soundalike

soundalike

soundalike is a command-line program that tries to find similar audio files by comparing acoustic fingerprints. Its main focus is identifying duplicate songs in music collections.

Fingerprints are generated using the fpcalc utility from the Chromaprint library (which does basically all of the heavy lifting). No network requests are made to AcoustID or other APIs.

Installation

You may be able to download a precompiled soundalike executable for your operating system and architecture from the Releases page.

Otherwise, to compile and install the soundalike executable, run go install from the root of this repository. You will need to have Go installed.

Chromaprint's fpcalc program must also be in your path. On a Debian system, it can be installed by running:

sudo apt install libchromaprint-tools

Otherwise, you can download the latest Chromaprint release for your operating system and architecture and place the fpcalc or fpcalc.exe executable in your path (e.g. in C:\Windows\System32 on Windows).

Usage

soundalike scans all of the audio files that it finds in the specified directory and then prints groups of similar files.

Usage: soundalike [flag]... <DIR>
Find duplicate audio files within a directory.

  -compare
    	Compare two files given via positional args instead of scanning directory
    	(increases -fpcalc-length by default)
  -compare-interval int
    	Score interval for -compare (0 to print overall score)
  -db string
    	SQLite database file for storing file info (temp file if unset)
  -exclude
    	Update database to exclude files in positional args from being grouped together
  -file-regexp string
    	Regular expression for audio files (default "(?i)\\.(aiff|flac|m4a|mp3|oga|ogg|opus|wav|wma)$")
  -fpcalc-algorithm int
    	Fingerprint algorithm (default 2)
  -fpcalc-chunk float
    	Audio chunk duration in seconds
  -fpcalc-length float
    	Max audio duration in seconds to process (default 15)
  -fpcalc-overlap
    	Overlap audio chunks in fingerprints
  -log-sec int
    	Logging frequency in seconds (0 or negative to disable logging) (default 10)
  -lookup-threshold float
    	Threshold for lookup table in (0.0, 1.0] (default 0.25)
  -match-min-length
    	Use shorter fingerprint length when scoring bitwise comparisons
  -match-threshold float
    	Threshold for bitwise comparisons in (0.0, 1.0] (default 0.95)
  -print-file-info
    	Print file sizes and durations (default true)
  -print-full-paths
    	Print absolute file paths (rather than relative to dir)
  -skip-bad-files
    	Skip files that can't be fingerprinted by fpcalc (default true)
  -skip-new-files
    	Skip files not already in database given via -db
  -version
    	Print version and exit

Example output when scanning the current directory:

% soundalike .
2022/02/12 08:49:49 Finished scanning 67 files
eva02.mp3    3.07 MB  120.45 sec
Eva_Two.mp3  1.84 MB  120.46 sec

hedgehogs_dilemma.mp3  3.82 MB  167.37 sec
Hedgehogs_Dilemma.mp3  2.55 MB  167.35 sec

...

-compare can also be used to compare two files:

% soundalike -compare \
  fly_me_to_the_moon_instrumental_version.mp3 \
  Fly_Me_To_The_Moon_Instrumental.mp3
0.972
% soundalike -compare \
  fly_me_to_the_moon_instrumental_version.mp3 \
  Fly_Me_To_The_Moon_Instrumental_2.mp3
0.202

Add -compare-interval 100 to instruct -compare to print a score after comparing every 100 value pairs within the two fingerprints instead of printing an overall score. This can help approximate the point at which differences between two songs occur:

% soundalike -compare -compare-interval 100 instrumental.mp3 vocals.mp3
 100: 0.983
 200: 0.984
 300: 0.983
 400: 0.989
 500: 0.819
 600: 0.757
 700: 0.751
 800: 0.753
...

The -match-min-length flag may be helpful for detecting songs that have been truncated or that contain added silence at their beginnings or ends.

How it works

Chromaprint's fpcalc utility splits audio into overlapping frames and uses Fourier transforms to identify (Western) notes, eventually generating fingerprints consisting of sequences of 32-bit unsigned integers (in my experience, ~100 for a 15-second sample and ~462 for a minute).

soundalike runs fpcalc on each file and maintains a lookup table from values from fingerprints (truncated from 32 to 16 bits) to files. Fingerprint values (irrespective of order) are compared between files. If -lookup-threshold or more of the values match, the file pair moves on to the next phase.

The two fingerprints are then compared in their original order, with all possible alignments, to see how many bits match between the two. If more than -match-threshold of the bits are identical, the file pair is retained.

Pairs are treated as edges in an undirected graph, and files are grouped into components. Finally, each group is printed.

When -compare is passed, only the final bitwise comparison is performed.

Accuracy

Vocal and instrumental versions of the same song can be problematic. With the -fpcalc-length flag's default value only the first 15 seconds of each file are fingerprinted, so if the vocals only come in later in the song the files will appear to be identical. A larger value can be passed at the the expense of performance and memory consumption.

I've also sometimes seen false positives between tracks that start with silence, or between electronic tracks that start with the same chords or with similar drumbeats. Passing a higher -match-threshold value may prevent some of these.

When running soundalike repeatedly with the -db flag, -exclude can be used to specify false positives that should be excluded from future runs.

fpcalc isn't able to generate fingerprints for very short files, so they're silently skipped. The cutoff seems to be somewhere around 2 to 3 seconds.

Performance

Performance is largely dependent on the -fpcalc-length flag's value.

On a laptop with an Intel Core i5-8250U CPU 1.60GHz processor, soundalike takes about 10 seconds to scan 99 MP3 and WAV files totalling 266 MB using the default flags (i.e. fingerprinting up to 15 seconds of each file).

A much slower system with an Intel Celeron 2955U 1.40GHz processor takes roughly 3 minutes to scan 1,000 "song-length" MP3 files with default flags (so, about 20,000 songs per hour).

When running against a large music collection, the -db flag can be passed to save fingerprints and other file information for future runs. Note that the database will not be reusable if you pass different -fpcalc-* flags in the future, since those change how fpcalc computes fingerprints.

When running soundalike repeatedly, -skip-new-files can be used to avoid repeatedly trying to fingerprint bad/corrupted files.

Memory usage

Memory usage grows with the number of files that are scanned and the fingerprint length, since (truncated) values from fingerprints are stored in memory to enable searching for collisions. I saw the resident set size (RSS in ps, RES in top) grow to 78,116 KB while scanning a bit over 20,000 MP3 files with 15-second fingerprints.

More information

The following pages contain additional technical details that may be of interest:

• Acoustid's Chromaprint page
• Lukáš Lalinský's "How does Chromaprint work?" post
• "question about using chromaprint to identify the same tracks" from Acoustid mailing list

See also the Picard music tagger from the MusicBrainz project.