This is a slightly detailed list describing the main algorithms you can find in Essentia as well as a small description.
Please note that this is not an exhaustive list but rather a quick overview of the most used algorithms in Essentia. For the full list of algorithms and their detailed documentation, please see the complete reference.
Audio input / output¶
Essentia has a variety of audio loaders to provide a very convenient way to load audio files from disk:
AudioLoader: this is the most basic audio loader. It will load the given audio file and return the stream of stereo samples. If the audio file was mono, only the left channel is used. It will load virtually any format existing as it uses the FFmpeg libraries to load audio. It will also load the audio from a given video file (works even with Flash files!).
MonoLoader: probably the most used audio loader, the
MonoLoaderwill downmix the file to a mono signal and resample it to the given sample rate (default 44.1 KHz).
MonoLoaderwhich can also trim the audio to a given start/end time and normalize the resulting samples using a given ReplayGain value.
EasyLoaderthat applies an Equal-loudness filter to the resulting audio
MetadataReader: reads the metadata tags stored in the given file (e.g.: mp3 tags, …).
Essentia can also write audio files, using the following algorithms:
AudioWriter: a versatile audio writer that can write a stream of samples to any format supported by FFmpeg. Takes as input a stream of stereo samples.
AudioWriterthat takes a stream of mono samples as input.
AudioOnsetsMarker: writes an audio file given the stream of samples adding beeps at the given onset times.
Standard signal processing algorithms¶
The following algorithms are quite standard in any signal processing library:
FrameCutter: takes an input audio stream and cuts it into frames.
Windowing: returns a windowed frame (supports a lot of standard window types, such as Hann, Hamming, Blackman-Harris, …).
FFT: the omnipresent (complex) Fast Fourier Transform algorithm.
Spectrum: returns only the magnitude part of the
DCT: the type II Discrete Cosine Transform.
Envelope: an envelope follower.
PeakDetection: detects peaks in an array of values.
Resample: resamples a given audio signal.
Slicer: returns the given slices (start/end times) of an audio stream.
ReplayGain: returns the ReplayGain value for the given audio.
CrossCorrelation: computes the cross-correlation of two signals.
AutoCorrelation: computes the auto-correlation of a signal.
PolarToCartesian: do the conversion of complex arrays between the polar and the cartesian coordinate systems.
The following algorithms compute statistics over an array of values, or do some kind of aggregation:
Median: do what their names hint at.
Energy: computes the energy of the array.
RMS: computes the Root mean square of the array.
SingleGaussian: computes a single Gaussian estimation for the given list of arrays (returns the mean array, its covariance and inverse covariance matrices)
RawMoments: compute the moments up to the 5th-order of the array.
Kurtosis: are often used to describe probability distributions.
DistributionShape: combines the
Kurtosisvalues in a single vector.
InstantPowerare some others lesser used algorithms.
Essentia implements a nice variety of audio filters:
IIR: a generic algorithm that does IIR filtering, where you can specify the coefficients manually.
HighPass: do 1st or 2nd order low-, band- and high-pass filtering.
BandReject: do band-pass and band-rejection filtering.
DCRemoval: removes the DC component of a signal.
EqualLoudness: filters the signal using an equal-loudness curve approximating filter.
The following algorithms compute low/mid/high-level descriptors frequently used in the Music Information Retrieval community.
BarkBands: computes the Bark band energies.
MelBands: computes the Mel band energies.
ERBBands: computes the energies in bands spaced on an Equivalent Rectangular Bandwidth scale.
MFCC: computes the Mel-frequency cepstral coefficients of a frame.
GFCC: computes the gammatone feature cepstrum coefficients similar to MFCCs.
LPC: computes the Linear Predictive Coding coefficients of a frame as well as the associated reflection coefficients.
HFC: computes the High-Frequency Content measure.
SpectralContrast: computes spectral contrast of a spectrum.
Dissonance: both try to estimate whether an audio frame “sounds” harmonic or not.
SpectralWhitening: whitens the input spectrum.
Panning: computes the panorama distribution of a stereo audio frame.
