is only needed if you intend to use debug mode.
-
asm.js: Named
libav-<version>-<variant>.asm.js
. No modern browser excludes support for WebAssembly, so this is probably not necessary. -
Plain WebAssembly: Named
libav-<version>-<variant>.wasm.js
andlibav-<version>-<variant>.wasm.wasm
. Used in most situations. -
Threaded WebAssembly: Named
libav-<version>-<variant>.thr.js
,.thr.wasm
, and.thr.worker.js
. Used only when threading is supported by the browser andyesthreads
is set. If you don't intend to use threads (setyesthreads
), it is safe to exclude this. Used only when threads are activated and supported.
At a minimum, it is usually sufficient to include only the .js
, .wasm.js
,
and .wasm.wasm
files. To include threads, you must also include .thr.js
and
.thr.wasm
. Again, use mjs
instead of js
if using ES6 imports.
The file libav.types.d.ts
is a TypeScript types definition file, and is only
needed to compile TypeScript code with support for libav.js's types. It should
never be necessary to distribute.
Note that, independently of what files are available to end users, you are
contractually obligated to release the source code of libav.js and all of its
dependencies if you provide the compiled version. If you are using a compiled,
released version, it is sufficient to provide the sources
directory.
libav.js is published to NPM as libav.js
, and each released variant is
published in a much smaller NPM package as @libav.js/variant-<variant>
. The
CDN example above uses the @libav.js/variant-default
package, for example.
Caching is Hell.
libav-<variant>.*
is also available in the releases and repository, but you're
highly recommended not to use this name on any web installation, as caching
will cause strange nonsense to happen. Use a full versioned name to avoid
caching madness.
Emscripten's implementation of an in-memory filesystem has severe limitations.
You're recommended to use virtual devices, implemented by libav.js
, for most
I/O. See IO.md for more details. libav.js itself imposes no
restriction on file sizes so long as you use asynchronous, device-backed I/O
(thus, the only restriction to size is JavaScript's number type).
ffmpeg was never designed to work asynchronously, and was only designed to work with blocking I/O. Still, it's possible to use libav.js with asynchronous input through these devices.
Type definitions for libav.js are provided by libav.types.d.ts
. You can
either copy this file and import it:
import type LibAVJS from "./libav.types";
declare let LibAV: LibAVJS.LibAVWrapper;
or import it from the npm package:
import type LibAVJS from "libav.js";
declare let LibAV: LibAVJS.LibAVWrapper;
With all of its bells and whistles enabled, FFmpeg is pretty large. So, I disable most bells and most whistles and build specific versions with specific features.
The default variant, libav-<version>-default.js
, includes support for the most
important (and timeless) audio codecs and formats: Opus, FLAC, and wav, in WebM,
ogg, FLAC, or wav containers. It also has a set of common audio filters.
Built-in variants are created by combining “configuration fragments”. You can find more on configuration fragments or making your own variants in CONFIG.md.
Use make build-<variant>
, replacing <variant>
with the variant name, to
build another variant.
Most of the variants provided in the repository are also built and available in NPM and as binary releases. The notable exception is all variants that include codecs controlled by the Misanthropic Patent Extortion Gang (MPEG). They are not built by default, and if you have any sense, you should not use them. MPEG is a cancer on the digital media ecosystem.
The included variants and their codecs and formats are:
-
default, default-cli: Opus (via libopus), FLAC, and wav in ogg, WebM, FLAC, and wav containers, plus audio filters. The
-cli
subvariant additionally includes the CLI (ffmpeg
andffprobe
functions). -
opus, opus-af: Opus in ogg or WebM.
-af
additionally includes audio filters. -
flac, flac-af: FLAC in ogg or FLAC.
-af
additionally includes audio filters. -
wav, wav-af: PCM wav (16-bit or 24-bit) in wav.
-af
additionally includes audio filters. -
obsolete: Same as default with the addition of two obsolete codecs, Vorbis (via libvorbis) and MPEG-1 Layer 3 (MP3) (via libmp3lame). Also includes the MP3 container format.
-
webm, webm-vp9¹, webm-cli, webm-vp9-cli¹: Same as default with the addition of VP8 (via libvpx) and video filters.
-vp9
additionally includes VP9,-cli
additionally includes the CLI.-vp9
is separated due to the rather significant size of the VP9 codec. -
webcodecs, webcodecs-avf: Designed to serve as a demuxer/muxer for codecs supported by WebCodecs. Pairs well with libavjs-webcodecs-bridge. Includes codecs for Opus, FLAC, and wav. Includes parsers (but not codecs) for AAC, VP8, VP9, AV1, H.264, and H.265. Includes the ogg, WebM, MP4, FLAC, and wav formats. This means that it can demux files including, e.g., H.264, but cannot decode the frames. If your WebCodecs supports H.264, you can then use it to decode.
-avf
additionally includes audio and video filters. -
vp8-opus, vp8-opus-avf: VP8 and Opus in WebM (or ogg).
-avf
additionally includes audio and video filters. -
vp9-opus¹, vp9-opus-avf¹: VP9 and Opus in WebM (or ogg).
-avf
additionally includes audio and video filters. -
av1-opus¹, av1-opus-avf¹: AV1 (via libaom) and Opus in WebM (or ogg). Note that AV1 support is currently so slow in WebAssembly even with threads that these variants are effectively unusable.
-avf
additionally includes audio and video filters. -
aac², aac-af²: Reprobate codec AAC in MP4 or AAC/ADTS.
-af
additionally includes audio filters. -
h264-aac², h264-aac-avf²: Reprobate codec H.264 (via libopenh264) and reprobate codec AAC in MP4 (or AAC/ADTS).
