TLDR: This paper presents a machine learning method for converting MIDI musical parts into guitar tablature. It uses a deep neural network to generate a “probabilistic tablature” and a search algorithm that applies guitar playability rules to select the best string-fret combinations, considering previous finger positions. The system was trained on the DadaGP dataset, with an augmented version to handle music not originally meant for guitar. Results show that training with augmented data improves performance, though future work aims to incorporate future musical context and refine the playability search algorithm.
Guitar tablature, often called “tabs,” offers a unique way for musicians to learn and play guitar. Unlike traditional music scores that focus on abstract sound characteristics like pitch, tablature directly shows where to place fingers on the fretboard – specifying the string and fret number for each note. This makes it particularly valuable for beginners and self-taught players, providing an active connection with the instrument.
However, generating accurate and playable guitar tablature from a musical piece, especially one represented in MIDI format, is a complex task. The main difficulty arises because the same musical pitch can often be played at multiple positions on a guitar’s fretboard. This inherent ambiguity makes it challenging for automated systems to decide the “best” string-fret combination, especially when considering playability and natural finger movements.
Researchers have explored various methods for tablature transcription. Some approaches focus on directly converting audio into tablature by analyzing sound characteristics like inharmonicity to detect pitch and string. Others tackle the problem by first detecting pitches (from audio, MIDI, or sheet music) and then converting these pitches into plausible string-fret combinations, often relying on playability constraints. Techniques like Hidden Markov Models, weighted directed acyclic graphs, dynamic programming, and even genetic algorithms have been employed to find optimal fingering sequences.
More recently, machine learning, particularly Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs), have been applied to this challenge. These methods aim to learn complex patterns from data to predict chord recognition, hand positions, or even compose tablature. This paper, titled “A Machine Learning Approach for MIDI to Guitar Tablature Conversion,” introduces a novel method that leverages deep neural networks to address this problem. You can read the full research paper here.
The proposed method focuses on converting MIDI pitch information into guitar tablature for standard 6-string guitars with 24 frets. It simplifies the transcription task by considering only the MIDI pitches to be transcribed, disregarding aspects like velocity or duration. The system operates in two main stages: first, a deep neural network generates a “probabilistic tablature” indicating the likelihood of certain string-fret combinations; second, a search algorithm uses a concept of guitar “playability” to select the best actual fingering, taking into account previous finger positions and the network’s probabilistic output.
The deep neural network architecture, which includes convolutional and transposed convolutional layers, is designed to learn fingering shapes and temporal dependencies between successive tablature frames. It takes as input not only the current MIDI pitches but also the four previous tablature frames, providing context for finger movement. This input is a binary vector representing active MIDI pitches and string-fret combinations. The network then outputs a “probabilistic tablature” – a non-binary representation where values indicate the probability of a string-fret pair being active.
After the network generates this probabilistic tablature, a crucial analytical step converts it into actual, playable fingerings. This involves several considerations: adjusting pitches to fit within the fretboard range, deciding how many pitches (up to six for a 6-string guitar) can be played simultaneously, and generating all possible “playable” binary fretboards. A fretboard is considered playable if it has at most one pitch per string and all non-open string pitches fall within a six-fret window (to account for finger stretching limits). From these playable options, the system selects the one that best aligns with the network’s probabilistic output, maximizing a calculated inner product.
The researchers used the DadaGP dataset, comprising over 26,000 guitar pieces in GuitarPro format, for training and testing. To make the system more robust and capable of transcribing music not originally intended for guitar, a data augmentation method was introduced. This involved artificially adding extra pitches (e.g., an octave up or down, a fifth, a third) to the MIDI input, forcing the network to learn to be selective and ignore pitches that don’t fit a playable guitar context. This augmentation aimed to make the system more adaptable to complex or non-guitaristic musical parts.
Results showed that training the system with this augmented dataset generally led to better performance, even in simpler, single-pitch scenarios where augmentation might not have directly occurred. The augmented training improved both partial and exact matches when tested on both guitar-only and augmented datasets. However, the study also highlighted areas for improvement. One identified limitation is the system’s current inability to incorporate “future” information, meaning it only considers past tablature frames when making decisions. Incorporating future context could lead to more musically intuitive and playable tablature sequences.
Also Read:
- LadderSym: A New AI Model for Advanced Music Practice Error Detection
- Uncovering Sampled Music: A New Approach to Automatic Identification
Another observed issue was the “greedy” nature of the analytical step, which prioritizes fitting the maximum number of requested pitches onto the fretboard, even if those combinations have very low probabilities according to the neural network. Future work aims to refine this step to allow for solutions with fewer pitches if they correspond to higher probability fingerings. Further improvements could include examining octave alterations for unplayable pitches, supporting different guitar tunings, and generating exercises with varied fingering characteristics, ultimately making the system more versatile and musically intelligent.