EffectiveDuration: compute the total duration of a signal and the duration of the signal being above a certain energy level.
ZCR: computes the Zero-crossing rate of the signal.
LoudnessVicker: are different loudness measures.
PredominantMelody: compute pitch salience function and estimate the fundamental frequency of the predominant melody by the MELODIA algorithm.
PitchYinFFT: estimates pitch of a signal by YinFFT algorithm.
HPCP: computes the Harmonic Pitch-Class Profile of a spectrum (also called Chroma features).
TuningFrequency: returns the exact frequency on which a song is tuned and the number of cents to 440Hz.
Key: returns the key and scale of a song.
ChordsDetection: computes the sequence of chords in a song.
ChordsDescriptors: computes some descriptors associated with the sequence of chords, such as its histogram, etc.
BeatTrackerDegara: the beat tracker based on complex spectral difference feature.
BeatTrackerMultiFeaturethe multifeature beat tracker (combines 5 different beat trackers taking into account the maximum mutual agreement between them.
RhythmExtractor2013: сomputes BPM of a song in addition to the estimated beat positions (using either
BpmHistogramDescriptors: computes statistics of the BPM histogram of a song.
NoveltyCurve: computes the novelty curve for the audio signal.
OnsetDetectionGlobalestimate various onset detection functions useful for beat tracking and onset detection.
Onsets: computes the list of onsets in the audio signal.
RhythmTransform: computes a rhythmical representation based on the FFT over temporal windows of MFCC frames.
BeatsLoudness: computes the loudness of the signal on windows centered around the beat locations.
Most of the algorithms are designed to deal with music audio files, but the following ones are intended to be used with short sounds instead of full-length music tracks:
LogAttackTime: returns the logarithm of the attack time for the sound.
MinToTotal: return a measure of whether the max (respectively, min) value of a sound is located towards its beginning or end.
PitchSalience: returns whether the pitch is salient (that is, strongly marked) in a sound.
TCToTotal: computes the normalized position of the temporal centroid.
Other high-level descriptors¶
Essentia also contains the following mid- and high-level descriptors:
Danceability: returns whether a song is “danceable”.
DynamicComplexity: returns whether a song has a high dynamic range.
FadeDetection: detects fade-ins/fade-outs in a song.
SBic: segments a song using the Bayesian Information Criterion.
PCA: does the Principal Component Analysis of the given list of arrays.
GaiaTransform: applies the given Gaia transformation to a Pool. This is mostly used for classifiers which have been trained with Gaia.
You can train your own classifier models.
In addition, we provide a number of pre-trained classifier models:
- musical genre (4 different databases)
- ballroom music classification
- moods: happy, sad, aggressive, relaxed, acoustic, electronic, party
- western / non-western music
- tonal / atonal
- voice / instrumental
- gender (male, female singer)
- timbre (dark, bright)
These models were trained on a number of annotated music collections, including various inhouse collections created at Music Technology Group. To run them, use the standalone Music Extractor available in Essentia 2.1_beta2.
As Essentia algorithms can themselves be composed of multiple algorithms (see Composite algorithms), a few useful extractors have been written as algorithms. They are the following:
LevelExtractor: computes the loudness of a music track.
LowLevelSpectralExtractor: computes a lot of low-level features from a music stream.
LowLevelSpectralEqloudExtractor: computes a lot of low-level features which require preliminary equal-loudness filter from a music stream.
TuningFrequencyExtractor: computes the tuning frequency of a music track.
KeyExtractor: computes the key and scale of a music track.
TonalExtractor: computes the tonal information of a music track (key, scale, chords sequence, chords histogram, …)
RhythmDescriptors: computes the rhythm information of music track (beat positions, BPM and related histogram statistics).
Extractor: extracts pretty much all the features useful as descriptors for doing music similarity.
For more information about
Algorithms, please see the complete reference.
For information on the other types and classes of Essentia which are not
Algorithms, see the Design Overview page.
For more advanced examples, you can also look at the src/examples directory of Essentia’s git repository.