bliss music analyzer - Rust version

bliss-rs is the Rust improvement of bliss, a library used to make playlists by analyzing songs, and computing distance between them.

Like bliss, it eases the creation of « intelligent » playlists and/or continuous play, à la Spotify/Grooveshark Radio, as well as easing creating plug-ins for existing audio players. For instance, you can use it to make calm playlists to help you sleeping, fast playlists to get you started during the day, etc.

For now (and if you're looking for an easy-to use smooth play experience), blissify implements bliss for MPD.

There are also python bindings. The wheels are compiled used maturin; the sources are available here for inspiration.

Note: the features bliss-rs outputs are not compatible with the ones used by C-bliss, since it uses different, more accurate values, based on actual literature. It is also faster.

Dependencies

To use bliss-rs, you'll need a few packages: a C linker, ffmpeg, and the associated development packages (libavformat, libavutil, libavcodec, libavfilter, libavdevice), as well as the clang development packages. These steps are necessary to e.g. run the examples below.

On Ubuntu:

$ sudo apt install build-essential pkg-config libavutil-dev libavformat-dev \
libavfilter-dev libavdevice-dev libclang-dev

On Archlinux:

$ sudo pacman -S base-devel clang ffmpeg

If you want to use the library trait, in order to e.g. make a bliss plugin for an existing audio player, and build the examples associated with it, you will also need libsqlite3-dev on Ubuntu, or sqlite on Archlinux.

Note: it is also possible to use the Symphonia crate instead of FFmpeg to handle the decoding of songs. See the Decoders section below for more information.

The Symphonia decoder is slightly slower, but using it makes bliss-rs easier to package, since it removes a big dependency to FFmpeg.

The examples section below shows how to use both Symphonia and FFmpeg. Take a look at the decoders for more information on the two.

Examples

For simple analysis / distance computing, take a look at examples/distance.rs and examples/analyze.rs.

If you simply want to try out making playlists from a folder containing songs, this example contains all you need. Usage:

    cargo run --release --example=playlist --features=serde /path/to/folder /path/to/first/song

Don't forget the --release flag!

To use Symphonia instead of FFmpeg to decode songs, use the symphonia-all feature flag instead:

    cargo run --release --example=playlist --no-default-features --features=serde,aubio-static,symphonia-all /path/to/folder /path/to/first/song

By default, it outputs the playlist to stdout, but you can use -o <path> to output it to a specific path.

To avoid having to analyze the entire folder several times, it also stores the analysis in /tmp/analysis.json. You can customize this behavior by using -a <path> to store this file in a specific place.

Ready to use code examples:

Compute the distance between two songs

use bliss_audio::decoder::bliss_ffmpeg::FFmpegDecoder as Decoder;
use bliss_audio::decoder::Decoder as DecoderTrait;
use bliss_audio::BlissError;

fn main() -> Result<(), BlissError> {
    let song1 = Decoder::from_path("/path/to/song1")?;
    let song2 = Decoder::from_path("/path/to/song2")?;
        
    println!("Distance between song1 and song2 is {}", song1.distance(&song2));
    Ok(())
}

Make a playlist from a song

use bliss_audio::decoder::bliss_ffmpeg::FFmpegDecoder as Decoder;
use bliss_audio::decoder::Decoder as DecoderTrait;
use bliss_audio::{BlissError, Song};
use noisy_float::prelude::n32;

fn main() -> Result<(), BlissError> {
    let paths = vec!["/path/to/song1", "/path/to/song2", "/path/to/song3"];
    let mut songs: Vec<Song> = paths
        .iter()
        .map(|path| Decoder::song_from_path(path))
        .collect::<Result<Vec<Song>, BlissError>>()?;

    // Assuming there is a first song
    let first_song = songs.first().unwrap().to_owned();

    songs.sort_by_cached_key(|song| n32(first_song.distance(&song)));
    println!(
        "Playlist is: {:?}",
        songs
            .iter()
            .map(|song| &song.path)
            .collect::<Vec<&String>>()
    );
    Ok(())
}

Decoders

bliss-rs is decoder-agnostic, meaning that, if fed the proper audio format (raw single-channeled, floating-point (f32), little-endian PCM samples sampled at 22050 Hz), the analysis of the same track should return the same analysis results.

