Guitar playing techniques dataset (GPT)

Guitar is probably the most popular medium for music among people. Widely used in popular music, rock music, folk music…etc, learning to play guitar is a worldwide interest among various people. In particular, the use of various guitar playing techniques is essential to express different emotions, make different sounds, and, thus, create diverse music.

For guitar, a practical, interpretable automatic transcription system should provide information about playing techniques in addition to information about pitch or onset. To investigate the guitar playing technique recognition problem, we established a database consisting of 6,580 clips across 7 frequently used playing techniques and 7 different tones.

The 7 techniques are described below:

Technique Description # of clips
Normal Normal sound 2,009
Muting Sounds muted (by right hand) to create great attenuation 385
Vibrato Trilled sound produced by twisting left hand finger on the string 637
Pull-off Sound similar to normal but with the smoother attack created by pulling off the string by left hand finger 525
Hammer-on Sound similar to normal but with the smoother attack created by hammering on the string by left hand finger 581
Sliding Discrete change to the target note with a smooth attack by left hand finger sliding through the string 1162
Bending Continuous change to the target note without an apparent attack by bending the string by left hand fingers 1281

To make the quality of the sound recordings akin to that of real-world performance, we augment the single clean tone source to 7 different guitar tones, which are described below based on the differences in effect and equalizer settings:

Tone name Effect Equalizer
1 (Normal tone) moderate distortion no modification on EQ
2 (Solo tone) moderate distortion and moderate reverb mid-frequency is emphasized
3 (Solo tone) moderate distortion, intense chorus, slight reverb mid-frequency is emphasized
4 (Solo tone) moderate distortion, intense delay, moderate reverb mid-frequency is emphasized
5 (Riff tone) Intense distortion mid-frequency is suppressed while high-frequency and low-frequency are emphasized
6 (Country tone) very slight distortion no modification on EQ
7 (Funk tone) slight distortion, slight delay, and slight reverb high-frequency component is emphasized a little

Note that the 7 tones are different in other ways apart from effect and equalizer. For example, each tones are generated through different amplifier simulators, which affect a lot but hard to describe clearly.

The dataset is recorded by a professional guitarist using a recording interface, PreSonus’ AudioBox USB, with bit depth of 24 bits and frequency response from 14 Hz to 70 kHz. We directly line-in the guitar to recording interface to catch every nuance of sound and exclude environmental noise. The guitar for recording is ESP’s MII with Seymour Duncan’s pickup and Ebony finger board, which is a high-quality guitar especially for metal and rock music. The tuning of different tones is done in the post-production stage using music production software Cubase.

If you make use of this dataset for academic purposes, please cite the following publication:

Li Su, Li-Fan Yu, and Yi-Hsuan Yang, “Sparse Cepstral and Phase Codes for Guitar Playing Technique Classification”, in 15th International Society for Music Information Retrieval Conference, Taipei, Taiwan, Oct. 2014.

Please download the full datasethere(3.4 GB).

For more information, please contact Li Su (