Deep Polyphonic ADSR Piano Note Transcription

Abstract

We investigate a late-fusion approach to piano transcription, combined with a strong temporal prior in the form of a handcrafted Hidden Markov Model (HMM). The network architecture under consideration is compact in terms of its number of parameters and easy to train with gradient descent. The network outputs are fused over time in the final stage to obtain note segmentations, with an HMM whose transition probabilities are chosen based on a model of attack, decay, sustain, release (ADSR) envelopes, commonly used for sound synthesis. The note segments are then subject to a final binary decision rule to reject too weak note segment hypotheses. We obtain state-of-the-art results on the MAPS dataset, and are able to outperform other approaches by a large margin, when predicting complete note regions from onsets to offsets.

Multitask Learning for Polyphonic Piano Transcription, a Case Study

“…Most hyperparameters, such as minibatch size (64), choice of optimizer (gradient descent with momentum (0.9) and Nesterov correction [16]), parameter initialization (drawn from a uniform distribution according to [17], colloquially called "Glorot, Uniform") as well as kernel sizes were chosen based on prior knowledge that they lead to acceptable performance for the smaller model described in [10]. The number of feature maps for the scaled up version of the model was chosen as a trade off between capacity and training time.…”

“…The latter is conditioned on the onset predictions from the first network as an additional input, but there is no shared representation and there are no shared parameters for the two tasks. The approaches in [9], [10] introduce an additional note offset prediction task, but differ in their parameter sharing strategy. No parameters are shared in [9] (Figure 1a), and three separate networks are trained to predict onsets, intermediate note frames and offsets respectively.…”

“…No parameters are shared in [9] (Figure 1a), and three separate networks are trained to predict onsets, intermediate note frames and offsets respectively. In contrast, [10] use hard parameter sharing (Figure 1b), where almost all network capacity is shared among the three prediction tasks, with little room for individual specialization. The models in [2], [8] have an additional output that predicts key strike velocity.…”

Joint Multi-Pitch Detection and Score Transcription for Polyphonic Piano Music

“…To obtain data to develop our model, we create a synthesized dataset 2 using scores collected from the MuseScore website 3 . We do this as a starting point and because there is a lack of AMT datasets that provide score ground truth on both physical and musical time.…”

“…A large part of work in AMT falls under the tasks of multi-pitch detection and onset/offset detection (e.g. [2,3]), which are often referred to jointly as note tracking.…”