For convenience's sake, two different decoder suites have been implemented:

• FFmpeg, used in most examples and used by default. It is fast, and supports the widest array of esoteric containers. However, it requires the user to install libav* (see Dependencies), which might be inconvenient for maintainers looking to package bliss-rs.
• Symphonia, which is a pure rust audio decoding library. This means that users won't have to install a third-party program to decode songs. While it makes the decoding of songs slightly slower, its impact is neligible when compared to the time spent running the actual analysis (the whole process gets ~5-10% slower compared to decoding with FFmpeg. For instance, it took around 56 minutes to decode & analyze 10k files with FFmpeg, and around 1 hour 5 minutes with Symphonia). It might also yield very slightly different analysis results (which shouldn't impact the playlist-making process).

To turn off the FFmpeg dependency, and use Symphonia instead:

    cargo build --release --no-default-features --features=aubio-static,symphonia-all

Otherwise, FFmpeg will be used automatically.

It is also possible to turn both decoders off, and implement your own decoder pretty easily - see the decoder module.

Further use

Instead of reinventing ways to fetch a user library, play songs, etc, and embed that into bliss, it is easier to look at the library module. It implements common analysis functions, and allows analyzed songs to be put in a sqlite database seamlessly.

See blissify for a reference implementation.

If you want to experiment with e.g. distance metrics, taking a look at the mahalanobis distance function is a good idea. It allows customizing the distance metric, putting more emphasis on certain parts of the features, e.g. making the tempo twice as prominent as the timbral features.

It is also possible to perform metric learning, i.e. making playlists tailored to your specific tastes by learning a distance metric for the mahalanobis distance using the code in the metric learning repository. blissify-rs is made to work with it, but it should be fairly straightforward to implement it for other uses. The metric learning repo also contains a bit more context on what metric learning is.

Cross-compilation

To cross-compile bliss-rs from linux to x86_64 windows, install the x86_64-pc-windows-gnu target via:

    rustup target add x86_64-pc-windows-gnu

Make sure you have x86_64-w64-mingw32-gcc installed on your computer.

Then after downloading and extracting ffmpeg's prebuilt binaries, running:

    FFMPEG_DIR=/path/to/prebuilt/ffmpeg cargo build --target x86_64-pc-windows-gnu --release

Will produce a .rlib library file. If you want to generate a shared .dll library, add:

    [lib]
    crate-type = ["cdylib"]

to Cargo.toml before compiling, and if you want to generate a .lib static library, add:

    [lib]
    crate-type = ["staticlib"]

You can of course test the examples yourself by compiling them as .exe:

    FFMPEG_DIR=/path/to/prebuilt/ffmpeg cargo build --target x86_64-pc-windows-gnu --release --examples

WARNING: Doing all of the above and making it work on windows requires to have FFmpeg's dll on your Windows %PATH% (avcodec-59.dll, etc). Usually installing FFmpeg on the target windows is enough, but you can also just extract them from /path/to/prebuilt/ffmpeg/bin and put them next to the thing you generated from cargo (either bliss' dll or executable).

Acknowledgements

• This library relies heavily on aubio's Rust bindings for the spectral / timbral analysis, so a big thanks to both the creators and contributors of librosa, and to @katyo for making aubio bindings for Rust.
• The first part of the chroma extraction is basically a rewrite of librosa's chroma feature extraction from python to Rust, with just as little features as needed. Thanks to both creators and contributors as well.
• Finally, a big thanks to Christof Weiss for pointing me in the right direction for the chroma feature summarization, which are basically also a rewrite from Python to Rust of some of the awesome notebooks by AudioLabs Erlangen, that you can find here.