Skip to main content


A Dataset of Music Production Metadata

4 min read

A Dataset of Music Production Metadata

The Semantic Audio Feature Extraction Dataset (SAFE-DB) is a continually updating database of semantically annotated music production metadata, taken from an international user group of sound engineers. The data is taken from 4 audio effects: a dynamic range compressor, an overdrive distortion, a parametric equaliser and an algorithmic reverb. Each entry into the dataset contains the following items:


A string representing a user’s description of the audio transformation (e.g. warm, bright, fluffly).

Audio Features

An array of audio features extracted from the audio file, taken before and after processing is applied to the signal. Feature list:

  • Mean
  • Variance
  • StdDev
  • RMS_Amplitude
  • Zero_Crossing_Rate
  • Spectral_Centroid
  • Spectral_Variance
  • Spectral_Standard_Deviation
  • Spectral_Skewness
  • Spectral_Kurtosis
  • Irregularity_J
  • Irregularity_K
  • F0 (autocorrelation)
  • Smoothness
  • Spectral_Roll_Off
  • Spectral_Flatness
  • Spectral_Crest
  • Spectral_Slope
  • Peak_Spectral_Centroid
  • Peak_Spectral_Variance
  • Peak_Spectral_Standard_Deviation
  • Peak_Spectral_Skewness
  • Peak_Spectral_Kurtosis
  • Peak_Irregularity_J
  • Peak_Irregularity_K
  • Peak_Tristimulus (x3)
  • Inharmonicity
  • Harmonic_Spectral_Centroid
  • Harmonic_Spectral_Variance
  • Harmonic_Spectral_Standard_Deviation
  • Harmonic_Spectral_Skewness
  • Harmonic_Spectral_Kurtosis
  • Harmonic_Irregularity_J
  • Harmonic_Irregularity_K
  • Harmonic_Tristimulus (x3)
  • Noisiness
  • Bark_Coefficients (x25)
  • Parity_Ratio
  • MFCCs (x13)

Audio Effect Parameters

A variable sized array of plug-in parameters used to apply processing to the signal.

User Data

A series of strings representing the user’s age, location, language and experience, along with the genre of the music and the instrument being processed. These fields can be left blank.

Note: To limit the file-size of the download, feature-sets are averaged. For data from a specific date, or for multi-channel time series features please get in touch.


  • Dynamic Range Compressor – User Data | Feature Data
  • Overdrive Distortion - User Data | Feature Data
  • Parametric Equaliser – User Data | Feature Data
  • Algorithmic Reverb - User Data | Feature Data
  • Download All (.zip)


R. Stables, S. Enderby, B. De Man, G. Fazekas, and J. D. Reiss, “SAFE: A system for the extraction and retrieval of semantic audio descriptors”, The International Society for Music Information Retrieval (ISMIR), 2014.

Data Visualisation

Using multidimensional scaling, we can map the descriptive terms from the dataset into a more intuitive space. The interactive visualisation below shows the feature space of the SAFE-DB in 3-dimensions.

Live demo coming soon

Find more applications