-avf
additionally includes audio and video filters. -
hevc-aac², hevc-aac-avf²: Reprobate codec H.265 (decoding only) and reprobate codec AAC in MP4 (or AAC/ADTS).
-avf
additionally includes audio and video filters.
¹ These builds are not included in the full NPM release for space reasons, but are
included in GitHub releases, and are available on NPM as
@libav.js/variant-<variant>
, e.g. @libav.js/variant-vp9-opus
.
² Includes technologies patented by the Misanthropic Patent Extortion Gang (MPEG). You should not build these, you should not use these builds, and you should not support this organization which works actively against the common good.
This is intentionally designed so that you can add new variants without needing to patch anything that already exists. If you want to create your own variants, see CONFIG.md.
You can also build against different versions of FFmpeg than the version built
by default. To build against, for instance, FFmpeg 4.3.6, use make FFMPEG_VERSION_MAJOR=4 FFMPEG_VERSION_MINREV=3.6
. Note that you must use
FFMPEG_VERSION_MAJOR
and FFMPEG_VERSION_MINREV
, not just FFMPEG_VERSION
,
because FFMPEG_VERSION_MAJOR
is used to direct the process of patching FFmpeg.
libav.js should generally build against any version of FFmpeg in the 4, 5, or 6
series, but is not heavily tested against older versions; you should use the
default version unless you have some specific compatibility issue that forces
you to use a different version.
FFmpeg is big, so libav.js is big. But, it's not ludicrous; the WebAssembly is usually between 1.5 and 3 MiB for fairly complete builds, and the asm.js is about double that.
You can estimate the size of variants based on the size of the constituent
fragments. As of version 5.0.6.1.1, an empty build is approximately 589KiB
(WebAssembly). The sizes of each additional fragment can be found in
fragment-sizes.csv. The data in that CSV file can be
recreated by tools/fragment-sizes.sh
, but note that the CSV file in the
repository is after further processing (in particular, normalizing to KiB and
subtracting away the empty size).
The asm.js versions are much bigger, but will not be loaded on WebAssembly-capable clients.
The wrapper (“glue”) code is about 292KiB, but is highly compressible.
Generally speaking, the performance of audio en- and decoding is much faster than real time, to the point that it's simply not a concern for most applications. The author of libav.js regularly uses libav.js in live audio systems.
Video is a different story, of course.
Video is nowhere near as slow as you might imagine. On reasonable systems, faster-than-real-time performance for decoding of up to 1080P is achievable if you use a threaded version of libav.js. If you're willing to use older, simpler video codecs and lower-resolution video, even real-time encoding is possible. But, for complex codecs, real-time en/decoding is not realistic. One of the revolutions of video en/decoding is hardware en/decoding, and libav.js cannot do that, so its performance ceiling is already low.
Muxing and demuxing are bound by I/O time, not software performance. libav.js will always mux or demux faster than you can use the data.
On some modern browsers, the WebCodecs API is availble for hardware-accelerated (or at least, CPU-specific) en/decoding of various codecs. When it is available, it is better to use it than libav.js. However, WebCodecs does not mux or demux, and which codecs it supports varies based on the moods of its implementor (if it is even present), so generally, it is necessary to support WebCodecs but fall back to libav.js when necessary.
To make this easier, two companion projects to libav.js are provided that connect it to WebCodecs:
-
libavjs-webcodecs-polyfill is a polyfill for the WebCodecs API using libav.js. Even if WebCodecs exists on your browser, this polyfill allows the user to guarantee a certain set of supported codecs; any codecs not supported by the built-in WebCodecs can simply fall back to libav.js, using only one API.
-
libavjs-webcodecs-bridge is a bridge between libav.js and WebCodecs, converting between the two data formats. This makes it easy to use libav.js for demuxing and WebCodecs for decoding, or WebCodecs for encoding and libav.js for muxing. Of course, the WebCodecs used with the bridge can easily be the polyfill if needed.
Generally speaking, because libav.js needs to adjust its loading procedure based on the environment it's being loaded in, it's not a good idea to bundle libav.js. However, if you have to bundle it, it can be done if necessary.
libav.js has a frontend (libav-<version>-<variant>.js
), a factory
(libav-<version>-<variant>.wasm.js
or .thr.js
), and, if using WebAssembly, a
backend (libav-<version>-<variant>.wasm.wasm
or .thr.wasm
). Any of these can
be overridden, and any of them can be object URLs to bundle everything, though
this will destroy libav.js's ability to load the correct version for the system.
To override the frontend, simply load a different frontend!
To override the factory, you have two choices:
-
Pass
toImport
, a string, toLibAV.LibAV
's options, e.g.,LibAV.LibAV({toImport: "libav-but-better.wasm.js"})
. -
Load the factory yourself, and pass the factory function as the
factory
option toLibAV.LibAV
, e.g.,LibAV.LibAV({factory: LibAVFactory})
. By default, the factory function is exported asLibAVFactory
, or for ES6 modules, it is the default export of the module.
To override the backend, you can pass the full URL (or object URL) to the
WebAssembly as the option wasmurl
to LibAV.LibAV
, e.g.,
LibAV.LibAV({wasmurl: URL.createObjectURL(...)})
.
Be careful about which versions of things you bundle. The ES6 module version of
libav.js assumes that it will actually be an ES6 module, and so will be able
to use, e.g., import
. If your bundler transforms it into a non-ES6 module, you
must explicitly tell it to import some other way by passing the option noes6
to LibAV.LibAV
, e.g., LibAV.LibAV({noes6: true})
. Also, if your bundler
converts the ES6 frontend to non-ES6 and you intend to explicitly specify
toImport
, you must specify the non-ES6 factory.